The interesting thing was watching what happened to the headline number on the way through. It was introduced in early August at 2.5 per cent. The Act that passed says 2.75 per cent. That change landed somewhere in the final stage of amendments, and days after it passed, journalists and the trade press were still reporting the superseded figure.

In the end, though, what the number was matters less than what the scheme means, and that’s where all the coverage went: platforms and publishers. Nobody asked the question this edition is about. Who actually pays.

The machine parliament actually built

To make sense of the design, you have to go back to the failure it replaces. The 2021 Bargaining Code was, at its heart, a threat: do deals with publishers or the Treasurer designates you, and designation meant forced arbitration over what news is worth. And it worked, for a while. Three years of deals worth an estimated $200 to $250 million annually. Then in March 2024 Meta called the bluff and declined to renew, on the basis that “people don’t come to Facebook for news and political content”. The bluff was callable because everyone had watched what happened in Canada, where Meta responded to a similar law by blocking news entirely. That block is still in place today. If Australia designated Meta, the same button was sitting right there. So for two years we had a law that couldn’t touch the one company most publicly refusing to pay.

The NBI is built to remove that exit. The charge lands on Australian advertising revenue whether or not the service carries a single news article, which means blocking news no longer saves a cent. It’s worth pausing on the name for a second, too, because the word bargaining really only survives in the title. There’s no designation power in the Act, no arbitration, no negotiation over what news is worth. What there is, is a charge, and a discount for settling privately.

The easiest way to understand the machine is to run an imaginary platform through it. Picture one booking a round $1 billion a year in Australian ad revenue, comfortably clear of the $250 million threshold where the law switches on (in practice the list is Meta, Google, TikTok and Microsoft’s LinkedIn). The meter reads 2.75 per cent, so our platform owes $27.5 million for the year. Then come the discounts. Every dollar it spends with a large publisher takes $1.50 off the bill, and every dollar with a small or medium publisher takes $2 off. There are two house rules: the deals have to span at least eight Australian news businesses, and no single agreement can wipe out more than a quarter of the bill (a ceiling News and Nine managed to push up from 16.6).

Which brings us to the first thing worth understanding about the NBI: it’s a tax designed to collect nothing. If the charge is ever actually paid, the money flows into a distribution scheme for newsrooms. Notice the conditional doing the work in that sentence, though. The whole distribution architecture only switches on if a platform chooses the charge over the deals, and every number in the design exists to make sure none of them do. The Acts are in force now, by the way, and the first charge year is 2026-27, the financial year we’re already in.

Follow the money in

So the platforms will do deals, the publishers will get paid, and the press releases will say big tech is paying for news. My question is what happens one step further back.

You already know this machinery, as it happens, from every cafe counter in the country: the small sign saying cards incur a 1.5 per cent surcharge. Google runs the same sign at invoice scale. Its own help pages list the matching invoice lines across half a dozen countries, and the cleanest case is Canada. When Canada imposed a 3 per cent digital services tax, Google added a 2.5 per cent Canada DST fee to every invoice from October 2024, and only removed it once Ottawa moved to repeal the tax. The sequence never varies: a government levies a percentage on platform ad revenue, the platform converts it into a line on the advertiser’s invoice, and the advertiser pays it.

The NBI can reach you down one of two paths, and the difference between them is really just whether you can see it. The levy path is the visible one. A platform that refuses to deal pays the 2.75, and if Canada is any guide, most of that lands on invoices as something like an Australia Regulatory Operating Cost, with Canberra there to blame. The deals path is the invisible one. The Code produced $200 to $250 million a year and nobody ever saw a surcharge line for News Showcase, because deal costs dissolve into the price of the product the same way the cafe’s rent dissolves into the coffee. No line item, no announcement, no one to blame.

And here’s the design point that took me a while to see. The discounts don’t just make deals attractive, they make the invisible path the cheaper one: $18.3 million in large-publisher deals, or $13.75 million through the small end, against $27.5 million for the visible route. The statute prices visibility.

The solver in the middle

Now look at what the Act hands each platform, because it’s a shopping list with rules. Spend whatever you like across at least eight suppliers, none earning more than a quarter of the credit, small ones counting double. The obvious next question is the cheapest way to tick every box, and in software that question goes to a solver, the routine that finds the lowest-cost answer inside a set of rules. I think solver is the honest name for what each platform’s deal desk becomes under this Act.

For our $1 billion platform, the three routes price like this.

1. Paying the charge costs the full $27.5 million which is the 2.75%.

2. Offsetting entirely through large publishers costs $18.3 million, which works out at 1.83% of revenue instead of 2.75.

3. Offsetting entirely through the small and medium publishers costs $13.75 million, or 1.375%

So compliance comes out between half and two-thirds of the headline rate, with the cheapest mix running through the small end of the market. Meanwhile the quarter cap means four big cheques can carry an entire offset, with four smaller handshakes making up the required eight signatures. Scott Purcell made exactly this criticism from the small-publisher side before it passed: the cap stops the money concentrating in one publisher, and does nothing to stop it concentrating in four.

Which is the point of this edition, and honestly of this newsletter. The government didn’t decide which newsrooms get funded here. It wrote the rules of a game in which the platforms decide. Canberra wrote the constraints. The platforms run the solver. Which mastheads, at what value, on what terms, all of it is solver output, computed on the deal desks of four companies, none of them Australian. Follow the control, not the conversation.

The open questions

This is usually the part of an edition where I’d tell you what to do about it all. Not this time. The machine is barely a fortnight old, and I think the honest position right now is a set of questions rather than answers. So here are the four I’m carrying into the next twelve months, with an early lean on each. Treat the leans as provisional, because they are.

Does subsidised journalism become more inventory?

This is the one the industry will want to be true: a couple of hundred million a year into newsrooms becoming more content, and more display and video supply for buyers. The thing is, the Code years already ran this experiment. The money flowed from 2021 to 2024 and newsroom contraction continued right through it, which suggests it behaved as a floor under existing cost bases rather than a growth fund. And none of it re-competes for your budget anyway. The deals fund journalism, and the money left publishers because the ad product fell behind, which this scheme doesn’t touch. More pages only become more inventory if traffic follows, and the referral channels that used to reward volume are the ones AI answers are closing.

Does news come back to Meta?

There’s a strange loop in the design here. The economics that justified walking away are gone, and the cheapest route from a $27.5 million bill down to the $13 to 18 million range runs through deals with the very publishers Meta spent two years winding back from. But the offset rules cover payments for the production of news content, or for news being made available on the platform, and production alone qualifies. Which means Meta can reach the cheap end of the bill without restoring a single link. Nothing in the Act forces the comeback, so the first deal round becomes the tell. Production-only deals would mean the market’s largest social platform funding journalism it never shows, which would be the purest expression of this whole design. I’d watch Meta’s first reporting period the way you’d watch a product launch, because commercially that’s what it is.

What does the small end do with its first platform money?

The double credit and the eight-signature floor push cheques toward regional, niche and multicultural publishers the Code never reached, and it happens to land just as client briefs keep asking for communities the usual schedule misses. The question for a planner is which of those publishers convert a cheque into ad product and a working sales capability.

And then the question I think decides how this era gets remembered: pension or runway?

A pension plugs the P&L and defers the reckoning. Runway funds the rebuild this newsletter keeps returning to, inventory an agent can discover, read and price, the seller-side plumbing that decides whether a publisher exists to the next generation of buyers at all. Nothing in the Act directs the money either way, so the split will happen publisher by publisher, and it should be visible within a year or two in who ships what. The Code money mostly bought time. Whether this round buys anything else is, for my money, the question that matters most.

The base is the exemption

One more thing about that 2.75 per cent, because the number it multiplies is quietly the biggest editorial decision in the Act. The charge is calculated on advertising revenue. Which means any company that consumes Australian news without selling Australian ads sits outside the instrument entirely, by construction. OpenAI, Anthropic, Perplexity: services that read the news, summarise it, and increasingly answer the reader with it, most of them at zero local ad revenue. And the exclusion is doubled in the drafting, because a service that solely or primarily uses large language models is, by definition, not a search service under this scheme. Australia answered the ChatGPT question before Brussels has managed to, twice over.

Picture the scheme as a tollbooth on the twenty-year highway between platforms and publishers, while the bypass goes up beside it: AI systems reading the journalism, answering the reader directly, and sending nothing back down the old road. No traffic, no toll. It’s why News Corp is running its “woo and sue” campaign against the AI companies in parallel, on private paper, rather than waiting for Canberra. The public instrument reaches the platforms that used to send publishers traffic. It can’t reach the systems now consuming the journalism without sending anything back. The meter is bolted to the road that is closing.

None of which makes the NBI pointless, to be clear. Two hundred-odd million a year is real money for Australian newsrooms, extracted by a design no other country has managed. But it's worth knowing it for what it is: a levy on Australian advertisers, allocated by platform deal desks, indexed to the economics of an era that's ending. Budget accordingly.

Share

Systems That Decide publishes on the infrastructure of decision-making in marketing and advertising. If this edition was useful, share it with someone who still thinks big tech is paying for news.

For a good stretch of my career I worked at Integral Ad Science, selling ad verification into this market. Every pitch rested on one word: independence. You can trust the number because the company producing it has no stake in what the number says. It’s a clean argument, I made it for years, and I still believe it, which is why I’ve read one announcement from last Thursday about five times now.

Two things happened on 6 August, hours apart. First, The Trade Desk reported revenue of $715 million for the June quarter, up 3 per cent on a year earlier. For a company that spent a decade compounding above 20 per cent, three is a repricing. The stock fell as much as a fifth in after-hours trading, and Jeff Green told analysts the result was “below our expectations and below the standard we hold ourselves to”. His diagnosis is worth reading twice: brands, he said, are “falling prey to low-cost, low-decisioning methods”, and he named programmatic guaranteed and fixed-price buying. Guidance for the current quarter implies revenue actually declining, which has never happened in the company’s reported history.

Hold onto that phrase, low-decisioning, because I think it’s doing more work than Green intended.

The second announcement is the one I keep coming back to: Nielsen agreed to buy DoubleVerify for $2.15 billion in cash. Neither company mentioned the other, and the coverage treated them as separate stories, which is fair enough on the surface. I don’t think they’re separate at all. One is the market repricing the visible decision. The other is a company buying up the invisible one.

The second edition of this newsletter split programmatic into two functions, decisioning and execution, and argued the value sits with whoever holds decisioning. The market has now run the experiment, and the result is stranger than the thesis. Decisioning is being commoditised where you can see it and concentrated where you can’t. The price of a decision made in the open, at trade time, for a disclosed fee, is collapsing, while the value of owning what decisions are made from, the signals, the algorithms, the measurement that grades the outcome, keeps climbing. Thursday was both halves of that trade printing on the same day.

The visible decision gets cheaper

A DSP’s product, stripped to its essence, is a decision. An impression appears, the platform decides in milliseconds whether it’s worth buying and at what price, and the buyer pays a fee for that decision. The Trade Desk’s fee has sat within a point or two of 20 per cent for a decade, and Green’s own explanation for the stall is that buyers are declining to pay it. Programmatic guaranteed and fixed-price deals don’t remove the decision, they just move it out of the auction and into a negotiation that happened weeks before your brief existed, and they price it differently.

Priced differently is the polite version. A fee is a price you can see on an invoice. A spread is the gap between what your agency paid for the media and what you paid your agency for it, and if you’ve never gone looking for that gap, that’s rather the point. Both are prices for decisions. Only one survives an audit.

The spread has a name, principal media, and 2026 is the year it stopped being whispered. In March, Publicis told clients it would no longer recommend The Trade Desk, citing a third-party audit it says found billing and consent violations. TTD’s denial was flat: “Any notion that TTD failed an audit is not true.” Omnicom opened its own audit within the week. And Green, in the middle of that fight, pointed at what he called “the most ironic programmatic practice, principal-based buying”, his phrase for agencies attacking his transparency while moving client dollars into transaction methods with none.

Five months later, his earnings-call diagnosis described the same migration without naming it. Make of that what you will. My read is that it’s diplomacy rather than amnesia, given the holdcos he’d be naming are also his largest customers.

The filings describe it fine anyway. Omnicom’s half-year accounts carry a line for third party service costs, almost $2.9 billion for the half against $1.7 billion a year earlier, and the footnote attached to it says those costs include supplier costs incurred when Omnicom acts as principal for its clients. Now, that comparison spans the Interpublic merger, so a large part of the jump is consolidation rather than growth, and I don’t want to oversell the delta. The more interesting fact is simply that principal is now big enough to need a footnote.

The buyers can see it happening, too. The ANA found in March that 90 per cent of marketers name one worry above all others, whether the principal media recommended to them is actually in their best interest, up from 79 per cent two years ago, while only 57 per cent have any governance rules for the practice. And Forrester says 81 per cent plan to buy more of it this year anyway, which tells you the discount is doing the persuading. Publicis, for its part, says principal is “not a strategic priority” and that it is far from the biggest player, and I’m happy to take that at face value, because the structural point doesn’t need Publicis: the dollars are moving from a disclosed fee to an undisclosed spread, and every incentive on a holdco earnings call points the same direction.

To be honest about the stall itself, the spread is only one of three forces squeezing The Trade Desk right now. Big CPG and auto advertisers are cutting, the company has had a year of self-inflicted execution problems, and Amazon’s DSP is taking share at prices The Trade Desk won’t match. Green’s reasons are real, they’re just not complete. Underneath all of them sits the structural one: a decade of growth was built on the premise that auction-time decisioning is worth a fifth of the media, and that premise is now being tested from three directions, the spread you’ve just read about, and two more further down this page.

The invisible decision gets bought

Which brings me back to the announcement I’ve been re-reading.

Nielsen is paying $13.60 a share for DoubleVerify, all cash, at a premium of roughly 30 per cent, with the deal expected to close in the first quarter of 2027 and DV keeping its brand. On its face this is measurement consolidation, the audience currency buying the verification referee, and that alone would deserve a column, because it’s the second such exit in twelve months. Integral Ad Science, the company I used to sell for, went private in a $1.9 billion deal with Novacap within the past year, and Nielsen itself has been privately held since 2022. The companies the industry pays to referee its numbers now share a characteristic with the deal rooms: no public filings.

But DoubleVerify stopped being just a referee some time ago. In 2023 it bought Scibids, which builds custom bidding algorithms, actual buying decisions made by software. It also owns Rockerbox, which sells the attribution and mix modelling that grades those decisions afterwards.

ExchangeWire sketched what the combination enables within days of the announcement, and their mapping is the clearest thing published on this deal so far. In their read, the merged company becomes what they call a “Super Signal Aggregator”: Nielsen’s panel-verified, person-level reach data joins DV’s verification, attention, contextual and quality signals; those signals train buying algorithms through Scibids, activated into DSPs and into sell-side curation; and the outcomes are then measured by DV’s attention products and by Rockerbox models that train on the same Nielsen reach signal. To be clear, that’s their projection rather than an announced roadmap, and the deal hasn’t closed. But look at the shape of it: the signal, the decision and the verdict, under one owner.

This is the layer I know from the inside, so let me say plainly what the product actually is. The product is independence. A verification company’s signal is worth something because the company selling it has no stake in what the signal says. So the response this deal deserves isn’t outrage, it’s three questions, the same three I’d put to any company that ends up owning an appraisal layer. Will the signals be available to rival platforms on the same terms as its own activation paths? Who audits the loop from outside it? And can the referee’s call ever cost the owner money? If the answer to all three turns out to be yes, this is a formidable and honestly quite defensible product. If any one of them comes back no, the industry has recreated, at the measurement layer, the exact conflict it spent a decade litigating out of the media layer. Who grades the homework when the tutor, the student and the examiner bill to the same account?

What changed on Thursday, in my view, is the burden of proof: independence used to be the default assumption about a verifier, and it’s now a claim the merged company will need to keep demonstrating. The three questions are how this market should test it.

And to be fair to the deal’s logic, because there is one. Fragmented measurement is the single most common complaint of the people who read this newsletter; roughly a third of marketers measure across all their channels in one model, and the rest stitch. A unified signal stack is a real answer to a real problem, which is exactly what makes this hard. Consolidation really does solve fragmentation; it just charges auditability for the privilege. Whether that price is worth paying depends entirely on those three answers, and nobody outside the deal knows them yet.

The industry starts naming the prize

The tell that this is a repricing of the whole market, rather than one company’s rough year, is who else wants the same asset. Magnite, the largest independent sell-side platform, has been explicit that it doesn’t want to become a DSP; what it wants to own, per AdExchanger’s coverage, is “the decisioning layer”. When the sell side starts bidding for the decision, the decision has left the building it used to live in. And the buy side is quietly saying the same thing from the other direction: even the DSPs are building the agents that will replace their own interfaces, which is a company telling you where it thinks the decision is going.

Google, covered here two weeks ago, has Buyer Direct in open beta: agency buyers booking guaranteed inventory at fixed prices directly inside the ad server, with no DSP, no SSP, no exchange, and no auction-time decision to pay anyone for. The workaround to the remedy is also a bid for the decision, made from the infrastructure layer, while the court that would reshape that infrastructure continues to deliberate.

So that’s three bidders across three layers: agencies pricing the decision as spread, a measurement owner pricing it as signal, and a platform pricing it as infrastructure. What nobody seems to want to sell anymore is the decision itself, in the open, at a visible price. That was the product of the last fifteen years, and 6 August was the day its repricing became hard to unsee.

What to do with this in a concentrated market

The ANZ lens makes all of this arrive faster. In a market where two grocers, a handful of banks and a short list of holdcos set the terms, ownership changes at the signal layer and shifts in how demand clears compound quickly, and whichever seat you’re in, you’re likely the person your leadership team will ask to explain it. So let me split this by seat.

If you’re brand-side, start with a map you can build in an afternoon: every signal your buying decisions run on, verification, audience currency, attribution, bidding algorithms, and who owns each one today. That map is the one-pager for your leadership team, because for most advertisers the ownership column changed this year and nobody has updated the assumptions, and the version of that conversation you start is a plan, while the version someone else starts is a review. Then take three things to procurement at the next contract cycle: change-of-control language, so a sale of your measurement partner reopens access and pricing terms rather than quietly inheriting them; signal and data portability, in writing, for the day you move platforms; and on principal, the terms the ANA and Forrester both suggest belong in the agreement itself, opt-in at the overall contract level, an agreed ceiling as a share of spend, Forrester puts it between 10 and 20 per cent, and clear reporting. None of that is a judgement on principal. Done well it trades flexibility for price, and plenty of advertisers take that trade with their eyes open; the provisions simply keep the choice explicit.

If you’re agency-side, the same migration is a commercial story you’re better placed to tell than anyone, so tell it first. Clients are reading the same coverage you are and the questions about principal will come; the useful posture isn’t defensive, it’s fluent. Know what your desk offers, when it genuinely beats the fee model for a client, and how you describe the economics when someone asks directly. Do the same signal map for the tools you resell, and know your answers to the three loop questions for each of them, because your clients will inherit whatever those answers turn out to be. The agency that can explain the new economics in plain terms usually ends up leading that conversation rather than responding to it.

If you’re a publisher, this edition is better news than it probably reads. Demand is migrating toward deals, guaranteed buys and direct paths, and buyers with fixed-price appetites and principal commitments need packaged, guaranteed product to spend against. The work is readiness. Audit where your revenue actually clears today, open auction versus deals versus direct, because a yield strategy built for open-auction density ages quickly in this shift. Then look at the signal layer from your side of it: your inventory is increasingly scored and curated by the same signal stacks this deal consolidates, so ask your SSP and curation partners how those signals are sourced and what happens to the scoring of your inventory if ownership or terms change. And talk to your biggest agency partners about what their desks actually want to buy directly, because someone is going to package the inventory those desks are looking for, and it may as well be you.

How long does this run? My honest answer is that the migration itself doesn’t end. It’s been moving since about 2020, as the chart above shows, and the direction is set. The window for setting terms is narrower. The Nielsen deal closes around the first quarter of 2027, so commitments on access and auditability are easiest to get before the combined product exists, and with most media contracts in this market renewing annually, the next cycle or two of renewals is where terms get set for the market that’s arriving. That window is the same for all three seats. None of this requires believing anyone is behaving badly. It requires noticing that the structures that would let you check are being renovated, and that the renovation is not being done for your convenience.

Where the decision lives now

So, an answer to both halves of the title. The decision is leaving the place it lived for fifteen years, the open auction with the visible fee, and taking up residence in deal rooms as spread, in ad servers as fixed prices, in algorithms as trained weights, and in the signals that train them, which as of Thursday have a new concentration of private ownership.

There’s a harder question underneath all of it. A growing share of these decisions are executed by software agents, and some of those agents belong to parties holding positions in the very trades they execute. Whose agent is the agent? I don’t have a tidy answer to that one yet, which is exactly why it gets its own edition.

For now, take the question this newsletter will keep asking and run it over your own stack this week: where did the decision just move?

Next edition: the auction itself. Deals, direct paths and agents are pulling demand out of the open exchange, the market structure programmatic was built on and the one a federal judge is still writing remedies for. Does it die, or does it get demoted?

On 17 April 2025, Judge Leonie Brinkema of the Eastern District of Virginia ruled that Google illegally monopolised two markets: the publisher ad server market, the software that manages which ads show up on a website, and the open-web ad exchange market, the real-time auction that decides which advertiser wins each impression. She also found Google illegally tied the two together, forcing publishers who wanted access to Google’s exchange to also run Google’s ad server. The DOJ alleged a third monopoly, in advertiser ad networks. Brinkema ruled they hadn’t proven it. During trial, former News Corp ad-tech executive Stephanie Layser testified that her company estimated in 2017 it would lose roughly $9 million by moving its business off Google’s ad tech products, the kind of captivity the tying claim was built to describe.

We need to be precise about what Google actually owns here, because the whole case lives in the plumbing. When an ad loads on a website, three machines fire in sequence: the advertiser’s buying tool decides what to bid, an auction house picks the winner, and the publisher’s booking system serves the ad. Google owns the dominant version of every layer. This case is about the last two. The scandal was never any single fee. The auction house and the seller’s booking system are the same company, and that company also runs the biggest buyer.

Google’s framing afterwards was “we won half of this case and we will appeal the other half.” That’s true and also beside the point. The half it lost is the half that built the modern programmatic stack, the automated machinery that buys and sells most of the ads on the open web.

That ruling is over a year old. Nothing about it is contested anymore except through appeal. The actual fight, the one nobody outside the trade press is paying attention to, is what happens next. And it has been sitting, unresolved, for months.

Which is where this edition parts ways with the coverage. A verdict in Washington, a fine in Brussels, five years of consultation in Canberra, all saying the same thing, none of it yet changing anything. And while everyone deliberates, Google is already rewiring the machine they’re deliberating about. A direct line between agency buyers and its ad server went into open beta last month. No exchange in the path. If it works, whatever finally gets ordered lands on a system that has already moved, a decision rendered null and void by product roadmap rather than by appeal. The back half of this edition is about that.

The part everyone is still waiting on

Closing arguments in the remedy phase wrapped on 21 November 2025. The DOJ wants Google to divest AdX, its ad exchange, and to open-source the auction logic, publishing the code that decides who wins each impression, and at what price, for anyone to inspect and run. As a fallback, it wants Google Ad Manager, the publisher ad server, phased out or sold off entirely. Google’s counter is behavioural, not structural, conduct rules instead of asset sales: pricing flexibility for publishers, more interoperability between its tools, no forced sale of anything.

Brinkema set herself a 31 March 2026 deadline for the ruling. She missed it. As of early August, there is still no decision filed. The reporting that once pointed to a final order by the middle of the year has quietly shifted to any week now, and the appeals that will follow regardless of which way she rules are expected to run for years past that.

Part of the delay is structural and worth sitting with, because it’s the most interesting part of the whole case. At closing arguments, Brinkema pressed the government on a practical problem with its preferred remedy: it still hasn’t found a plausible buyer for AdX, and any company big enough to actually run it, Microsoft was the name raised in court, would face its own lengthy antitrust review the moment it tried. A court can rule that a company has too much control. Building a mechanism that actually removes that control, without just handing it to whoever’s next biggest, is a different and much harder problem. That gap, between a verdict and an enforceable remedy, is the real story here, and it isn’t unique to this courtroom. Two other regulators have already hit the identical wall, worth seeing side by side before going further.

This is the argument this newsletter opened with

The first edition of this newsletter argued the Publicis vs. The Trade Desk fight wasn't about transparency, it was about who controls the systems that decide where money flows. This is the same argument at maximum scale: the company accused of owning too much of the system is the biggest company in the industry, and this time it's a federal courtroom with an actual remedy on the table. A live case, already decided on liability, stuck on the only question that matters: can anyone build a remedy big enough to fit the problem.

Three jurisdictions, same diagnosis, three different speeds

Here's what gets missed in most of the US coverage: Washington isn't the only place that's reached this conclusion, and it isn't even the place that reached it first.

The European Commission fined Google €2.95 billion in September 2025 for self-preferencing in ad tech, conduct it says has been running since 2014. Google’s response, filed in mid-November, rejected a forced asset sale and offered product changes instead: different minimum prices for different bidders, more interoperability across its tools. Brussels market-tested that offer through the winter, asking publishers and rivals whether it would actually change anything, while competition chief Teresa Ribera said publicly that the only effective way for Google to end its conflict of interest may be structural, selling part of the ad tech business. Google, meanwhile, is appealing the fine and the findings in the EU courts. The first half of 2026, when a decision was expected, has come and gone. If the Commission decides the behavioural offer isn’t enough, it can force a divestiture on its own authority, separately from whatever Brinkema decides in Virginia.

Australia got there before either of them. The ACCC’s Digital Advertising Services Inquiry, finalised in September 2021, found that more than 90 per cent of ad impressions traded through the Australian supply chain touched at least one Google service, and that ad tech providers were keeping at least 27 per cent of advertiser spend as something close to a structural toll. The recommendation was an ex ante regime, rules that apply up front to designated platforms instead of lawsuits after the damage, with ad tech named one of the first services slated for designation. Nearly five years on, the enabling legislation still hasn’t made it into Parliament, and the designation investigation, a six-month process once it starts, is not yet underway for ad tech.

Put the three side by side and the pattern isn’t about whose regulators are smarter. Australia diagnosed the structural problem with sharper numbers, earlier, than either Washington or Brussels. What it didn’t have was a mechanism already built to act on the diagnosis once it was made. The US has a court with the power to order a divestiture and still can’t identify how to execute one cleanly. The EU has a Commission willing to threaten one and is negotiating over whether a lighter remedy will do instead. Australia has the clearest evidence of harm and the least developed enforcement instrument to act on it.

That’s the actual lesson sitting inside this story, and it’s bigger than Google. A finding of excessive control isn’t the same thing as a working method for removing it. Most of the regulatory conversation treats the verdict as the hard part. It is not. The hard part, in every one of these three cases, is building a remedy that actually changes who holds the system once the ruling says they shouldn’t.

What to actually watch from here

Four things, none of which are “will Google lose.” It already has, in two of the three forums that matter, and the third has effectively reached the same conclusion without the mechanism to enforce it yet.

Watch whether Brinkema orders a structural remedy or a behavioural one, though here I’ll make the call instead of just framing the question. At closing arguments she pressed the government on its failure to name a plausible buyer for AdX, and any company big enough to run it would face its own antitrust review, so a forced sale to whichever bidder is next biggest just relocates the monopoly instead of removing it, that’s not happening. But she has also shown no appetite for theatre. “This is to some degree window dressing,” she said mid-trial during the remedies phase, waving off another round of abstract industry testimony to demand the actual mechanics of a separation, “let’s get to the mannequins.” Most of the smart money, the analysts and antitrust lawyers who read this docket for a living, expects a behavioural order, and at closings she noted how much of Google’s offer overlaps with the government’s own behavioural asks.

My read is still that she splits it, a court-ordered structural separation that doesn’t require an outside buyer, AdX and Google Ad Manager run and audited as genuinely independent units inside Google, with binding interoperability and pricing mandates attached. A forced sale would still be the first time a US court has ordered a breakup of a major tech platform’s ad tech assets. A pure behavioural remedy would mean the company that built the monopoly keeps running it exactly as before, just with more disclosure attached. Those are the two poles this newsletter has been describing since the first edition, and my bet, now the contrarian one, is that she lands in the middle neither side asked for. That’s a prediction, not a fact, and it’s the one line in this edition that could age badly.

Watch whether the EU accepts Google’s interoperability commitments or decides they’re insufficient and forces a sale anyway. My read on Brussels: nobody wants to move first. If Virginia goes behavioural, the Commission takes an upgraded version of Google’s offer and declares victory. If Virginia orders structure, Brussels hardens. Either way the first mover sets both, and how that plays shapes what every other regulator watching, including the ACCC, believes is actually achievable.

And watch the Australian designation timeline specifically, because it’s the one piece of this story local advertisers and publishers can actually act on. If ad tech designation moves from consultation to investigation this year, ANZ moves from being a market that diagnosed the problem to one that’s finally building the instrument to do something about it. If it doesn’t, the gap between evidence and enforcement that’s defined the local conversation since 2021 just gets wider while two other jurisdictions move past it.

Watch what Google ships while she deliberates, because the workaround is already in public beta. Buyer Direct, quietly announced as an Ad Manager feature late last year, shown off at Cannes in June and expanded into open beta in July, lets agency buyers book guaranteed inventory, sold in advance at a fixed price the way print pages used to be, directly inside Google Ad Manager. No DSP, no SSP, no exchange in the path, none of the intermediaries that sit between an agency and a publisher clipping the ticket on the way through. Agentic capabilities are reportedly next. Read that against the remedy under deliberation: this case is about the tie between AdX and the ad server, and the leading structural fix is built around AdX. Buyer Direct routes demand through the ad server directly. If it works, Google doesn’t need the exchange to keep the demand, and the remedy everyone has spent two years arguing about starts aiming at a layer Google is already moving past. One agentic ad tech founder, Gamera’s Gareth Glaser, wrote in AdExchanger that he’s not sure any outside agentic system can outperform it. The counterargument is the DOJ’s fallback ask made flesh: none of this survives if the ad server itself has to go.

And the layer it's moving past is shrinking on its own. Google Network, the Alphabet reporting line that houses the sell side, the tools publishers use to sell their space, AdSense and AdMob for independent websites and apps, Ad Manager for everyone bigger, has now declined year on year for six straight quarters and is down 9 per cent from its 2022 peak, running at about $7.3 billion a quarter. It's 6 per cent of Alphabet's revenue, down from roughly 13 per cent in 2020, and it shrank in a quarter where Search grew 17 per cent, YouTube 13, Cloud 82, and the earnings call mentioned TPUs 28 times without saying the word publisher once. Google has already modelled the endgame twice. Internal studies surfaced at the remedies trial, Project Sunday in 2020 and Project Monday in 2021, examined fully shuttering AdX and licensing the ad server out through Google Cloud. Put it together and the delay has a quiet arithmetic to it. Every quarter without an order, the asset Brinkema would have to find a buyer for gets smaller, matters less to Google, and drifts closer to the outcome Google sketched for itself five years ago. A judge worried there's no plausible buyer for AdX may end up ruling on an exchange nobody needs to buy.

None of this resolves quickly. But the question stopped being whether Google has too much control over the system a long time ago. The question now, in every jurisdiction running this case, is whether anyone has actually built something capable of taking it back.

My honest answer: not yet, and maybe not in time. The order will land, the appeals will run, Brussels will calibrate against Virginia, Canberra will keep consulting, and somewhere in a product roadmap the system will quietly finish moving. If you buy or sell media for a living, that's the version to plan for. Not the breakup. Not the behavioural settlement. A Google that had solved the case before the court could. The first edition of this newsletter asked who controls the systems that decide where money flows. A federal judge answered it fifteen months ago. The question that matters now is who'll control them by the time the paperwork catches up.

It isn’t ignorance. The people writing the specifications know exactly what the word means. They’re just not in the room with you, and nothing they’ve written has made it as far as your next vendor meeting.

Here is the range that word is currently being asked to cover. In July, an audio campaign for a Dutch retailer went from brief to booked with no human anywhere in the transaction: the buying agent sent the request, the selling agent negotiated the terms, and the booking landed in the ad server, machine to machine, according to the companies that ran it. At the other end of the same word, the most talked-about agent in the industry, The Trade Desk’s Koa, will build your audience strategy, set up your campaign and troubleshoot your delivery. It cannot buy a single ad.

Both are agentic. So is the tool that drafts your campaign copy, and the black box moving your budget at 3am. It is the most overloaded term in advertising, and the vendors like it that way.

Gartner put numbers on the inflation back in June 2025. Of the thousands of vendors claiming agentic capability, it estimated only about 130 were the real thing, named the practice agent washing, and predicted more than 40 per cent of agentic AI projects would be cancelled by the end of 2027. Thirteen months on, the label has spread faster than the capability.

So before the next pitch, one question matters more than any demo. What have you authorised this thing to do without asking you first?

Everything in this piece hangs off that question. An agent is defined by delegated authority. The fluency is decoration.

Subscribe now

The line underneath the buzzword

There is a real distinction underneath the buzzword. It does not come from the ad industry, which is part of why it has not reached us. Anthropic drew it in late 2024. A workflow is software that runs on rails a human laid down. The steps are fixed and the model fills in the blanks. An agent decides its own next step. It holds a goal, picks its actions, uses tools, checks the result and adjusts.

If you run programmatic, you already live with automation at scale, which is exactly why the word confuses. Your bidder, your pacing logic and your smart bidding all execute a target you configured inside one system, and engineers have loosely called that kind of software an agent for twenty years. The new thing is delegation: you hand over an outcome and a slice of authority, and the software works out the steps itself. Nothing wrong with rails. When the cargo is your media budget, rails are often exactly what you want. The claim still deserves a second look.

The easiest way to hold the distinction is to stop thinking about software and start thinking about staff. There are three characters you’ll meet.

The Intern

The Intern does the task you hand it and gives the work back. It drafts the schedule, builds the plan you’ll rebuild, produces the supply path analysis, flags the pacing gap and suggests the reallocation. It touches nothing. Every output crosses your desk before it goes anywhere.

This is where most of the market actually sits, whatever the label says. Google’s Ads Advisor suggesting campaign changes. A planning tool producing three schedule options from a brief. A synthetic persona telling you what a 44-year-old category buyer might think of your concept. Useful, fast, and in most cases not an agent in any meaningful sense.

The Intern is also where agent washing lives. If the pitch deck says agent and the product description says recommends, surfaces, suggests or drafts, you’re hiring an Intern. Pay Intern rates.

What gets mislabelled

There’s a shorthand doing the rounds: AI answers, agentic acts. Keep it, it’s the fastest filter you’ll get. Just notice who benefits from stopping there. A binary lets every vendor above the line claim the same word, which is how a tool that stages changes for your approval ends up in the same category as software negotiating with a publisher. The line is a door, not a destination.

Four things most often wear the title without earning it.

A chat window on automation. Your bidder with a conversational skin is still your bidder. Natural language is an interface, and interfaces hold no authority.

Rules with a rebrand. Pause the line when the CPA passes fifty dollars is an instruction you wrote, executed on a schedule. Obedience isn’t judgement, and Gartner’s census suggests obedience is most of what’s in market.

A dashboard that talks. Reporting that speaks fluently is still reporting. The agent question is what acts on the report.

A persona with a job title. A synthetic customer panel is a focus group you can query at midnight. Useful in research, and it signs nothing.

The mislabel also runs the other way, and this one costs more. Plenty of teams file their approval-gated deployments under automation because a human still signs everything. If the software chose the plan and prepared the action, the delegation has already started, whatever you call it. Which brings us to the promotion.

The Intern gets the company card

The first real step up is a signature.

Give the Intern a company card and it can act. It builds the campaign, loads the budget, stages the changes. But every transaction needs your approval before it clears. The software prepares the action. A human signs it.

This is where the serious 2026 deployments sit, and the honest vendors say so. Koa is the example from the top of this piece: The Trade Desk was explicit in Digiday’s reporting this May that the agent is not yet buying ads. Google’s advisor products implement changes on your behalf, with permission. Amazon’s Ads Agent turns a media plan spreadsheet into launch-ready campaigns, then presents them for review.

Read those verbs closely. Prepares. Proposes. Presents for review. That’s the card-with-a-signature model, and it’s a sensible place to be. It’s also a much softer claim than autonomy, whatever the keynote implied.

The Autopilot

The Autopilot is the first character that acts without asking. You set the destination and the boundaries. It flies the plane continuously: adjusting bids, pacing spend, rotating creative, reallocating between audiences, thousands of decisions a day with no approval step on each one.

This tier is real and has been for a while. Closed-loop optimisation fits the definition of an agent better than anything else in market. Observe, decide, act, learn, repeat. If a vendor’s agent claim sits here, it’s usually defensible.

The question to ask about an Autopilot is different. Whose aircraft is this, and whose flight plan is loaded?

Meta’s Advantage+ and Google’s Performance Max are autopilots that live inside the platform’s own cockpit. When paired with their own analytics platform like GA4, it creates a two-black box problem. The agent is automating, its measurement is telling you it worked.

Reuters reported in June 2025 on Meta’s stated ambition: an advertiser supplies a product image and a budget, and the platform’s AI produces the creative, picks the audience and optimises delivery. That is an autonomous loop with no approval step anywhere in it. It is also a loop that flies toward the platform’s definition of arrival, measured with the platform’s instruments, on inventory the platform sells. You didn’t hire this autopilot. You boarded it.

None of that makes platform automation bad. It makes it a control question, and the cost of never asking it is specific: performance you can’t decompose, on inventory you can’t verify, graded by the seller. An autopilot you configure through open APIs, with your logs and your definition of success, is a different proposition from one you access through a toggle and a dashboard, even if both are correctly called agents.

The Dealmaker

The Dealmaker is the tier the trade press is excited about, and the one your team is least likely to have touched. Your agent talks to their agent. A buying agent takes a brief, discovers a publisher’s inventory through the publisher’s selling agent, negotiates terms, executes the buy and reports back. Machine to machine, end to end.

Is the Dealmaker real? In small numbers, yes. FreeWheel and NBCUniversal ran what they called the first cross-platform agentic media buy in January, and a cluster of live agent-to-agent transactions has followed. The frontier is the campaign this piece opened with. The Netherlands has produced national firsts in video, display and audio inside five months, and the phrase according to the companies is doing work: no no-human claim has been independently audited yet.

Worth asking why a small market got there first. The fewer counterparties in a market, the fewer deals anyone has to strike before a way of working becomes the way of working. Australians should recognise the shape: two grocery chains, a handful of banks, three commercial networks. Concentration is usually discussed here as a problem, and for pricing it is. For standards it’s an accelerant. It means these shifts tend to arrive early, settle fast, and be effectively decided before the market has finished debating them.

Everywhere else the Dealmaker still carries the company card. The agents discover, evaluate and agree the deal, then a person approves it before it clears, and the approval gate is written into the protocol as a feature rather than a stopgap. The harder truth sits underneath. Discovery, curation and activation are shipping. Software that bargains over price and terms against an independent seller agent is the piece that mostly doesn’t exist yet, and picking from a rate card doesn’t count as negotiating.

None of it works unless both sides accept the same paperwork, which is why the Dealmaker is the one character with a written job description. Both rulebooks are public documents, and reading them is the fastest education available in this category. The AdCP protocol, published through the Agentic Advertising Organisation, names the roles in specification language: buyer agents, sales agents, signals agents. The IAB Tech Lab is building a rival rulebook, AAMP, extending the standards the industry already runs on.

These are competing answers to the same question, and the contest is unresolved. The agencies running pilots today are mostly running them on one protocol. The other is positioning itself as the governance layer above all of them. Anyone telling you the standards have converged is selling something.

Two bodies. Two rulebooks. One market that can only use one of them at a time. Watch the paperwork fight more closely than the product demos. Whoever writes the schema the Dealmakers must speak will hold more of this market than any single vendor.

Same word, two dictionaries

Read the two rulebooks side by side and something awkward surfaces, and it is the thing to take into your next vendor meeting. The characters measure authority: how much you’ve delegated. The standards measure something else entirely: job functions and the language agents must speak. AdCP names buyer agents, sales agents and signals agents. AAMP arranges orchestrating and specialist agents into a hierarchy. Neither maps one-to-one onto the tiers, and putting them side by side forces two corrections.

First, nothing in either standard requires a Dealmaker to run without approval. Job function and authority are separate dials. A buyer agent can operate at company-card level or at full Autopilot trust, and today it almost always runs at the former. Second, under the standards’ definitions the platform Autopilots aren’t agents at all, because a walled garden has no counterparty and needs no shared protocol. The autonomy definition says agent. The protocol definition says automation. When a vendor says agent, ask which dictionary they’re holding.

The new-hire test

The practical takeaway fits on a sticky note. Stop asking whether something is an agent. Ask what you’d ask before delegating to a new hire:

1. What goal has it been given, in whose words?

2. What’s its spending authority, and what’s the cap?

3. Who signs off, and at what threshold?

4. Where’s the audit trail of what it did and why?

5. When does it escalate to a human, and how fast?

A vendor selling a real agent answers these in one breath, because the answers are the product. A vendor selling an Intern in a Dealmaker’s suit changes the subject to the model.

The numbers say most of your peers are still deciding whether to extend any authority at all. Comcast Advertising’s January report found fewer than a third of advertisers trust AI to do advertising tasks, and 61 per cent say they haven’t yet seen meaningful results from it. That’s not a technology gap. It’s a delegation gap, and it will close one signature at a time. Intern, card, Autopilot, Dealmaker.

Where this goes next

The verdict in one line: your 2026 exposure sits at the card and Autopilot rungs, where authority is already changing hands quietly, and your 2027 leverage is being decided in the paperwork fight, where almost nobody in your building is watching. Manage the first. Watch the second.

The people writing the rules still aren’t in your meetings. Until they are, someone has to keep bringing the documents back into the room, and translating them into the only question that matters at your desk: what has this been authorised to do without asking you first.

Everything above assumes the agent works for you. The stranger shift is already visible: agents that work for your customer. Shopping assistants that screen brands, compare claims and complete purchases before a human ever sees an ad. When the audience is partly software, persuasion starts to look less like attention and more like eligibility.

Follow the control, not the conversation.

That year, IAB Tech Lab shipped a text file. ads.txt: a few lines a publisher drops on a web server, listing who is authorised to sell its inventory. No product. No platform. No fee.

Watch how the power actually worked, because it’s subtler than the story usually gets told. IAB Tech Lab wrote the schema. Each publisher wrote its own declaration. And the decisive move belonged to neither of them: it came when the big DSPs, Google’s most consequentially, started refusing inventory that didn’t match the file. The domains we’d spent years flagging found themselves outside the fence. I saw it from the verification side, then from the demand side when I joined Yahoo DSP at the end of 2019, by which point checking the file was simply how the pipes worked. Three hands. The schema author, the declaration maker, the gate holder. The format was worthless until the gate enforced it.

Three layers, one question

Announcing the Government’s national AI framework last Wednesday, Assistant Minister Andrew Charlton described the goal as “a clear and enforceable social licence” for AI. Sit with that phrase. Social licence has always been the informal layer, the approval that lives outside law and contract. Read at face value, making it enforceable means someone must write the rules, someone must administer them, and someone must certify who meets them. That’s my inference from political language, not an announced legal instrument, but it’s the inference the rest of the week’s evidence supports.

The drafting is happening at three altitudes simultaneously. At the national layer, a new Office of AI will coordinate the design of mandatory national AI standards, with legislation due early next year. At the industry layer, two protocol families, AdCP and AAMP, are competing to standardise how agents transact advertising, and the contest is now running on Australian soil. At the business layer, the Privacy Act’s automated decision-making transparency obligation arrives on 10 December, and the declaration it requires is one you write yourself.

Notice the asymmetry. The attention each layer received this week is inversely proportional to how soon it binds you. The national layer has the headlines and no drafted rules. The business layer has a statutory deadline five months out and almost no coverage. So we will work down the stack, from the loudest layer to the one with your name on it.

Canberra: a licence it intends to enforce, eventually

Australia has written rules for Big Tech before, and everyone in this industry remembers how it went. The News Media Bargaining Code arrived after the platforms already held the audience, and when Canberra pushed, Meta switched off news for a week to demonstrate precisely where the gate sat. The Prime Minister’s speech reads like it was drafted by someone who remembers that week. The stated logic this time is to set terms while the infrastructure decisions are still live and Australia still holds pieces worth negotiating with.

Be precise about what was actually announced, because the ambition runs ahead of the machinery. The Office of AI’s announced role is coordination: pulling AI policy into one place and steering the design of national standards through a National Cabinet process. The concrete mandatory rules described so far concern large data centres and AI training. No registry, no certifier, no licensing system for anything resembling marketing was announced, and the Office’s own powers are yet to be legislated. Most of the release is data centre policy, and none of that moves a media budget.

Two details do touch this industry. The first is copyright: the PM declared that training AI on Australian creative work without the creator’s control is theft, and the market answered immediately, with Nine flagging more AI content deals behind its Microsoft agreement and the collecting societies declaring the licensing market open. The operator question arrives early: does your generative creative pipeline sit on licensed training data? Get that answer before your board asks for it.

The second detail is quieter and more important. Albanese said explicitly that the goal is not to legislate every eventuality, and that the framework must deal with issues in real time without bureaucracy contemplating every modification. That is principles-based regulation, and it has a specific consequence. When the statute stays deliberately thin, the operative rules get written downstream, in standards, guidance notes and compliance processes. Parliament sets the direction. The pen sits downstream.

Beyond those two details, stay conditional. The standards are undrafted, the legislation is six months away, and the promised AI consumer safety priorities do not exist yet. I will cover them when they do. Anyone selling you certainty about this layer in July is selling vapour. One altitude down, though, the pens are already moving.

The industry: two schemas, two registries, and a contested map

Two protocol families are competing to standardise agent-to-agent advertising. AdCP comes from the AAO, backed by a founding consortium including Yahoo, PubMatic, Optable, Scope3, Swivel and Triton Digital. AAMP comes from IAB Tech Lab.

How much they overlap is itself contested, and both maps deserve stating. AdCP’s documentation characterises the two as composable layers, AAMP as agentic bidding and AdCP as agentic buying. IAB Tech Lab describes AAMP more broadly, extending from bidstream execution into agent discovery, planning, negotiation, deals and campaign transactions. On the first map they stack. On the second they compete for the same ground. No joint roadmap exists to settle it, only parallel development and warm public language.

What is not contested is what each side is building, and here the ads.txt anatomy earns its keep. Both camps now operate registries, and they embody different architectures of authority. IAB Tech Lab, which wrote the ads.txt schema and knows exactly what format authorship is worth, launched its centralised Agent Registry in March: a list maintained at the centre, in the administrator’s house. The AdCP ecosystem runs the other pattern, in the direct lineage of ads.txt itself: publishers declare their authorised agents in an adagents.json file at their own domain, with an AAO-operated registry providing discovery and validation around those declarations. Two answers to the same question: who decides which agents are legitimate?

And in both architectures, the ads.txt lesson still applies. A schema plus a declaration is inert until someone holds the gate, refusing to transact with agents that aren’t on the list. Whoever’s registry the market’s orchestrators and platforms choose to enforce wins, whatever the specification documents say.

The race is accelerating locally. IAB Australia announced a local AAMP pilot at its AdTech & Ops Summit in Melbourne in early June, and has since reiterated it at its Data Summit, making this one of the first markets where the IAB path gets tested in production. Last week Adform announced ARTF support and AAMP 2.0 workflows, the clearest sell-side commitment to that track so far. The case for the IAB path, stated plainly: decades of standards administration, a membership spanning every major player, and now a live local pilot. The case for AdCP is a consortium moving fast on open, edge-declared infrastructure with serious sell-side and data weight behind it. I will not tell you which prevails. The prize, though, is the one Canberra is reaching for one level up: the authority to define compliance and certify who meets it. Same contest, different altitude.

Your business: December is the declaration you write yourself

From 10 December 2026, APP 1.7 of the Privacy Act requires entities to disclose automated decision-making in their privacy policies. The test has three limbs. A computer program makes, or does a thing substantially and directly related to making, a decision. The decision could reasonably be expected to significantly affect an individual's rights or interests. Personal information is used in the process. Meet all three and the obligation applies. That is a statutory deadline. Not a consultation.

The industry conversation treats advertising as an edge case here. The drafting record says otherwise. The Explanatory Memorandum to the Act itself gives the targeting of individuals with content and advertisements as an example of decisioning that may significantly affect a person, where it produces differential pricing for significant goods or services or limits access to employment opportunities. That is parliamentary material, written when the provision was drafted. Advertising was in contemplation from the start.

The regulator has gone further. The OAIC’s May Issues Paper, the consultation feeding guidance due by September, contains four fictional edge cases. Two are advertising scenarios. In one, an e-commerce platform charges a customer nearly triple for a book based on his postcode, and the OAIC asks at what percentage a price difference starts to significantly affect someone’s interests. In the other, a job ad is algorithmically promoted to male candidates, and the OAIC asks, in writing, whether it is a ‘decision’ that the female graduate never saw it. Two quiet provisions do heavy work underneath. Refusing or failing to make a decision counts as a decision. And effects count whether adverse or beneficial. Withholding is in scope, and “we only show people better offers” is not a defence.

The industry has answered. IAB Australia, the local trade body, lodged a submission supporting meaningful transparency while arguing that a broad reading would capture routine advertising activity, ranking, personalisation and pricing that involves no discrete decision about an individual, and urging a targeted, objective threshold. ADMA has argued along similar lines. For most of the stack we run today, that position is credible. An auction allocating impressions across millions of users is hard to describe as a discrete decision about a person, and the effect of seeing one ad instead of another is usually trivial.

Now hold that defence up against the previous section, and be precise about where it breaks, because it does not break everywhere. An agent that plans campaigns or buys inventory may never touch personal information or decide anything about an individual; protocol adoption alone triggers nothing. The collision comes with agentic personalisation and agentic commerce: systems that determine eligibility, personalise a price, negotiate terms, complete a transaction, or decide which opportunities an individual never sees.

Make it concrete. It is 2027, and a household’s commerce agent requests insurance renewal quotes. The insurer’s agent assembles one using the household’s claims history and postcode, prices it accordingly, determines this customer does not qualify for the retention discount others receive, and returns the terms. A computer program made the decision. Personal information ran through it. A price and an eligibility determination on an insurance contract sit inside the Explanatory Memorandum’s own examples of significance. Every limb of the statutory test, crossed in one automated exchange no human reviewed.

I went looking for anyone connecting these two facts, across the trade press, the legal briefings and every consultation position made public so far. I found nothing.

The narrow reading may well hold for the display stack of 2026. It cannot describe the personalising, transacting stack the standards bodies’ protocols are built to enable. The collision date is whenever that end of the stack reaches production scale.

Two more provisions land directly on operators. The obligation attaches to whoever arranged for the program to make or assist decisions, and the OAIC’s own example of arranging is procuring another entity’s AI system. Others in the chain may carry their own obligations, but the deployer cannot outsource the question: the OAIC expects the arranging entity to understand and oversee how third-party systems use ADM. And registered by that same date, though not yet confirmed to be in force on it, is the Children’s Online Privacy Code, still in draft, which proposes consent before children’s personal information is used for targeted advertising and reaches personalisation and recommender systems too. Its commencement date hasn’t been set, but 10 December is already the outer edge for both. If your properties or campaigns touch child-accessible services, that date is doing double duty.

Calibration, because precision matters more than alarm. Parts of adland can already feel the temperature change; TBWA’s chief AI and innovation officer read the week as compliance risk “shifting from soft to hard.” He is right about the direction. December’s obligation is transparency only. No notification duty, no right to human review. Western Australia’s new regime already requires both, which tells you the direction of travel, as does a national framework whose stated ambition is an enforceable licence. And the disclosure itself is categorical, the kinds of information used and the kinds of decisions made, a public map rather than a system-by-system register. But you cannot draw the map without the territory: writing the disclosure forces the inventory of every system that decides things about consumers using their data.

Which brings the ads.txt anatomy home. Under APP 1.7, you become the publisher. The privacy policy is your declaration file, drafted by you, published at your own domain, describing what your machines decide. And the OAIC holds the gate, with the power to audit the declaration against what your stack actually does. The other two rulebooks in this piece are being written for you. This one you write yourself, and you will be held to it.

Here is the call I will mark my own scorecard on. September’s guidance will sharpen the mechanics, the ‘substantially and directly related’ test, the treatment of human oversight, but it will not settle the significance threshold for advertising; that line gets drawn later, case by case. And when regulatory attention eventually reaches this industry, it will arrive through personalised pricing and eligibility, not media buying. If I’m wrong, you’ll read it here first.

Guidance lands around September, the interpretive question stays open until then, and this is a map rather than legal advice. Have privacy counsel across your position before December, with the inventory in hand.

What to sign off before the next planning cycle

Three moves. All fit inside a quarter. None requires a transformation budget.

Keep the agentic layer thin and swappable. The standards contest is unresolved and no convergence roadmap exists. An architecture bet placed entirely on one protocol family carries re-platforming risk if the other prevails or the two never merge. Abstraction costs a little now. Migration costs a lot later.

Make governance a procurement question, not just protocols. Ask every vendor and your agency, in writing: which standards are you building against, what happens if the other framework wins, and who bears the switching cost? Then go past the protocol layer to the authority layer: what spend limits, approval thresholds and revocation rights constrain your agents, and how do their systems make automated decisions about individuals? Under APP 1.7, that last answer feeds your disclosure. The paper trail is the point.

Run the ADM inventory before the OAIC guidance lands. Every system making automated decisions that touch consumers, the personal information feeding each one, and what your privacy policy currently says about them. The organisations that answer September’s guidance with an inventory will look very different from the ones that answer it with a scramble.

And when you put those questions to your vendors, I want to know what comes back. Reply to this email. I read everything, and the anonymised patterns will make their way into a future edition.

Three layers, one operator

Three layers, three sets of authors, one operator who has to live under whatever all of them produce.

The ANZ advantage is that this market is small enough to watch every layer at once. Anthropic data reported by Mi3 last week put Australia first among 121 countries on Claude usage per capita, at 6.4 times what population alone would predict. Adoption here is running ahead of nearly everywhere, and the rules are being written in three places simultaneously. You can see all of it from where you sit. Operators in larger markets cannot.

So keep the ads.txt lesson in your pocket as you watch. The schema author, the declaration maker and the gate holder are rarely the same hand, and the gate is where the power settles. At the national layer the gate is not yet built. At the industry layer it will belong to whichever registry the orchestrators, platforms and payment rails actually enforce, and that may end up being decided by the walled gardens and commerce ecosystems rather than either standards body. At your layer, the gate is the regulator, and it opens in December.

ads.txt was a few kilobytes, and it reorganised a market once the gate began enforcing it. I watched it happen from both sides. The documents being drafted now are considerably longer, and the market is considerably bigger.

Follow the control, not the conversation.

Sources

Government and regulatory

• Prime Minister Anthony Albanese, “AI in Australia’s Interests,” address at the University of Sydney, 15 July 2026

• OAIC, Automated Decision-Making Transparency: Issues Paper, 18 May 2026

• Explanatory Memorandum to the Privacy and Other Legislation Amendment Act 2024, paras 337 and 343, as quoted in the OAIC’s Issues Paper

• Western Australia’s Privacy and Responsible Information Sharing Act 2024, Principle 10, as cited in the OAIC’s Issues Paper

• OAIC, Children’s Online Privacy Code (exposure draft): must be registered by 10 December 2026; commencement date not yet confirmed

Industry and standards

• IAB Australia, public submission summary to the OAIC’s ADM consultation, July 2026

• IAB Tech Lab, Agent Registry launch, 1 March 2026

• Adform, announcement on ARTF support and AAMP 2.0 workflows, 8 July 2026

• IAB Australia’s AAMP pilot: announced at the AdTech & Ops Summit, Melbourne, and reiterated at the Data Summit, June 2026 (from my own attendance; I could not find independent published coverage to cite)

• AdCP documentation and the AAO’s registry infrastructure (as AAO Chapter Lead)

Trade and market reaction

• B&T, industry reaction to the national AI framework, including TBWA’s Lucio Ribeiro

• Mi3, Mumbrella, SBS and Inside Story coverage of the 15 July announcement

• Nine’s AI content agreement with Microsoft, and public reactions from ARIA and Man of Many to the government’s copyright commitments

• Anthropic usage data as reported by Mi3: Australia ranked first of 121 countries on Claude usage per capita

This is a working reference list, not a formal bibliography. Two items above are drawn from direct professional knowledge rather than a published source, and are marked as such.

“Hold on, where are the tokens?”

That was Vera Skocic, head of customer and strategy at Techtronic Industries, midway through planning next year’s SaaS budget, realising the line item does not exist yet. She has company. CBA’s Matt Comyn told an AFR conference in June that as reasoning, tools and context grow, “your token costs do not scale on a linear basis”, and that spending scrutiny will only tighten through 2026. Forrester’s Sam Higgins says as many as one in three Australian executives now raise it as a major concern. And the bills arrived offshore first: Uber reportedly burned through its 2026 AI coding budget by April, and UBS says around 60 per cent of enterprises are now throttling AI spend, a shift its own analysts read as discipline returning rather than retreat.

Mi3 pulled these threads together in a strong piece on Friday: token costs are migrating from the IT budget to the marketing P&L, nobody can forecast them, and the sensible advice is to find the meter before you rebuild the marketing machine around agentic AI.

Right advice. And in advertising, here is what you find when you go looking. The meter is not lost. It is held. It changes hands at every phase of the campaign lifecycle, and where you can see it, you can manage it. Where you cannot, someone else’s incentives are managing it for you.

This edition walks that lifecycle with the meter in view, and ends with two tools: a one-question test for whether anything sold to you as AI deserves the name, and one structural fix per phase that works for both sides of the partnership.

Follow the control, not the conversation.

Subscribe now

Four parties, one meter

Every token burned on your advertising is paid for by one of four parties.

1. Marketers using LLMs directly see the meter. It is an API bill or a seat licence, sitting in your cost base, arguing with your CFO.

2. Agencies building agentic delivery hold the meter, and under a fixed fee you inherit their incentives without ever seeing their bill.

3. Platforms running LLMs inside orchestration, the category last week’s edition mapped and the holdco builds beside it, show no meter at all: the cost is real, but it returns to you as take rate, margin, platform fee, or model quality.

4. Publishers pay the meter for the privilege of being queried. Being discoverable to buying agents is now a cost centre.

This is not a distant problem for this market. Australians use Claude at 6.4 times the rate their working-age population would predict, first among the 121 countries in Anthropic’s Economic Index, and up from 4.1 times just two months earlier. The meters here are spinning faster than anywhere.

Hold one distinction for the rest of the piece. Cost risk is temporary: an underpriced token bill gets repriced at the next renewal, and everyone moves on. Quality risk is immediate: the party paying the meter controls the architecture, and architecture offers savings the other side will never see. A cheaper model here. A truncated context window there. One less reasoning pass before the answer ships. Forrester’s researchers found quality, latency and cost degrade together as the context window fills. Mi3 documented the mechanism. The incentive structure is what this edition is about.

And be clear about who that structure serves: nobody. Holding a meter under a fixed fee is a burden as much as a power. It hands the holder a cost it cannot forecast and two private options, absorb the volatility or economise quietly, neither of which a good operator wants. This is also why tokens are different from every input agencies have ever bought on a client’s behalf. You meet the team. You never meet the model. A staffing shortcut is socially visible; a model shortcut is not. Which is why the fix in this edition is structural, not personal.

Six phases, three questions

Last week’s edition mapped the platforms now running the brief-to-proof loop: brief in, everything else autonomous. This week, walk the same loop with a meter in hand. Because “everything else, autonomous” has a second reading. Everything autonomous is everything metered.

For each phase, three questions. What work happens here? Who holds the meter? And what does holding it make them want to do? Two notes before we start. No public dataset measures token consumption by campaign phase, so what follows is a directional read, anchored to the research that exists on agent architectures. And this walkthrough follows the media loop: creative production, the other great token sink (a single Coca-Cola campaign reportedly consumed 70,000 prompts), runs alongside it and will get an edition of its own.

1. Brief and intake

Your team uses LLMs for brief development, research synthesis and audience work. You hold the meter, and you can see it, mostly.

Token intensity is modest, though it climbs quickly once deep-research agents enter brief development: multi-agent research is the same fifteen-times pattern you will meet again downstream. Quality risk is low, because you control the architecture.

The trap here is bundling. Brief work done inside SaaS tools with AI baked into the subscription has no meter you can see, and a flat-fee AI feature carries the same quiet-downgrade incentive as any platform. Otherwise this is the phase the Mi3 piece speaks to directly, and its advice applies as written. Find the meter and own the line item. T2 has already moved its AI costs out of the IT line into a standalone P&L line, with token limits that apply all the way up to the managing director. That is what control looks like when you hold the meter yourself.

Then the brief leaves your building, and the meter leaves with it.

2. Strategy and planning

Your agency’s agentic workflows plan communications, allocate channels and build scenarios. The agency holds the meter. Under a fixed fee, you inherit their incentive to economise.

This phase is far more token-hungry than it looks. Strategy work is almost pure context: loading briefs, market data, historical performance and audience research into models, then iterating scenarios against them. The research on agent architectures explains why exactly that kind of work is expensive: 56 to 72 per cent of an agent’s total tokens go to planning and context handling, not to executing anything. Strategy is the phase that is nothing but planning and context.

Which makes the quality risk here high and invisible. If token costs squeeze your agency’s margin this year, the fee conversation happens at renewal, in daylight. A quietly downgraded planning model never makes it onto the agenda. Not every agency hides this meter, to be fair: some already pass tokens through at cost, as a line item alongside every other cost. That is the behaviour to reward at the next review. And if you have in-housed this work, the meter sits on your own P&L, where you have traded the disclosure problem for the discipline one.

The plan goes to market, and the meter changes hands again.

3. Discovery and negotiation

This phase barely existed two years ago. It is now where the deepest thinking in the chain happens. Buyer agents query seller agents. Inventory is discovered and terms are negotiated, agent to agent. The standards defining how this works are being written right now (more on those below).

This is the most intense token burst in the chain, per episode. Anthropic’s own engineering research puts a single agent at roughly four times the tokens of a chat interaction, and multi-agent systems at around fifteen times. Negotiation is multi-agent by definition. It is a burst rather than a burn: it runs for hours or days, where the optimisation loops downstream run for weeks.

And the meter is split. Your agents burn your side’s tokens. The seller’s agents burn theirs. Which creates something advertising has not priced yet: reasoning asymmetry. Each side economises on its own agents, and the side running deeper reasoning into the negotiation wins it. Token budgets become bargaining power. I have no study to cite for that claim, because nobody has run one. But the logic only needs one renewal cycle to start showing up in outcomes.

Publishers sit at the other end of every one of these exchanges. Each query a buying agent sends costs the publisher tokens to answer. Discoverability now has a unit cost.

Notice the slope, three phases in. The meter has moved from your own P&L, to your agency’s, to an exchange between machines that nobody in your building will ever watch. Each phase deeper, the meter gets harder to see. It keeps going.

4. Activation and buying

The transaction itself. Orders placed, deals executed. The phase everyone assumes is the expensive one, because it is where the money moves.

It is the cheapest token event in the chain. The architecture research found actual tool execution “constitutes a negligible fraction of the total cost across all configurations”. Once the negotiation is done, the buy is a single instruction from one agent to another, and the new standards work on agentic real-time bidding is engineered to keep it that way, because per-impression decisioning cannot afford frontier-model calls.

The intelligence in your buying was spent, or skimped, upstream. This section is short because that is the point.

The campaign goes live, and the meter disappears entirely.

5. In-flight optimisation

Two different layers hide under this one label, and they have opposite token profiles.

Per-impression decisioning happens on two paths: the live auction, and serve-time matching of pre-negotiated packages against each impression as it happens, across web, CTV, AI assistants and retail media. Most of this layer is not new at all. It is the machine learning that has priced bids and paced budgets for a decade, plus newer compiled and distilled logic where agentic systems feed it. No LLM sits in a sub-100-millisecond loop across billions of impressions; the economics forbid it. Tokens per event: effectively nil, and the compute sits with the publisher and platform, not with you. But the intelligence is frozen at whatever quality was built in, and you have no way to inspect it.

Campaign-level optimisation is where the burn lives. LLM agents monitor delivery, evaluate performance, adjust strategy, and loop. Every turn of the loop re-feeds the context window. Forrester calls it context debt, billed by the token. The research on self-refinement adds a warning of its own: without stopping rules, “make this better” is an open-ended cost instruction. Negotiation was the peak burst; run for a full flight, this loop is the largest cumulative burn in the chain.

The meter at both layers belongs to the platform, and it is fully invisible to you, unless you run the loop yourself. An in-housed team operating its own optimisation agents holds this meter directly, and that control is much of what the in-housing decision was always about. For everyone else, this is the strongest degradation incentive in the lifecycle. A continuous background token spend the platform must recover, applied to decisions you cannot audit. If your optimisation loop moves from a frontier model to a distilled one on Tuesday, nothing in Wednesday’s reporting will tell you.

The loop finally turns toward proof, and the meter follows it into the reporting.

6. Measurement and reporting

Two layers again, and the token test separates them.

The econometrics itself, the mix models and the incrementality maths, is statistical compute, not token flow. It was never AI in the fashionable sense and does not need to be. What has changed is the layer on top: measurement agents that ingest and standardise the data, configure and validate the models, then sit in the room answering questions about what drove growth and where the next dollar should go. Mutinex’s MAITE is the local example, an agent trained on the client’s own mix model. Mi3 reports Lendi reallocated around 20 per cent of its search spend on its analysis.

That is not report formatting. It is analytical reasoning over large contexts, it loops, and it moves real budgets. Token intensity runs high, and even the polish is expensive: in off-the-shelf agent clients, 75 to 85 per cent of latency concentrates in final answer synthesis.

The meter sits with the measurement vendor, inside the subscription. But notice the structural difference from everything upstream. An independent measurement vendor’s reasoning quality is its entire product, so its incentive points toward spending tokens well, not sparingly. That matters, because the alternative narrator is the party that spent your tokens. Independent measurement agents are how ask six, below, stops being a demand and becomes a purchase.

The inversion

Last week’s edition ended in a debrief room. Six weeks after the campaign, someone asks what the agent actually did in week three, when the system shifted budget across channels, and nobody can answer with precision.

Run the sequel. Same room, same campaign, one question deeper: which model made the week-three call, and what was it allowed to spend thinking about it? A frontier model reasoning across your full delivery data, or a distilled one running on a truncated context window because someone’s cost target tightened in March? The report does not say. The invoice does not exist. The silence in that room is the meter you cannot see.

Now notice what the walkthrough did to your intuitions. Buyers worry most about the phases where AI costs are visible: their own usage, their agency’s fees. They worry least about the phases where no line item appears. The framework says that is exactly backwards. The invisible-meter phases, negotiation and in-flight optimisation, are where token intensity runs highest, where the incentive to degrade is strongest, and where no audit path exists.

No line item does not mean no cost. It means the cost reaches you through a channel you cannot see: margin, take rate, or the quality of the model deciding where your money goes.

It also reprices last week’s fork. Build the orchestration capability and you hold the meter, with everything that control costs and buys you. Buy it, and the meter sits inside the platform fee. That is bucket three, and everything this edition says about bucket three now applies to the vendor decision.

The token test

One more distinction before the asks, and it will save you money this year.

Most of what platforms will sell you as AI in 2026 is not new. Bid optimisation, pacing, budget allocation, audience expansion, click prediction: that is machine learning, it has run inside every major platform for a decade, and it is already priced into your fees. It is AI in the textbook sense. It did not become something else because the industry found a more fashionable word for it.

The tell is the token. Language models meter their work in tokens. The ML that has always run your campaigns does not. So when a platform presents an AI-powered anything, ask one question: what in this product consumes tokens, and who pays for them?

If the answer is nothing, you are looking at last decade’s machine learning wearing this year’s name. It works. It always did. You should not pay a new premium for it, and it belongs in none of the phases above.

If the answer is a genuine token flow, you are in this edition’s territory. Find whose meter it runs on, and ask what holding it makes them want to do.

What to do with this

One structural fix per phase. None of them need a data science team, and all of them work better agreed between partners than demanded across a table. The point is not to catch anyone out. It is to stop fixed fees forcing good operators into quiet choices.

1. Brief and intake. Put tokens on their own P&L line. You cannot manage what sits inside a SaaS bundle.

2. Strategy and planning. A model mix schedule in the agreement. Agencies already disclose staffing mix in every fee proposal: senior strategist hours here, coordinator hours there. The same schedule for machine labour serves both sides. What share of the work runs on frontier models, what runs on cheaper or open-weight alternatives, which tasks get which, and notification when the mix changes. To be clear, a mixed fleet is good practice, not a red flag. Omnicom calls it rocket ships versus bicycles, and nobody needs a frontier model to reformat a report. The point is not frontier everywhere. It is a mix that is declared, fit for purpose, and stable without notice, which also lets the agency price its token risk instead of burying it.

3. Discovery and negotiation. Agree what reasoning budget your agents carry into a negotiation relative to the sell side. Today, almost nobody in the industry can answer that question. The partnerships that can will win negotiations neither side ever watches.

4. Activation. Nothing, except the token test. If nothing in the product consumes tokens, it is the ML you were already paying for. Execution is the cheap part.

5. In-flight optimisation. Agreed right-to-know terms covering model and configuration changes on optimisation loops, plus documented stopping rules on refinement. Last week’s edition argued the governance log is the source of authority on what the agent did. Add two columns to that log: which model made each decision, and what token budget it ran on. The same two columns protect the platform when the question comes, and the client when it doesn’t.

6. Measurement. Separation. Whoever spends the tokens should not be the sole narrator of whether they were well spent. Independent measurement agents now make that separation buyable rather than aspirational.

And one prediction to hold me to. Within two renewal cycles, model mix and token disclosure schedules will be standard in ANZ agency agreements, tabled as often by agencies as requested by clients. The current arrangement serves neither side: under fixed fees the meter sits inside the margin, which means the agency carries a cost it cannot price and the client carries a quality risk it cannot see. Disclosure converts a hidden conflict into a priced service. The operators who move first, on either side of the table, set the standard everyone else signs.

A note for readers tracking the standards. The infrastructure described above is being formalised now. AdCP, from AgenticAdvertising.org, defines the agent workflow from discovery through to reporting, including serve-time package matching through its Trusted Match Protocol. IAB Tech Lab’s AAMP covers agent governance and identity, and its ARTF specifies the real-time execution layer alongside OpenRTB. Each deserves its own edition, and each will get one.

The meter moves

The Mi3 piece was right that the bill is arriving, right that nobody can forecast it, and right to send buyers looking for the meter. For your own usage, that search is enough.

For the rest of the chain, the meter is held by someone else, and the question changes. Not where is it, but who holds it, what does it cost them, and what they will quietly do to your outcomes to make that cost go away.

The meter moves. Control follows it.

Follow the control, not the conversation.

References

Mi3, token economics feature, 10 July 2026. https://www.mi-3.com.au/04-07-2026/token-shock-agentic-ai-burns-spend-deciding-what-do-which-tools-inspect-and-how-explain#login-nd

TechCrunch, “The token bill comes due: inside the industry scramble to manage AI’s runaway costs”, 5 June 2026. https://techcrunch.com/2026/06/05/the-token-bill-comes-due-inside-the-industry-scramble-to-manage-ais-runaway-costs/ (Uber, Microsoft and Cursor reporting)

UBS equities research note on enterprise AI spend throttling (Keirstead, Arcuri, McGinnis), as reported by Mi3.

Mi3, “Australia ranks first of 121 countries in Anthropic usage index”, 7 July 2026. https://www.mi-3.com.au/07-07-2026/australia-ranks-first-our-121-countries-anthropic-usage-index-64x-expected-rate-less

Anthropic Economic Index, including “How Australia uses Claude”. https://www.anthropic.com/economic-index and https://www.anthropic.com/research/how-australia-uses-claude

Mi3 podcast with T2’s Peter Randeria and Omnicom’s Alex Pearcy, 22 June 2026. https://www.mi-3.com.au/22-06-2026/dirty-little-secret-operating-model-agentic-ai-enters-its-cambrian-age-focus-shifts

Mi3 interview with Techtronic Industries’ Vera Skocic, 2 June 2026. https://www.mi-3.com.au/02-06-2026/zero-readiness-ai-search-retail-media-and-product-intelligence-reshape

Reuters via Yahoo Finance, “Australia’s CBA flags surging AI costs as tasks grow complex”. https://finance.yahoo.com/sectors/technology/articles/australias-cba-flags-surging-ai-041848066.html

Mi3, “MMM grows: Lendi shifts 20% of search spend and banks 40% ROI lift with Mutinex’s agentic-infused econometrics”, 18 March 2026. https://www.mi-3.com.au/18-03-2026/mmm-grows-lendi-shifts-20-search-spend-and-banks-40-roi-lift-mutinexs-agentic-infused

Digiday, “Agencies grapple with the economics of a new marketing currency: the AI token”. https://digiday.com/marketing/agencies-grapple-with-economics-of-a-new-marketing-currency-the-ai-token/ (agency token pass-through practice; Coca-Cola 70,000-prompt campaign, per its makers)

ProMCP: Profiling Token Flows and Latency Costs in Model Context Protocol-Based LLM Agents, Findings of ACL 2026. https://aclanthology.org/2026.findings-acl.1967.pdf

Forrester, “Your AI Bill is a Context Problem”. https://www.forrester.com/blogs/your-ai-bill-is-a-context-problem/

Anthropic engineering, “How we built our multi-agent research system”, June 2025 (single agent ~4x chat tokens; multi-agent systems ~15x, verbatim from the post). https://www.anthropic.com/engineering/multi-agent-research-system

Self-Refine: Iterative Refinement with Self-Feedback. https://arxiv.org/abs/2303.17651

Kearney, “AI is Everywhere Except in the Returns” (Sengar, Gupta, Livolsi). https://www.kearney.com/service/digital-analytics/article/ai-is-everywhere-except-in-the-returns

AdCP documentation, AgenticAdvertising.org, including the Trusted Match Protocol. https://docs.adcontextprotocol.org

IAB Tech Lab, AAMP (Agentic Advertising Management Protocols). https://iabtechlab.com/standards/aamp-agentic-advertising-management-protocols/

IAB Tech Lab, ARTF v1.0 announcement. https://iabtechlab.com/press-releases/iab-tech-lab-announces-agentic-rtb-framework-artf-v1-0-for-public-comment/

The claim lands differently depending on where you sit. If you are a holding company, it confirms an investment thesis already underway. If you are not, it finds you in a more familiar place: a reporting pack assembled across five platforms, each with its own attribution model, none talking to the others. A brief written last quarter for a world the stack is no longer executing in. A client asking whether the budget is working, and an honest answer that requires more manual consolidation than the original campaign. It names something you do not yet have a word for, and a decision you have not yet made.

What the advertising industry has been describing for twenty years without the infrastructure to deliver is now becoming real. The open standards that let agents buy, optimise, and report across the stack are live.

And into the gap, a new platform category is arriving. Not to close it entirely. The distance between what the open internet now makes possible and what the walled gardens still refuse to allow is the most important structural fact in what follows.

Subscribe now

What the category actually looks like

The platforms in this category are not DSPs with better dashboards. They are not campaign management tools with an AI layer bolted on. They occupy a distinct position above the existing buying infrastructure, connecting to DSPs, social platforms, publishers, and measurement partners through the protocol stack above, and managing the full loop from brief to outcome without requiring a human operator inside any downstream system.

Five funded platforms are competing for this starting position. They have not converged on a single design, which is the first thing worth understanding about this category: these are not versions of the same bet. They disagree on where the leverage in the loop sits.

Veylan launched in June 2025. Randall Rothenberg, the former CEO of the IAB, and John Frelinghuysen, former Bain partner and Disney/ABC media executive, published a piece in WARC in September 2025 coining the category label MAMMA, Marketing-Advertising-Media-Measurement-Analytics, and naming Veylan as one of its earliest examples. Their central argument: measurement has been an orphan dataset, disconnected from campaign execution, for 75 years. Veylan is designed to close that gap. Its design runs media buying and optimisation in the same workflow as strategy and creative, inside a single command centre. Anthony Iacovone, Veylan’s co-founder and CEO, who previously founded the DSP AdTheorent, describes the model:

PadSquad, an independent creative agency, reported 70 percent cost reductions and 10x faster campaign execution in early adoption.

Kana is the same thesis from a different direction: the human stays for approvals, not for the work.

Pencil, owned by Brandtech Group, is the holdco-scale version of the same idea: an Agentic Marketing OS that brings audience intelligence directly into the creative workflow and ships through to distribution. Oliver in Latin America built the entire operating model for Samsung’s Mobile Experience division around it. Not a pilot. A production operating model.

Gimmefy operates differently. Its model is white-label infrastructure: agencies license the platform and deploy it under their own brand, with 48 frontier models, brand memory that loads automatically into every output, multi-agent campaign playbooks, and governance across unlimited users. DBS, Singtel, Unilever, and Volvo are among the brands on it directly. Frontier Media, one of its agency partners, went furthest: productising the infrastructure as FrontierOS, their own branded operating system, with Gimmefy invisible underneath. That is not a capability tool. It is an agency product built on someone else’s infrastructure, sold to clients as their own.

Plurio, narrower in scope, processed $20 million in ad spend across EdTech and FinTech pilots and reported more than 20 percent lower customer acquisition costs. A performance component rather than a full stack, and further evidence that the category has more than one funded position.

Of the five, brief-native and white-label infrastructure are the most defensible long-term positions. The unified command centre competes directly with what holdcos are building internally. If that market consolidates the way DSPs did, the independent platforms will either find a niche the holdcos will not serve or be acquired by the ones that can. Whoever locks in the architecture before the market settles sets the terms for everyone that follows. The category is funded, in the hands of real clients, and has not yet found its form.

The divide

The holdco version of this orchestration layer already exists. It has been under construction for three years.

Publicis acquired LiveRamp for $2.17 billion. Not a data deal. A stake in the identity infrastructure that connects first-party data to the programmatic ecosystem. Omnicom, having absorbed IPG, is consolidating around Acxiom’s 2.6 billion verified identity dataset. Stagwell was the first holdco to integrate an agentic buying layer into its operating model. Horizon Media, the largest independent media agency globally, built its own orchestration layer from the ground up, explicitly against vendor lock-in.

The architecture of each of these bets is the same: a proprietary data layer, an AI orchestration capability running on top of it, and the brief entering the system with the agent managing everything downstream. The holdco is not describing this as a product it sells. It is describing it as an operating model, built for themselves and extended to the brands they work with.

For those brands, the orchestration layer arrives embedded in the agency relationship. The brand does not choose which platform runs its campaigns. The holdco’s infrastructure is the platform. The benefit is real: the engineering investment scales across every client on the roster. So does the dependency.

Independent agencies, in-housed brands, and mid-tier brands arrive at the same orchestrated future without the same infrastructure, and without the relationship that carries the holdco’s investment on their behalf. I have had the same conversation three times in the past few months, with three separate independent agencies, each at a different scale, each looking at the same structural problem from a different angle.

One was navigating a client relationship where reporting across a buy split between three DSPs and two retail media networks was consuming more internal hours than the strategy it was supposed to prove. No single view. No common attribution logic. Manual consolidation every week. The question the client was actually asking, whether the budget was working, was unanswerable with the infrastructure in place.

The second was a brief architecture problem. The brief was written for reach and frequency. The agents executing the plan were optimising against engagement signals. The gap between what the client had commissioned and what the platform was actually doing was invisible in the reporting layer, because the reporting surfaced the agent’s output but not its logic.

The third was running lean enough that the stack itself had become the overhead. More platforms than the team could actively manage. More decisions happening downstream of any human review than anyone was comfortable acknowledging.

Each was describing, without naming it, the problem the orchestration layer exists to solve. Each was arriving at the fork without a map.

If you did not choose the platform that manages your brief-to-measurement loop, something else is choosing it for you. That is not a warning about the future. It is a description of what is already happening to the agencies and brands operating without deliberate infrastructure decisions while the holdcos close the gap.

The fork

The choice facing independent agencies, in-housed brands, and mid-tier brands is not whether to engage with the orchestration layer. The layer is arriving whether they engage with it or not: through the infrastructure their holdco-aligned competitors are already operating on, through the vendor products their clients will eventually ask about, through the protocol stack that does not pause for market readiness.

The choice is whether to build the orchestration capability or buy it from the category now emerging.

Building is not a theoretical option. Gimmefy’s white-label model exists precisely because it is viable at agency scale: take the infrastructure, wrap a proprietary brand and methodology around it, and go to market with a platform that belongs to you commercially. Horizon’s Blu demonstrates what a larger-scale build looks like for a well-capitalised independent. The investment is real. The return on it is control: of the data layer, of the orchestration logic, of the measurement framework, and of the commercial relationship the platform enables with clients.

Buying from the vendor category is faster and cheaper upfront. Veylan is in market. Kana is live with a martech pedigree behind it. The case studies exist. The capital is behind the platforms. An independent agency or mid-tier brand that adopts one of these platforms gets the orchestration capability the holdcos built for themselves, at a price point accessible outside the holdco model. The tradeoff is unchanged from any platform adoption decision: operational speed now, dependency on the vendor’s roadmap later.

In-housed brands face a specific version of this tension worth naming separately. The decision to bring media in-house was a decision about control: of the data, of the buying decisions, of the attribution. An in-housed brand that now adopts an external orchestration platform is making a second decision about where that control sits. If the orchestration platform is what receives the brief, manages the buy, and produces the measurement output, then the control the in-housing decision was designed to secure has migrated to the platform vendor. That is not necessarily wrong. It is a decision that needs to be made explicitly, not by default, and with specific terms negotiated before the platform goes live rather than after.

Mid-tier brands sit in the clearest exposed position. Too large to be indifferent to the orchestration gap. Too small to carry the infrastructure investment internally, or to have holdco engineering resources closing the gap on their behalf. The vendor category is the realistic path. The question is which vendor, and what the commercial terms mean for who owns the governance log the platform generates.

Who owns the system

Here is the moment the question becomes concrete. A post-campaign debrief, six weeks after the campaign ran. The platform has produced a report: reach, attributed conversions, cost per acquisition, all within the targets set at brief. The numbers look right. Then someone asks what the agent actually did in week three, when performance softened and the system automatically shifted budget away from two channels and toward a third. Nobody in the room can answer with precision. The platform reports the outcome. Not the reasoning. Not what signal the agent weighted, what it evaluated and rejected, or what the walled garden’s API returned that drove the reallocation. The log that would answer the question lives somewhere inside the platform. Whether the brand has the right to access it, and whether it covers walled garden spend as well as open web, are questions nobody asked before the campaign went live.

For open internet buying, the orchestration platform produces an audit trail. Every decision the agent made, every signal it evaluated, every allocation it chose, visible and auditable through the platform layer. That trail is the proof. Not the DSP’s reported metric. Not the publisher’s own attribution. The decision log, end to end.

The governance log is not a compliance artefact. It is the source of authority in every conversation about whether the campaign worked, whether the budget was justified, and whether the agency or the brand gets credit for the outcome. Whoever generates that log defines the terms of that conversation. This is the commercial consequence of the infrastructure decision, and it is rarely named as such in any discussion of platform adoption.

A brand with access to that log is in a fundamentally different position when a client, a regulator, or an internal stakeholder asks what the agent actually did and why.

For walled garden buying, which for most brands remains the majority of digital spend, the proof the orchestration platform can surface is the metric the walled garden chooses to report. Google reports its own conversions. Meta reports its own. Amazon reports its own. Pulling those numbers into a single interface does not create cross-platform measurement. It creates a single interface for three separate attribution models that may be counting the same consumer multiple times across each.

Holdcos can absorb this gap because they have the data infrastructure, LiveRamp in Publicis’s case, to bring incrementality testing and media mix modelling (the statistical methods that measure walled garden spend without relying on platform-reported figures) to the walled garden question. Everyone else is paying for a measurement gap the walled gardens created deliberately and the orchestration layer cannot close. That is not a temporary limitation. It is the permanent commercial advantage of the platforms that refused to join the open protocol stack, and it will hold for as long as Google, Meta, and Amazon choose not to move.

Independent agencies, in-housed brands, and mid-tier brands choosing an orchestration platform are making a measurement decision at the same moment as an operational one. A platform that can only surface walled garden reported metrics is a better interface. It is not a measurement solution. The two are priced the same and worth very different things.

The question to ask of any orchestration platform before committing: what does your audit trail cover, and where does it stop? For open internet, the answer should be complete. For walled gardens, every honest answer will name the API ceiling. The gap between the two is not a vendor failure. It is the measurement bifurcation the walled gardens have built into their business models for twenty years, arriving at a new layer of the stack.

What this means in ANZ

The orchestration platform category does not yet have an ANZ entrant. Veylan and Kana are US-founded. Pencil is London and New York. Gimmefy is Singapore-based. Plurio is San Francisco.

The ANZ independent agency market, faces the same structural divide with a different set of available tools and a narrower window to act before the gap between them and holdco agencies becomes permanent. Holdco agencies in ANZ are deploying global AI platforms through their parent companies. Publicis, through Atomic 212 as its CoreAI vehicle, is furthest along. The others are at earlier stages, with the same underlying infrastructure advantage. The gap between what a holdco-aligned client is now operating on and what an independent-agency client has access to is widening at a pace that is not visible until it is already concrete.

In-housed brands in ANZ face the specific tension named above. A number of the most sophisticated marketing organisations in this market made deliberate decisions to bring buying in-house and build first-party data infrastructure. The orchestration layer was not part of the original rationale. It now is. An in-housed brand in ANZ looking at a Veylan or a Kana is not just evaluating a platform. It is revisiting the original in-housing thesis and deciding whether the control it was designed to provide extends to this layer or ends before it.

The mid-tier brand market has the fewest options and the least leverage in negotiating them. That is worth stating plainly.

The risk here is infrastructure tenancy: running your campaigns on a platform you had no hand in designing, on terms someone else can change. The brands and agencies that choose who builds that infrastructure, and on what terms, are in a different position from the ones that find it already in place when they arrive. IAB Australia’s Standards Council has been piloting AAMP locally from June 2026. That is the most relevant current mechanism for local operators to influence what the infrastructure looks like before it hardens.

The closing frame

Here is the question underneath all of this: what platform does your brief-to-proof loop run on? Who built it? Who owns the logic that takes the brief and produces the measurement output?

The orchestration layer is being built. The holdcos have built theirs. The vendor category is building for everyone else. The window for being deliberate about this decision, rather than inheriting someone else’s, is open. It is narrowing.

The right architecture is not yet settled. The platforms are a year old, the protocols newer still, and nobody has a complete picture of what the build-versus-buy decision means three years out. That uncertainty is real. What is not uncertain is that the question needs to be on the table now.

The brands and agencies navigating this well right now are not the ones with the answers. They are the ones that have put the question on the table. In procurement conversations. In agency briefs. In conversations with people who are tracking what this infrastructure actually costs, what it controls, and what the commercial terms mean before the platform is live rather than after. That is the difference between arriving at this with leverage and arriving without it.

If you did not choose the platform, something else is choosing it for you.

The complete series, The Frame, The Stack, The Brands, The Agencies, The Publishers, The Measurement, The ANZ View, is available at systemsthatdecide.substack.com. Three companion playbooks, one each for brands, agencies, and publishers, turn this series into a 90-day action plan for ANZ operators. They are available now.

Most of what’s been written about this shift is generic, US-centric, or written by someone with a platform to sell. Almost every consultancy advising on this has platform relationships of its own. Almost every agency earns margin on what it recommends. I wanted something built for the actual seat you sit in, with ANZ’s regulatory and market conditions built in, and no vendor relationship behind a single page of it.

So today I’m publishing three of them: The Brand Operator’s Briefing, The Agency Operator’s Briefing, and The Publisher’s Briefing. Free. Four modules each. Written for one seat, not three.

Download Your Briefing Here.

If you’re on the brand side, your stake is budget authority. Agentic systems are quietly moving decisions that used to sit with you, bidding logic, audience construction, increasingly creative selection, into platforms you can’t fully see inside. The brand briefing covers what changes in your agency contract, what your CFO will start asking once attribution stops being a clean answer, and a 90-day audit you can run yourself.

If you’re at an agency, your stake is the fee model. Either your agency writes the answer to what agentic buying means for your clients, or procurement writes it for you. The agency briefing covers what moves from activities to outcomes in your scope of work, who owns the guardrails when an agent executes the brief, and the governance conversation your biggest client is about to start having without you in the room.

If you’re at a publisher, your stake is the signal. Buying agents don’t browse your media kit, they read structured signals. If your inventory doesn’t expose what an agent is built to evaluate, you don’t lose the pitch, you never get invited to it. The publisher briefing covers what changes in floor price logic, what advertisers will start demanding in transparency, and the questions to put to your SSP partners now.

All three lead with the same operation underneath the surface. ARTF, AdCP, MCP and A2A, the standards defining how agentic systems actually work, and neither has won yet. Whichever your vendors build to decides what data you can see and what you’re able to ask for in an audit. IAB Australia started piloting ARTF this same month. This isn’t a future-tense argument. None of it is vendor-shaped either way, I wrote all three of these myself, with no platform relationship behind a single page.

Get the briefing for your seat below. tell it which one you are, it’ll send the right one.

Download Your Briefing Here.

If something in it maps to a decision actually sitting on your desk, the next step is a 30-minute scoping call, not a sales pitch. One business day to reply.

Andrew

On 27 April 2026, PubMatic confirmed that its AgenticOS platform was running live campaigns in Australia. It had been running in the United States since January. France and the Netherlands since February. Australia joined four months after the global launch, ahead of most of Europe and the rest of Asia-Pacific outside Singapore.

Global platform launches typically reach Australia eighteen months after the US, six months after the UK, three months after the announcement that “APAC is a priority.” By that standard, this is fast. Meaningfully fast. Fast enough that calling Australia a laggard in the agentic buying transition would be wrong.

But being an early recipient of global infrastructure is not the same as having a seat at the table where that infrastructure was designed. The standards governing how buying agents communicate, the protocols defining what data flows between systems, the governance frameworks specifying who is accountable when an agent makes a wrong decision. These are being written in San Francisco and New York and London. They are being written by the platforms, the holdcos, and the standards bodies that have the resources to participate in the working groups.

Australia is receiving the outputs. It is not shaping the inputs. Call it infrastructure tenancy: occupying a building you had no hand in designing, on a lease whose terms someone else can change. That distinction matters more than the deployment timing.

What is actually live

The honest account of agentic buying in ANZ starts with separating what the global platforms have deployed locally from what local market participants have built.

On the global platform side, the picture is materially further along than most practitioners in this market appreciate. It runs across four layers: the demand and supply-side platforms themselves, the ad-ops infrastructure underneath them, the standards bodies responding to both, and the retail media networks running their own AI investment in parallel.

PubMatic AgenticOS is confirmed live in Australia with active campaigns. Yahoo DSP’s “Yours, Mine, Ours” framework launched in Australia and Singapore at general availability in January 2026. Not a pilot, not a regional rollout following a delayed approval process. General availability, from day one. Amazon, Google, and The Trade Desk have all reached Australian advertisers on a similar timeline: Amazon’s MCP server since February 2026 and Ads Agent since June, Google’s DV360 Gemini integration since March, The Trade Desk’s Koa Agents in alpha with local access through partner agencies.

Ad-ops infrastructure is following the same pattern, one layer down from the DSPs and SSPs above. Mediaocean launched NIVO AI globally on 15 June 2026, an agentic layer built on Innovid’s ad-serving and measurement systems that automates campaign trafficking, QA, and reporting. The next day, Innovid’s APAC general manager, Georgia Brammer, was in Australian trade press speaking directly to the structure of local teams, lean operations where the same people own strategy and manual setup, with activation often split across offshore and global ops. The figure she cited was more specific than the global “90 percent” headline claim: a live test that cut one trafficking and QA workflow from an hour and forty minutes to forty-five. Global launch to Australian trade press commentary inside 24 hours is itself the data point worth noting.

Standards bodies have not been idle either. IAB Australia formally adopted the IAB Tech Lab roadmap in early 2026 and is piloting AAMP locally from June 2026 through its Standards Council. The Australian internet advertising market reached AUD 18.4 billion in calendar 2025, up eleven-point-five percent year on year. In absolute dollar terms, that is a market large enough to attract meaningful platform investment.

The retail media layer is further advanced than most discussions of agentic buying acknowledge. Cartology, Woolworths’ retail media network, generates approximately AUD 750 million in annual revenue, growing at nearly twenty percent per year. Coles 360 grew more than twenty percent in the 2025 financial year; growth moderated to ten-point-three percent in the half-year result reported in February 2026, an early sign that supermarket retail media growth in Australia is no longer purely additive. Combined, the two supermarket retail media networks are approaching one billion dollars in Australian ad revenue. Woolworths was first in the Asia-Pacific region to deploy Google’s Gemini Enterprise, announced in January 2026 for its Olive customer assistant, with the agentic features rolling out through the year. Coles signed with OpenAI in 2025 and established a data clean room arrangement in January 2026. Bunnings Buddy, showcased at Google Cloud Next 2026 in April, is Australia’s first consumer-facing agentic shopping assistant. The duopoly dynamic in Australian grocery retail is also a duopoly dynamic in Australian retail media. An agent that cannot transact at Coles or Woolworths cannot execute an Australian grocery brief. That is a structural fact the global agentic buying architecture was not designed around.

What local market participants have built

On the local market side, the picture is different.

I went looking for seller agents already live in the Australian market for this series, a crawl of publisher and brand websites, no announcement required, just evidence on the page. I expected to find a handful of quiet experiments. What turned up was one. Yahoo’s, detected in late April 2026. Yahoo was a founding member of AdCP, and its local seller-agent capability arrived on the same general-availability timeline as its DSP framework. That is not a coincidence. The companies building the standards are also the companies shipping them first, in every market, including this one.

Everything else on the local side is absence, and on both sides of the market, it is absence of the category itself. There is no Australian or New Zealand-founded DSP or SSP operating at meaningful scale. The IAB Tech Lab Agent Registry, which reached ten entries by March 2026, has none from this region. No major Australian broadcaster or publisher, not Nine, not Seven, not News Corp, not Foxtel, has publicly announced a seller agent capability or an AdCP-compatible inventory declaration.

The Seven West Media and SCA merger, completed in January 2026, concentrated national television, BVOD, radio, and audio into a single commercial entity without, by any public signal, an agentic infrastructure investment to match. That is not a criticism of a merger that had clear commercial logic in a difficult market. It is an observation about where the combined entity sits as agentic buying builds. A major seller without a seller agent, in a market where buyer agents are arriving, is negotiating from a structural disadvantage that relationships and bundling cannot fully offset.

Supply is consolidating from the demand side too, just not in Australia. Publicis’s USD 2.167 billion acquisition of LiveRamp, announced in May 2026, puts an agency holding company in control of the identity-resolution and clean-room infrastructure that Australian brands, including Coles, already rely on for retail media targeting. Neither consolidation was designed with an ANZ seller agent or an ANZ buyer agent in mind. Both will shape what Australian buyers and sellers can do with the data they hold.

The independent agency market, Hatched, Half Dome, This Is Flow, JOY, Ryvalmedia, has no public agentic buying pilot comparable to the Butler/Till credentialing campaign in the United States. The holdco agencies are deploying global AI platforms locally; Publicis is furthest along, through Atomic 212 as its CoreAI vehicle after the January 2025 acquisition. The others are at earlier stages.

REA Group, through its Gemini and OpenAI integrations, is the most publicly advanced local publisher on AI investment. That investment is principally on the consumer-facing product side. The extension to an advertising seller-agent capability is a logical next step for which there is no public announcement.

The regulatory threads nobody is designing for

The most differentiated part of the ANZ agentic buying story is not the infrastructure. It is the regulatory environment.

Here is the fact that should reorganise how every brand and agency in this market plans for the second half of 2026: two of the three regulatory threads converging on agentic buying in Australia commence on the same day, 10 December 2026. Not a compliance season. A single date. That is either useful clarity or a sign of how compressed this is about to get. Probably both.

The first thread is the Privacy Act automated decision-making transparency obligations, in effect from that date, creating explanation rights for individuals affected by automated decisions. A buying agent making decisions about which audiences to target, which content contexts to serve against, and which individuals to reach or exclude is making automated decisions at scale. The specific application of the ADM obligations to advertising agents is not fully settled. The direction is clear. A brand whose buying agent cannot produce a decision-level explanation of why a specific targeting decision was made will be in a poor position when the obligations are tested. My view: treat agent-executed targeting as in scope now. Waiting for the regulator to confirm the boundary costs more than assuming the conservative reading too early.

The second thread is the ACMA AI-disclosure rule for commercial radio, mandatory from 1 July 2026, earlier and separate from the December collision. AI-generated or AI-assisted content in commercial radio broadcasting must be disclosed. The extension of this principle to AI-agent-executed advertising decisions is not yet determined, specifically, whether an ad buy executed by an agent requires disclosure. But the regulatory appetite is evident. My read here too: assume disclosure requirements will extend to agent-executed buys before the regulator says so explicitly. The downside of disclosing early is minor. The downside of being the test case is not.

The third thread is the Children’s Online Privacy Code, mandated by the same 10 December date as the ADM obligations. An agent briefed to reach a broad demographic and optimising against engagement signals, including signals that skew toward younger audiences, could make targeting decisions that breach the code without the brand having made an explicit decision to reach children. The brief architecture question and the governance question are the same question under this framework.

None of the six major AI research tools consulted in the preparation of this series covered these regulatory threads in depth. They are among the most consequential local market dynamics, and they are almost entirely absent from the global conversation about agentic buying. That absence is itself a signal. The platforms building this infrastructure are not designing for the Australian regulatory environment. Someone local will need to be.

The risk worth naming

There is a version of the ANZ agentic buying story that is optimistic in a shallow way. Australia is further along than most of EMEA. The infrastructure is live. The retail media layer is sophisticated. Local publishers are experimenting. The market is large enough to attract platform investment.

That version is accurate and insufficient.

The risk is not that agentic buying arrives slowly in Australia. The risk is that it arrives as a product Australia consumes rather than a system Australia shapes. That the standards are written elsewhere, the protocols are set elsewhere, the governance frameworks are determined elsewhere, and Australian operators implement what they are given, at the speed at which global platforms choose to deploy, under the terms global platforms choose to offer, accountable to the regulatory frameworks global platforms prefer.

This is not a hypothetical. It is the current trajectory.

The IAB Tech Lab Agent Registry has no ANZ entries. No Australian company holds founding-member status at AgenticAdvertising.org in a working-group or governance capacity. The bodies setting AdCP and AAMP standards have no documented Australian company participation in their governance structures. The Privacy Act ADM obligations, which will apply to agent-executed buying decisions in December 2026, are being developed by Australian regulators largely in isolation from the platform infrastructure being built to execute those decisions.

The Australian market is big enough to be a meaningful test environment for the platforms. It is not, at current trajectory, shaping the systems being tested in it.

What local operators need to do

The window to influence the systems is not closed. It is narrowing.

IAB Australia’s commitment to the Tech Lab roadmap and its Standards Council pilot work is the most important current mechanism for local influence. It is under-resourced relative to the scale of what is being built and the pace at which the global standards are being set. The Australian operators with the scale to participate in working groups are not a mystery. The Seven West Media and SCA combined entity, the obvious next AdCP or IAB Tech Lab seat given it is already the major seller in this piece negotiating from the weakest position. REA Group, given how far ahead its product-side AI investment already runs. Nine and the ABC. The holdco agencies, Publicis, Omnicom, WPP, already running global AI platforms locally. All have both the interest and the standing to occupy a seat. Most are not.

A disclosure, because the argument above is about representation: I hold a seat in one of the working groups shaping these standards. One advisor holding a seat is not the same as the publishers, holdcos, and agencies named above occupying theirs, and it is exactly that seat they need to be filling before the standards finish hardening.

On the regulatory thread: the Privacy Act ADM obligations, the ACMA AI-disclosure rule, and the Children’s Online Privacy Code are all arriving in 2026. The brands, agencies, and publishers that have mapped their agentic buying exposure to those obligations before the compliance date are in a better position than those that discover the exposure after an enforcement action. The legal and procurement teams in most organisations have not yet had this conversation. The time to have it is now, before campaigns are running under agent autonomy, not after.

On the protocol question: the local publishers and broadcasters that build seller agents and establish machine-readable inventory declarations, under AdCP or AAMP or both, before buyer agents arrive in volume, are setting the terms on which that inventory is accessed. The ones that do not are accessed at the terms the buyer agent’s platform has set. Those are different commercial positions.

On the brief and governance question: Australian brands deploying agentic buying through global platforms or through holdco AI systems need to specify, in commercial agreements, what decision-level log access they retain. The December 2026 ADM transparency obligations create a legal basis for that access. The commercial leverage to negotiate it is highest before the campaigns run.

The closing frame

This series started with the argument that agentic buying is not primarily an AI story. It is a systems story. The question is not how smart the agent is. It is who drew the box the agent operates inside, whose interests that box was designed to serve, and what happens to the people outside the box when the machine starts working at scale.

Seven parts later, that frame holds.

The infrastructure layer is being built by platforms with their own commercial interests, on standards being written by consortiums representing specific parts of the industry, governed by bodies that are not required to represent the interests of Australian brands, publishers, or audiences.

The measurement framework that will determine how the agents are evaluated is being developed largely in the United States and Europe, for US and European regulatory environments, by companies whose primary commercial interests lie elsewhere. Mutinex, the Australian-founded platform now running multi-agent measurement for clients including Hershey, is the clearest exception: proof that ANZ-built measurement infrastructure can compete globally when it gets the chance to.

The regulatory obligations that will govern how agents make decisions in Australia arrive in December 2026, largely independent of the infrastructure being built to execute those decisions.

The consequence of not engaging with these systems, of receiving the outputs of decisions made elsewhere rather than shaping the decisions themselves, is not a technical outcome. It is a commercial and strategic one. The decisions being made in San Francisco about which protocols win, which data flows where, and who gets to read the log, will shape Australian media markets for the next decade.

On 27 April 2026, the gap between San Francisco and Sydney was four months. By the time the next platform launches, the question will not be how fast it reached Australia. It will be whether anyone in Sydney had a hand in building it before it did. Infrastructure tenancy is comfortable right up until the landlord changes the terms.

The window to have a view on those decisions, and to act on it, is open.

For now.

The complete series, The Frame, The Stack, The Brands, The Agencies, The Publishers, The Measurement, The ANZ View, is available at systemsthatdecide.substack.com.

Every agentic buying system generates two types of records, and the distinction between them is the accountability question the industry has not yet named clearly.

The first is the governance trail: what was proposed, what was approved, what constraints were attached, what actually ran. This layer is being built into the emerging protocol standards. AdCP, the open agentic advertising standard, includes a “get_plan_audit_logs” task as a protocol-level feature, a structured record accessible to the operator and the brand in a compliant implementation.

The second is the optimisation trail: why the agent chose one publisher over another, what inventory it evaluated and rejected, the reasoning behind each bid decision. This layer has no equivalent standardisation. It sits inside the DSP’s or platform’s optimisation logic, and the brand whose budget is being spent has no guaranteed right to read it. Most don’t know to ask.

For fifteen years, media buying has operated without either record. A human trader makes a bid decision in a platform interface. The decision is not recorded in a structured, auditable format. The DSP records the bid. It does not record why. The post-campaign report is a reconstruction from outputs, not a trace of decisions.

The gap between what the governance trail records and what the optimisation trail does not is the most commercially important accountability question in the next era of media.

What legacy measurement actually measures

The measurement infrastructure the industry built over the past two decades was designed to answer one question: did the ad reach the right people?

Impressions served. Viewability rates. Brand safety classifications. Frequency caps. Click-through and completion rates. These are all measures of delivery, not of effect. The attribution models layered on top, last-click, multi-touch, view-through, attempt to connect delivery to outcome, but they do so with the signals available, not the signals needed.

The result is a measurement system that is, by wide practitioner consensus, structurally broken in several places. Last-click attribution systematically overweights search and direct channels at the expense of brand-building media. Multi-touch attribution models produce results that differ dramatically depending on their design assumptions and are not empirically validatable at the decision level. The entire system assumes that the relationship between an ad exposure and a subsequent action can be inferred from observational data without controlled experiments.

The Privacy Sandbox collapse made this worse. Google’s plan to replace third-party cookies with privacy-preserving APIs consumed years of industry attention and was effectively abandoned by April 2025. Of the technologies Google launched to replace cookies, only three minor APIs survived to 2026. None touched the attribution problem the industry actually had. The industry spent three years preparing for a solution that did not arrive.

The answer the industry landed on instead is triangulation. Marketing mix modelling for budget allocation. Incrementality testing for causal ground truth. Platform attribution for in-flight tactical steering. None of these methods is new. What is new is that they are the answer by default rather than the answer by choice, because the alternatives degraded away. This is the context in which agentic buying arrives. Not a clean slate. A patchwork that works directionally, that practitioners distrust privately, and that nobody has managed to replace. Signal loss is permanent. No deterministic cross-site identity is coming.

The brand measurement problem

There is a constraint agentic buying cannot fix. Name it before describing what it can.

No real-time brand signal exists that an agent can reliably optimise against. Brand lift data, where it exists at all, arrives weekly at best and monthly in most deployments. It is survey-based, platform-siloed, and not comparable across campaigns or channels. Attention metrics are a proxy, not a measure of equity. Search query lift as a brand indicator is useful but lagging and correlational. There is no closed-loop transaction record for a change in brand consideration. There is no equivalent to a purchase event that confirms brand intent at the moment it forms.

An agent operating without a real-time brand signal will systematically optimise away from brand investment. Not because it is wrong about what it sees. Because what it sees is only lower-funnel outcomes. Brand outcomes arrive too slowly to appear in the reward function, so the agent behaves as if they do not exist. It reallocates budget toward whatever produces a measurable signal. Over twelve to twenty-four months of autonomous operation, this compounds. The agent is doing exactly what it was asked to do. What it was asked to do is incomplete.

The only current defence is a human-set objective hierarchy defined before the agent runs. Decide deliberately what the agent is optimising toward, how the brand and sales objectives relate, and what constraints prevent the agent from treating brand spend as inefficiency to be reallocated. That hierarchy is part of the brief. The brief is an accountability document. It belongs in the governance log.

The MMM renaissance

The measurement approach that has survived this collapse most intact is Marketing Mix Modelling. MMM is not new. It has been used in some form by major marketers since the 1990s. What is new is the cadence, the data inputs, and how it connects to the buying system.

The traditional MMM process ran three to five cycles per year, covering a subset of brands, with a lag of three to six months between data and insight. That cadence was built around human analysts producing outputs that human planners used in the next planning cycle.

The Hershey and Mutinex case study, the most thoroughly documented example of agentic MMM in market, shows what changes. Hershey manages total media and trade spend exceeding two billion dollars. Previously, it ran three MMM cycles per year across five brands, with a five-month data lag. Under the new architecture, a multi-agent system built on Claude and Gemini with distinct agents for econometrics, competitive pricing, and model-failure diagnosis, it runs monthly MMM across its full portfolio. That is not an incremental improvement. It is a structural change in the relationship between measurement and decision.

Google’s Meridian MMM framework, open-sourced in January 2025, has been downloaded hundreds of thousands of times on GitHub and now has more than twenty certified implementation partners. Sellforte launched a Media Buyer Agent in Q1 2026 with a capability that did not exist twelve months earlier: direct integration to push bid changes into Google, Meta, and TikTok in either self-drive or assisted mode. The loop between measurement and buying, which was previously a human-mediated weekly or monthly process, has been automated.

Lifesight’s Mia agent launched in March 2026. Mutinex, the Australian-founded platform, is running a conversational AI layer on its GrowthOS platform using Claude and Gemini as the agentic interface. Its ANZ clients include Domino’s, Samsung, Lion, Asahi, Optus, REA Group, and ING, which together represent a large share of the Australian marketing budget already operating with AI in the measurement loop.

What the cadence change means is not only faster insights. A model that runs continuously can drift, overfit, or compound a data error across months of recommendations before anyone runs an audit. The governance trail is the mechanism that makes a continuously-running MMM auditable. Without it, you have a measurement system that is faster, more integrated, and harder to check than the one it replaced.

The data quality constraint

The renaissance has a binding constraint: the data going into the models.

Gartner’s October 2025 survey found that forty-five percent of marketing technology leaders said vendor AI agents failed to meet their expectations. A subsequent Gartner study found that fifty percent of organisations lacked the data stack readiness for AI agents to function effectively. Sarah Martinez, Chief Commercial Officer at Tracer, Hershey’s named data partner in the Mutinex deployment, put it directly: “Most companies don’t have an AI problem. They have a data readiness problem.”

Agents amplify the data they receive. If the media investment data flowing into an MMM model is incomplete, delayed, or miscategorised, the agent’s recommendations will be precisely wrong rather than roughly right. The Hershey result, a forecast of four to five percent revenue lift attributable to media, is contingent on data quality that most organisations have not yet achieved.

The ANA’s Q1 2025 Programmatic Transparency Benchmark found that fifty-six to fifty-nine percent of programmatic spend still could not be clearly tied to quality-defined impressions. An agentic MMM that does not correct for that structural inefficiency, because the data it receives does not identify it, will model the inefficiency as baseline performance and recommend maintaining the spending pattern that produces it.

The log and the accountability gap

The structural measurement problem of agentic buying is speed mismatch. The agent clock runs in seconds. The MMM clock runs in months. The attribution gap, the lag between a buying decision and the outcome it produces, is not solved by faster agents. It is made more visible by them.

The log does not solve the attribution gap. That is not what makes it significant. What makes it significant is that it creates, for the first time, a decision-level record of what the agent was doing and that record enables accountability in a way that human buying has never systematically allowed. Those are different claims. One is about measurement. The other is about governance. The industry has been running them together, and that has obscured which problem is actually solvable.

The governance trail: what was proposed, what was approved, what constraints were attached, what actually ran, can be standardised and made accessible. That is what AdCP’s audit log captures: the plan, the escalations, the human approvals, the conditions, the outcome. In a compliant implementation, a brand can query the full governance record. What it cannot query, under any current standard, is the optimisation trail underneath: why the agent chose one publisher over another, what inventory it evaluated and rejected at the selection stage, whether those choices were consistent with the brief at the impression level.

Incrementality testing, the one reliable causal method in the stack, runs into a specific problem under agent autonomy: geo-holdout experiments contaminate when an agent continuously optimises against detected performance differences between test and control groups. The agent will route budget away from the holdout region when it detects lower performance, invalidating the experiment. The honest practitioner answer is policy-level validation: auditing the agent’s optimisation policy, not its individual decisions.

The governance log tells you that the agent was authorised to spend $50,000 on CTV sports inventory with a weekly cap and specific brand safety conditions. The optimisation log would tell you why it chose certain publishers over others, and whether those choices were consistent with the brief beyond the authorised parameters. Without the second log, you can verify compliance with the authorisation. You cannot verify alignment with the objective. Both are accountability questions. Only one has a current answer.

Who owns the audit trail

The two logs sit in different infrastructure, and the distinction determines what brands can and cannot currently know about how their money was spent.

The governance trail has a defined home in AdCP-compliant systems. The protocol specifies that the operator and buyer can query the complete governance record through a standard task. What was proposed, what was escalated, what a human approved, what conditions were attached, and what actually ran. This is accessible as a protocol-level feature, not something that requires special contract negotiation in a compliant implementation. The gap that earlier disputes like the Trade Desk and Publicis matter made visible, whether clients can see what their agents were authorised to do and whether those authorisations were followed, is the gap the governance trail closes.

The optimisation trail is a different question entirely. The buying decisions, the publisher-selection reasoning, the impression-level logic, this is the output of the DSP’s or platform’s optimisation engine. It is not covered by AdCP’s audit log, ARTF’s mutation logs, or any current standard. It is the record that would tell a brand whether the agent’s publisher selection was consistent with its brief, whether it systematically favoured certain inventory for reasons the brand did not intend, or whether its decisions drifted over time in ways that are commercially significant.

This is not a minor gap. The optimisation trail is the audit trail for the actual buying decisions an agent makes on a brand’s behalf. If a brand cannot read it, it cannot verify that the agent’s choices were aligned with the brief. It cannot detect systematic drift at the selection level. It cannot demonstrate to its board that autonomous media buying is operating in its commercial interest beyond the parameters that were formally approved.

The IAB Tech Lab’s Agentic Real-Time Framework (ARTF), which governs how agent containers integrate into the OpenRTB bid stream, requires that every proposed mutation declare its intent, adjusting a bid, activating a segment, curating a deal, and mandates that orchestrators log which mutations were requested and which were accepted or rejected. That is the closest thing currently in production to a structured optimisation trail. But it sits in the SSP’s or DSP’s infrastructure, and the standard is silent on buyer or brand access. AdCP’s governance log covers the workflow layer, what was proposed, approved, what conditions were attached, what actually ran, and in a compliant implementation this is accessible to the brand. Neither standard specifies that the full decision-level optimisation record must reach the advertiser. That is the gap. The brands that negotiate access to it in their commercial agreements before campaigns run, specifying what mutation-level log data they have the right to access, in what format, on what timeline, will have a different accountability conversation than the ones that raise it after.

The regulatory dimension

The measurement and accountability dimension of agentic buying is not only a commercial question. It is becoming a legal one.

The Privacy Act amendments create automated decision-making transparency obligations that take effect in Australia in December 2026. A buying agent making decisions about which audiences to target, which inventory to purchase, and which creative to serve is making automated decisions subject to explanation obligations under that framework. The direction of regulatory travel is clear, and the December 2026 deadline is fixed.

The Children’s Online Privacy Code, also mandated by December 2026, creates additional constraints on automated targeting in contexts where children may be among the reached audience (if you need to read more on the COPC, Dan Richardson provides a good breakdown here). An agent briefed to reach a broad demographic and optimising against engagement signals could, without explicit constraints, serve targeting decisions that breach the code. The brief architecture question from Part Three and the accountability log question are the same question, viewed from different angles.

Brands operating in Australia under agentic buying systems will, by December 2026, need to be able to explain to regulators how those systems make decisions. The governance trail covers what was authorised. If the optimisation trail is held by the platform and not accessible to the brand, the explanation of why specific decisions were made is not possible.

The measurement agenda worth building

The practical forward agenda for measurement in agentic buying is not elegant. There is no single solution that resolves the attribution problem, closes the speed gap, and addresses the optimisation log gap at the same time. What there is, is a sequenced set of questions that operators can begin working through now.

First: data readiness before agent deployment. The fifty percent of organisations that do not have the data stack readiness for AI agents to function effectively will not solve that problem by adding an agent to a broken data environment. They will produce a precise version of the wrong answer faster.

Second: distinguish governance access from optimisation access in commercial agreements. The governance trail, what was proposed, approved, and run, is accessible in compliant implementations. The optimisation trail, why the agent made specific buying decisions, is not currently standardised. Before deploying an agent or authorising an agency to deploy one, specify in the contract what decision-level optimisation data the brand has the right to access, in what format, and on what timeline. These are different asks, and most current agreements address neither.

Third: policy-level validation rather than impression-level testing. The geo-holdout experiment does not work cleanly under continuous agent optimisation. The question worth answering is whether the agent’s optimisation policy is aligned with the brand’s business objective, and that question is answered by reading the optimisation log and auditing the brief, not by running a traditional holdout.

Fourth: account for the measurement blind spot that sits before the first click. As purchasing increasingly routes through AI assistants, brands appear or do not appear in LLM responses before a user reaches any tracked surface. No platform reports on this. No marketing mix model includes it. Brands that optimise heavily on measurable conversion signals are building spend efficiency on a channel they can measure while remaining blind to the recommendation layer that is increasingly influencing which channel the buyer reaches first. That blind spot will not close by itself. It needs to be named in the measurement framework, even before the tools to fill it exist.

The governance trail is accessible today if you are running on a compliant protocol. The optimisation trail, the record of why your budget went where it went at the selection level, is the one that is not. Getting access to it is a procurement decision you can make before the campaign runs.

Part Seven closes the series with the ANZ view: what the infrastructure has delivered locally, what the regulatory clock means for operators in Australia and New Zealand, and what the risk is of becoming a market that resells decisions made elsewhere.

The programmatic advertising infrastructure that publishers built over the past fifteen years was designed around a specific assumption: that the entity making the buying decision is a human trader, sitting at a platform, exercising judgment about which inventory to purchase and at what price.

That assumption held for long enough that publishers built their entire commercial architecture on top of it. Direct sales teams that cultivate relationships. Private marketplace deals that reward trust and exclusivity. Attention metrics that can be explained in a presentation. Brand safety controls that can be communicated to a media buyer. Floor pricing logic that relies on a human to read the signals and respond.

None of that is worthless under agentic buying. Some of it is. Knowing which is which, and doing it before the agents arrive in volume, not after, is the most urgent commercial problem facing publishers in the Australian market right now.

How publishers built for human buyers

The layered yield stack that most premium publishers operate runs on a simple logic, a direct sales team negotiates guaranteed deals at the top, curated private marketplace deals sit below that, and the open auction absorbs whatever is left. Floor prices, private auctions, first-look arrangements, all of it exists to manage the relationship between supply quality and price in a system where a human on the other side can read the signals and pay accordingly.

The direct relationship matters in this model. An agency buyer who has met with the publisher, who understands the editorial context, who has been walked through an audience proposition. That buyer applies judgment the algorithm doesn’t have. They may pay a premium for a placement they understand. They will advocate internally for the publisher’s proposition when a planning conversation happens. The human relationship is a component of yield.

A buying agent does not have a relationship. It has data. If the data it has access to does not capture what makes a publisher’s inventory valuable, the agent cannot factor that value into its decision. It will buy at floor, or not at all.

What agents actually evaluate

The transition from human buyers to agent buyers is a transition in how inventory is evaluated. Human buyers use a mix of data, relationship, intuition, and precedent. Agents use signals that are either present in the bid request and associated metadata, or not present. There is no intuition and no relationship to draw on.

The signals that buying agents can access and act on are expanding rapidly: brand safety classifications, contextual relevance scores, audience segment matches against first-party data, verified attention metrics from providers like Adelaide (now integrated into Nielsen ONE from October 2025), viewability and completion rates, and even the carbon footprint of the supply path via Scope3’s emissions data. None of that list is exotic. It is the working vocabulary an agent already speaks, and the publisher that hasn’t translated its proposition into that vocabulary is, to the agent, simply silent.

Under AdCP and AAMP, the IAB Tech Lab’s rival agentic standard, a publisher can also expose richer capabilities directly to buyer agents. Through a well-known discovery file, a publisher’s seller agent can advertise: what inventory is available, what deal structures are supported, what content signals can be accessed, what audience data is addressable. A buyer agent reads that file before it begins negotiation. The publisher that has built a rich, machine-readable declaration of its inventory’s value is in a fundamentally different position to the one that has not.

The publisher that is not in that file, or whose file is sparse, will be evaluated at the level of the open auction bid request, a domain they have been losing ground in for fifteen years.

Scope3, writing about this same shift in March 2026, called it a move from biddable to discoverable, competition decided not by who wins an auction but by whether a buyer’s agent can find you, read you, and understand what you are offering before the auction even starts. That is a sharper way of naming what this section has been building toward. The publisher that wins here is not necessarily the one with the best inventory. It is the one whose best inventory an agent can actually see.

The signals publishers are not making available

The publisher proposition, as most premium Australian publishers tell it, has several elements that are difficult to encode in current bid stream signals.

Editorial quality is one. The contextual safety and trust of appearing alongside well-resourced, professionally edited journalism. This matters to brands and agency planners. It matters less to a buying agent whose brand safety check runs against a keyword blocklist and a domain safety score. The nuance of editorial quality, that a considered long-form feature produces a different reader state than an algorithmically assembled content feed, is a signal that the current measurement infrastructure does not reliably capture or transmit.

Exclusive audience access is another. The reader of the Australian Financial Review or the Sydney Morning Herald has demographic and professional characteristics that make her valuable to a specific category of advertiser in ways that a segment-match score does not fully represent. The publisher knows this. The agent, working from the bid stream, does not necessarily.

The relationships that have historically bridged this gap, the publisher sales team that communicates these propositions to human planners, become less load-bearing as planning moves toward agent systems. The agent does not take the briefing. It reads the data.

Picture the meeting that doesn’t happen. A publisher sales lead has spent two weeks preparing a deck on why a finance masthead’s audience is worth a premium: the household income data, the purchase-intent signals, the fact that readers stay an average of eleven minutes on a long-form market analysis. In the old model, that deck gets walked through a planner who nods, asks questions, and carries the proposition back to a media plan. In the agentic model, there is no meeting. The buying agent has already priced the inventory from the bid request alone, made its allocation, and moved on, three milliseconds before the sales lead opened the laptop. The deck was right. It was also invisible to the only party now making the decision.

The question for publishers is not whether their inventory is genuinely valuable. In most cases, it is. The question is whether they have made that value legible to a machine.

This is the same divide Part Two of this series mapped, entities with first-party data at scale are building agent-compatible walls, and everyone else is competing within open standards. Publishers with genuinely valuable but hard-to-encode first-party data, property intent, financial behaviour, professional audience composition, sit in the open-standards camp by necessity. That is not a consolation prize. AdCP gives them a route to buyer agents that does not require an SSP or DSP to translate their proposition for them. PubMatic’s AgenticOS, built deliberately on the supply side, had run more than thirty fully autonomous campaigns with more than a thousand direct publisher deals by May 2026. The infrastructure exists. The publishers that have not yet declared their inventory into it are not waiting for the technology. They are waiting for a reason to move first.

The concentration problem

The Seven West Media and SCA merger, completed in January 2026, concentrated national television, BVOD, radio, and audio inventory into a single commercial entity. The rationale was scale and cost efficiency. Combined sales, shared infrastructure, negotiating leverage with advertisers. No part of the public case for the merger referenced AI-native buying, seller agents, or machine-readable inventory declarations. That is not an oversight. It is a signal of where the industry’s attention was when the deal was struck, and a reason to expect the combined entity is further behind on agentic readiness than its scale would otherwise suggest.

That logic was built for a market where scale means more human sales teams, more relationships, more presence in more agency conversations. Under agentic buying, the logic changes.

When buying agents begin negotiating at scale, the counterparty is not a human sales team. It is a seller agent. The leverage in that negotiation is not relationships. It is the richness of the inventory data, the completeness of the deal structure declaration, and the speed and flexibility of the seller agent’s ability to match buyer requirements to available supply.

In a concentrated market, the entity with the largest and most diverse inventory pool has a structural advantage, if that inventory is accessible to buyer agents at the right level of signal granularity. If it is not, concentration creates a different problem: a large seller with underspecified inventory is still bought at floor price. The buyer agent does not pay a relationship premium.

The second issue with concentration in an agentic buying environment is negotiation dynamics at the protocol layer. AdCP’s seller-agent architecture is designed around a model where each publisher or publisher group runs a seller agent that negotiates with buyer agents. A merged Seven West and SCA entity with a single agentic sales infrastructure would be a formidable counterparty for buying agents negotiating CTV, BVOD, audio, and radio inventory in a single interaction. By any public signal, that architecture has not been built. The merger concentrated the inventory. It did not, on the evidence available, concentrate the agentic infrastructure to match it. The commercial logic for building it is strong, and the gap between the two is exactly the structural disadvantage a major seller carries into a market where buyer agents are arriving.

The case for Agentic Direct

Here is the argument I think the concentration story is actually pointing toward. Publisher concentration in ANZ creates the conditions for what I’d call “Agentic Direct,” agent-to-agent negotiation of programmatic guaranteed deals, fixed price, guaranteed volume, the programmatic equivalent of a direct insertion order, that bypasses the open auction entirely. The deal type is not new. What would be new is who negotiates it and how often. Today, a publisher and an agency strike that arrangement perhaps once a year, in a process that can run for months and ties up both sides’ senior commercial people. Agentic Direct replaces that human, annual cycle with something else entirely: a credentialed seller agent and a credentialed buying agent negotiating the same fixed-price, guaranteed-volume arrangement deal by deal, in close to real time, with AdCP supplying the authentication and audit trail that makes a machine-negotiated guarantee enforceable. A buying agent with a sufficiently detailed brief could lock in the inventory, the volume, and the rate it wants in a fraction of the time a human direct deal requires, and do it as often as the brief demands rather than once a year.

Here is the question that makes or breaks this argument: why would an agent built to continuously re-optimise against live signal ever choose to lock itself into a fixed commitment? On its own terms, it would not. But the agent does not set its own terms. It executes a brief, and the brief is where a brand’s long-term mandate lives, the part of the marketing function that has always cared about things current measurement infrastructure still cannot reliably price: category salience, memory structures, the slow accumulation of trust that shows up in a P&L years after the campaign that built it has finished running. None of that is visible in a bid-stream signal. An agent optimising purely on what it can measure will systematically underweight it, not because the value is not real, but because the agent has no instrument for it.

That is the gap a guarantee closes. A brief encoding a brand-building mandate cannot be satisfied by a heavily weighted bid, because a weighted bid is still a bid, and a bid can be outcompeted, moment to moment, by another buyer chasing the same scarce premium context. What that brief actually needs is sustained, consistent presence in that context over a defined period, the kind of repeated exposure that builds the memory structures the brand is paying for. Only a locked commitment, fixed inventory, fixed volume, fixed period, secures that against competitive pressure the agent cannot control deal by deal. The guarantee is not a habit the agent inherits from its human predecessors. It is the most precise instrument available for executing a brief about something the agent cannot measure but has been told, by the brand that wrote it, to protect regardless.

For premium publishers, Nine, Seven (now with SCA), News Corp Australia, the ABC, this would reorder the yield hierarchy in a way that favours them over the current open-auction dynamic. Direct guaranteed deals have always had better economics than open-auction CPMs. If the direct-deal layer becomes accessible to agentic buyers at scale and speed, the publishers that have built seller agents capable of negotiating those deals will capture disproportionate value, and the open auction shrinks to the residual layer it was always meant to be.

I also want to be honest about what kind of claim this is. Nobody has built Agentic Direct yet, and nobody has announced they’re building it. What exists is the infrastructure it would require, and that infrastructure has moved since Part Two of this series said the thing to watch was the schedule, not the announcements. AdCP v3.0, unveiled in late March 2026, gave publishers something they didn’t have before, a standard, publisher-owned definition of exactly what their inventory is, plus a governance layer for calibrating brand suitability, alongside a wider set of structural updates to how that inventory gets classified. The Curation Protocol that Part Two flagged as the likely bridge between AdCP and AAMP is still listed as coming in Q2 2026, on schedule but not yet shipped, so the open question that piece raised, whether it lands on time, is exactly as open as it was then. What v3.0 has already delivered is enough on its own: authentication, an audit trail, and now a publisher-owned definition of what is actually for sale. The pieces are on the table. The publisher that assembles them first does not just gain a new sales channel. It resets the yield hierarchy in its own favour before its competitors understand the game has changed.

REA Group is the most publicly advanced local publisher on AI integration, Gemini deployed into Mortgage Choice, OpenAI integration in the realEstimate valuation tool, conversational AI across the core realestate.com.au experience. The extension of that investment into a seller-agent capability for advertising would be a logical next step, and a logical next mover. Whoever moves first on Agentic Direct captures a lead the rest of the market will spend years trying to close.

REA Group is still the most publicly advanced local publisher on AI integration in general. But on the specific question this piece keeps returning to, whether anyone operating in the Australian market has actually built and declared an AdCP-compatible seller agent, REA Group is not first. Neither is Nine, Seven, News Corp, or the ABC. A crawl of Australian publisher and brand websites for this series found exactly one live, functioning seller agent, detected in late April 2026, Yahoo’s. That is not a coincidence of timing. Yahoo was a founding member of AdCP when the protocol launched in October 2025, alongside PubMatic, Scope3 and four others. It had the earliest possible access to the standard, and it appears to be the only operator in this market that has actually shipped something with it.

The publisher retail media already is

Every publisher discussed so far, REA Group included, is reasoning its way toward infrastructure built for a buyer that no longer exists in the same form. There is one category of seller that never had to make that case to itself. It already has what this piece argues traditional publishers need to build toward, owned inventory, authenticated audiences, first-party purchase data, and a closed measurement loop back to the transaction. That category is retail media.

Retail media networks are not a side conversation to this argument. They are advertising publishers, selling inventory against exactly the kind of rich, proprietary signal a buying agent is built to evaluate, if that signal is made legible to it. eMarketer forecasts US retail media spend will reach USD 69.33 billion in 2026, up from USD 58.79 billion in 2025. In Australia, Cartology and Coles 360 are approaching a combined AUD 1 billion in annual revenue between them. This is not an emerging category. It is a large and growing one, run by sellers who, on paper, hold the strongest hand in the agentic transition of any publisher type.

What is less clear is whether they are playing it. The most direct, recent proof point of a seller moving on this did not come from retail media at all. On 3 June 2026, InMobi and Scope3 launched a sell-side agent for InMobi’s premium commerce inventory, built on AdCP, arguing explicitly that context-rich inventory is poorly served by open auctions and needs agent-to-agent negotiation to express its real value. That is, almost word for word, the argument this piece makes about traditional publishers. It came from a mobile commerce platform, not a retail media network.

The retail media networks best positioned to win this transition, the ones holding exactly the asset this piece says publishers need, do not yet appear, on the public record, to be building the seller-agent infrastructure to capitalise on it. Coles’s data clean room arrangement, established in January 2026, is a step toward making first-party signal accessible to external systems. It is infrastructure adjacent to the problem, not yet a declared answer to it. The retailer that gets there first will not just be selling media well. It will be demonstrating, for every publisher watching, what it looks like to convert a genuine first-party advantage into agentic-era leverage before someone else’s buyer agent prices it at floor.

What publishers need to do

There are now two open routes to the same prize, and they run through different players. Agentic Direct rewards whichever premium publisher builds the seller-agent infrastructure to negotiate guaranteed deals at machine speed. REA Group is the likeliest of the five heavyweight local movers to get there.

But the infrastructure stopped being hypothetical the moment Yahoo, a founding member of the protocol, switched its seller agent on, and none of the five Australian-headquartered names this piece keeps returning to has matched it yet.

The retail media route rewards whichever retailer turns its first-party advantage into a declared, agent-legible proposition before a generalist platform like InMobi does it for commerce inventory at scale. Both prizes go to whoever moves first. Neither has been claimed. That is the gap this piece has been describing from every angle: not a future problem, but a live one, sitting open, right now, for someone to walk through.

If none of Nine, Seven, News Corp Australia, the ABC or REA Group has declared an AdCP-compatible seller agent of its own by the end of 2026, roughly eight months after one of the protocol’s own founding members proved the thing works in this exact market, that gap stops being a temporary head start. It becomes a structural feature of the market, the kind that takes years and a great deal of money to close from behind.

The practical implication is uncomfortable because it requires publishers to do two things simultaneously that have historically been in tension: invest in technical infrastructure and maintain the commercial relationships that generate direct revenue today.

On the infrastructure side: the minimum viable position is a machine-readable inventory declaration. A well-known URL with a complete and accurate description of available inventory, deal structures, audience data access, and content signals. An AdCP-compatible seller agent is the next step, and it is no longer the heavy, scaled-player-only build it looked like a year ago. Magnite launched a live seller agent in SpringServe on 6 January 2026. On 29 January 2026, Prebid took over stewardship of an open-source, AdCP-compatible seller agent, putting a neutral, low-barrier path to this infrastructure within reach of any publisher, including the mid-tier Australian mastheads and broadcasters that could never have justified building one from scratch. The excuse that this is too large a lift for anyone but a global platform no longer holds. Without either the declaration or the agent, a publisher’s inventory is accessed only through the open auction, at auction prices, by agents that have no signal of differentiated value to act on.

Go back to the sales lead with the deck nobody read. The infrastructure fix is not a better deck. It is making sure the eleven-minutes-on-page figure, the household income data, the purchase-intent signal, exists somewhere a buying agent can actually find it: in the bid stream, in the discovery file, in a declared audience segment the agent can match against. Build that, and the same proposition that went unread in a meeting that never happened gets priced into every single auction the publisher’s inventory enters, automatically, at scale, without anyone in the room. That is the trade publishers are being offered: give up the meeting, keep the value, but only if the value has been translated into a language the room’s replacement can read.

On the commercial side: the direct relationship with planners and buyers does not become worthless immediately. The transition from human to agent buying is uneven across categories, deal types, and budget tiers. The premium end of the direct market will remain human-mediated longer than the open-auction end. Publishers that use the current window to build agent-accessible infrastructure while maintaining the relationship-dependent revenue they have today are in the best position. Publishers that wait for agentic buying to become dominant before investing will be building infrastructure to catch up to a market that has already priced them.

The buyer’s agent does not care about the editorial quality you explained in a presentation three years ago. It cares about the signal it can read today. The question is whether that signal is in the data, or only in a relationship.

Part Six looks at measurement. The audit trail the agent generates at every decision point, who controls it, and why the question of who gets to read it may be the most commercially important unresolved question in the stack.

For the deeper dive on AdCP, AAMP, and the protocol fight underneath all of this, the earlier edition is here: “The Most Consequential Document in Advertising Right Now Is One Most People Haven’t Read”

The brief is about to become the most important document in advertising. Not because strategy suddenly matters more. Because it is the last place where human judgment enters a transaction that agents will increasingly execute without supervision.

In July 2024, we launched Afterpay’s commerce media network, the first payments-native ad product in ANZ combining Afterpay and Square’s transaction data. The pitch to agencies and brands was straightforward: we have purchase data that no DSP can replicate, connected to over 10 million verified buyers, 240,000 merchants, with a closed loop between media exposure and transaction in a way no publisher or retail media network in the market could match.

The briefs we received in response were, almost without exception, built for a different world. Audience segments expressed as demographics. Success metrics defined as vanity metrics. Optimisation goals that had nothing to do with what the data could actually tell us. Not because the agencies were bad at their jobs. Because there was no shared language for what the inventory could do, and no brief format that could encode it. The inventory and the brief were speaking different languages. The campaign ran. The data was largely ignored.

That is the failure mode the agentic era inherits, at scale. As agents take over execution, the brief becomes the only document the agent cannot rewrite. Write it badly, and the agent optimises beautifully toward the wrong thing. Write it well, and it is a precision instrument that performs without drift across every channel and placement the agent touches. The brief is no longer a creative document. It is a technical specification. And most agencies are still writing poetry when the system needs code.

How agencies respond to that problem will determine which ones are still relevant in three years.

What agencies actually do

The agency value chain has four layers, and they are not equally exposed to automation. Most discussion of agentic buying stops at that observation. The more useful question is which layers are collapsing, and how fast.

The first is research and analysis: audience understanding, competitive context, attribution. AI tools have compressed this layer dramatically over the past two years. Work that took a planning team three weeks now takes a well-prompted agent three hours.

The second is execution. Trading, trafficking, pacing, optimisation, reconciliation. This is the layer agentic buying directly automates. The Omnicom and IPG merger, which closed in November 2025 with a stated target of $750 million in annual cost synergies, should be read in this context. The combined entities have cut more than 10,000 roles across two years. These are not redundancies of convenience. They are the execution layer being handed to machines.

The third is strategy. The brief architecture. The channel allocation. The investment philosophy. The layer where experienced judgment, commercial relationships, and accountability for outcomes sit. It is the hardest to automate. Not because it is more creative, but because it requires owning a point of view in the face of uncertainty. Agents optimise. They do not commit.

The fourth is creative concept. The originating idea that makes the media plan worth executing.

The compression is concentrated in the first two layers. That is where the margin is going. Agencies absorbing those losses while assuming strategy and creative concept will hold are right that those layers are durable. They are wrong that the timeline is comfortable.

The compression is real

Forrester projects fifteen percent of agency jobs will go in 2026, concentrated in the clerical, administrative, and market research roles that sit squarely in the compressed layers above. This is happening into a growing market. The jobs are not disappearing because there is less work. They are disappearing because agents do the work.

When briefing-to-skeleton-plan work that took four days takes four hours, the agency has two options. It can charge the same fee and absorb a margin improvement. Or it can charge a lower fee and compete on price. Neither is comfortable. The first requires the client not to notice. The second is a race that ends somewhere agencies do not want to be.

The holdco earnings tell the story of what happens when you get the bet right versus wrong. Publicis posted strong 2025 results, winning Coca-Cola North America, Mars global, and PayPal. WPP pre-tax profits fell seventy-one percent in the first half of 2025; Dentsu reported a JPY 328 billion loss. The divergence is not random. Publicis made a specific bet on technology integration at scale, anchored by Epsilon and CoreAI. That bet is pulling ahead.

Three pressures, converging

The structural threat to agencies is not one thing. Three forces are compressing from different directions simultaneously. See them separately and each looks manageable. See them together and the urgency comes into focus.

The five major holding companies are all running versions of the same strategy: build or acquire an AI operating layer, use it to win new business, use the new business to justify the investment. The label each has given this effort differs. The underlying bet is the same.

Publicis has gone furthest. CoreAI, a three-hundred-million-euro three-year investment, uses Epsilon’s four-billion-person consumer database as its foundation. Its Microsoft partnership, expanded in April 2026, is the most concrete agentic-marketing alliance in market.

The clarifying move came in May 2026, when Publicis agreed to acquire LiveRamp, the data collaboration platform connecting 25,000 publisher domains and 500-plus technology partners, for roughly $2.17 billion. Most coverage treated this as a data acquisition.

It is something more specific.

LiveRamp authored the User Context Protocol (UCP) and donated it to the IAB Tech Lab’s agentic standards initiative. UCP is one of two protocols competing to become the standard grammar that buying agents use to communicate with publishers. The other is AdCP, led by Scope3’s Brian O’Kelley and built on Anthropic’s MCP. Whoever wins that standards race shapes how every agent in the market reads the world.

Publicis is not buying data. It is buying a stake in the grammar the agents speak.

The control thesis has shifted. It is no longer about having the best inputs. It is about writing the language.

Omnicom, having absorbed IPG, is consolidating around Acxiom’s 2.6-billion verified identity dataset. WPP is running a margin strategy: fewer people per pound of revenue through WPP Production and Open Pro, which carries the risk of looking less like a creative partner and more like a software-enabled production operation. Stagwell’s The Machine made it the first holdco partner on Trade Desk’s Koa Agents alpha via the Open Agentic Kit integration framework, positioning for a negotiating advantage when that programme scales.

The race is not for AI capability. It is for data infrastructure, and for the standards that infrastructure runs on.

In February 2026, WPP and Dentsu both quietly exited OpenPath, Trade Desk’s direct publisher-buying programme, framed as commercial decisions but also a signal. The FirmDecisions audit of Publicis then alleged improper DSP fees, unauthorised tool opt-ins, and the absence of cost validation. The Publicis memo of 17 March 2026 advised clients to reduce Trade Desk exposure.

Read as a transparency dispute, these are three separate events. Read as a margin war, they are one.

Holding companies at twelve to eighteen percent operating margin compete with platforms at thirty to forty-two percent. When the platform’s infrastructure can execute a function the agency charges for, the agency’s margin on that function is at risk. Trade Desk’s direct publisher relationships through OpenPath represent a structural threat to agency inventory access economics. Publicis used the FirmDecisions findings as grounds to advise clients to reduce Trade Desk exposure. The structural logic of client protection and margin defence are not mutually exclusive here. Both motivations were almost certainly present.

The holdcos are also building their own exit.

On its Q1 2026 earnings call, Omnicom told investors it has executed live client media buys using an agent-to-agent framework built on the Ad Context Protocol, buying directly from publishers to remove ad-tech intermediaries. John Wren described those intermediaries as extracting a “toll” paid ultimately by clients.

Forrester projects principal-based buying will rise to roughly a third of total agency billings in 2026. Agent-to-agent buying is the mechanism that scales it.

The sharpest threat is the platforms putting agentic buying directly into the hands of advertisers, with no agency in the loop.

Meta acquired Manus AI in December 2025, reportedly for more than two billion dollars, and embedded it in Ads Manager so advertisers can run campaigns conversationally. Advantage+ already manages billions in spend autonomously. Google has shipped Ads Advisor and released an open-source MCP server for its Ads API. Amazon Ads launched an MCP server in open beta that turns natural-language agent prompts into structured ad-system calls. eMarketer estimates roughly 70 percent of digital ad spend is now managed by agentic systems, concentrating in Alphabet, Amazon, and Meta.

Platform agents optimise inside their own walls. They cannot hold a cross-platform investment thesis, cannot arbitrate between channels competing for the same budget, and have no incentive to tell an advertiser to spend less on their own inventory. That is the space the surviving agency functions occupy. But the platforms have made the execution layer free and self-serve. The agency can no longer charge for it.

Three pressures, from three directions. The margin is going from below. The infrastructure race is pulling ahead above. The execution layer is being commoditised in between. What is left to sell when the compression completes is the question this piece is actually about.

Where the value actually goes

The agency functions that survive agentic compression are the ones that require owning an outcome rather than executing a task.

Brief architecture is the clearest case, and the one I watched fail in real time at Afterpay. The data we had was differentiated: verified purchase behaviour, closed-loop transaction signals, 10 million active buyers. The briefs we received were built for reach and frequency. There was no shared language for what the data could do, and no brief format that could encode it. That problem scales. As agentic buying takes over execution, the brief becomes the primary point where human judgment enters the transaction. An agency that can encode a brand’s objective hierarchy, its principal media guidelines, its measurement framework, and its audit access requirements into a form an agent can execute without systematic drift is providing something hard to replicate. The brief is no longer a creative document. It is a technical specification. And most agencies are still writing poetry when the system needs code.

Agent governance is a second durable function. It goes deeper than accountability structures.

An agency has always been the caretaker of how a brand shows up in market. That role now requires something more precise. An agent executing a media plan has no concept of brand. It has parameters, objectives, and guardrails someone wrote. Brand identity lives in the calls made at the margin: this placement is technically within spec but it’s not where we want to be seen; this creative works on performance metrics but it’s off-voice. The agent cannot read that ambiguity. The agency that has translated brand identity into a form a system can execute without drift is doing something that does not get cheaper as automation scales. It gets more necessary. That translation is a bespoke operating specification for every brand the agency services. Not a template. Not a standard offering.

There is a second structural limit the agent cannot cross. Brand growth metrics, awareness, consideration, affinity, show up three, six, twelve months after the investment that produced them. An agent optimising on real-time signals will find performance every time, because performance is visible. MMM and incrementality testing are improving, but they still cannot validate a brand bet inside the optimisation window an agent operates on. The brand strategist committing to brand spend when the data won’t validate it for a year is doing something an agent structurally cannot do. Not because agents lack sophistication. Because the signal the agent needs does not yet exist.

This is what the caretaker role actually means in an agentic environment. Not oversight of a system. Ownership of the thesis the system cannot hold.

Creative concept is a third durable function. Not production, not adaptation. The originating idea that gives a campaign its distinctive, its reason to exist. Springboards, a Sydney-founded creative strategy platform backed by Blackbird Ventures and used by 120-plus agencies globally, was built on this observation: frontier LLMs converge. In testing on Novelty Bench, an independent creativity benchmark, leading LLMs averaged 2.88 out of ten for functionally distinct outputs. Springboards’ own Flint model, purpose-built against the convergence problem, scored seven. That gap is the agency value proposition made measurable. Agencies that have collapsed concept origination into production and execution work have traded their most durable asset for short-term margin.

What this means in ANZ

The ANZ agency market is dominated by holdco subsidiaries without meaningful local data infrastructure. When the LiveRamp acquisition closes, that gap widens. LiveRamp operates in 14 markets including Australia, so the data co-creation layer Publicis is assembling will have direct local reach. The other holdcos are deploying global AI platforms without anything comparable.

The independent market is navigating this differently. Without the capital to build proprietary AI infrastructure, agencies like Hatched, Half Dome, This Is Flow, JOY, and Ryvalmedia are aggregating capability through partnerships. Those that specialise deeply in strategy or creative, and use AI for execution, are more defensible than those trying to replicate holdco scale at lower cost.

No Australian independent has run an agentic buying pilot at the scale of Butler/Till’s December 2025 credentialing campaign. Here is the call: the window for the independent market is probably 18 months. Not to build data infrastructure — that ship has sailed. But to establish a clear position in brief architecture and agent governance before those functions are claimed by platform tools or in-house brand teams. The independents that will hold their ground are the ones already translating brand standards into agent-executable specifications, before their clients ask them to. The ones waiting for the client brief will find someone else already wrote it.

The question worth carrying forward

What does an agency sell when the agent does the buying?

Strategy, governance, and originating creative. None of those are small things. All of them require a different commercial model than agencies have historically used to price execution.

Software engineers are a useful analogy. They are not paid by the hour for code that may or may not work. They are paid for functional output and for the reduction in technical debt. A good engineer who writes clean, well-documented code that runs without intervention is worth more than one who produces the same feature count with three times the maintenance overhead. The brief architect’s value is structurally identical: priced against the quality of what the agent produces downstream, and against the reduction in systematic error. Not for hours spent on the document.

Agent governance is a retainer structure tied to accountability for outcomes the agent generates, not a project fee. Creative concept justifies the media plan existing at all, and should be priced accordingly. And principal media, the model the holdcos are racing toward, only holds if the agency owns the strategy that makes the buy worth making. Agencies that try to migrate upstream while keeping execution-era pricing will find the margin problem follows them.

Over the next twelve to twenty-four months, the agencies that move to these positions and reprice as they go, before they are occupied by platform tools or in-house brand teams, will have a durable answer. The ones that wait for the compression to complete before repositioning will be answering it from a much weaker place.

That work is available to any agency willing to do it. Not many are moving fast enough.

And it points at a question this series will keep returning to: when the buyer is an agent, and increasingly when the consumer is buying through an agent too, the closed loop between media and transaction stops being a nice-to-have and becomes the contested ground.

________________________________________

Part Five looks at what agentic buying means for publishers — specifically, the problem of having built an entire sales infrastructure for human buyers when the buyers are no longer human.

“Drive consideration among 25-to-54-year-old Australians. Purchase intent index above 110.” That is a brief line. A human planner reads it as a starting point for a conversation. A buying agent reads it as an executable instruction.

The agent finds the cheapest inventory meeting those parameters and buys at scale. Cost-per-view comes in under forecast. Consideration doesn’t move. Nobody made an error. The agent worked exactly as designed. The brief was the problem.

Part One of this series made a blunt observation: agents amplify the system they are given. They do not correct it. Part Two asked which side of the infrastructure divide brands are on: open standards or walled garden, AdCP or a proprietary MCP. This piece asks a different question. Whatever stack you are running, whatever protocols govern how your agent transacts, what are you telling it to do? And is that actually what you mean?

The question being asked in most CMO conversations I am aware of is some version of: how do we govern the agent? What guardrails do we put in place? Who is accountable when it makes a bad buy? These are real questions. They are also, in a precise sense, the wrong starting point. The agent does not create the accountability problem. The brief does.

That scenario has not played out at scale yet. Agentic buying has run on a small number of closely supervised campaigns in the United States and United Kingdom, proof of concept, humans in the loop, limited autonomy. In Australia and New Zealand, it is essentially pre-deployment. This piece makes a structural argument, not a post-mortem: the brief problem, the principal media problem, the liability gap, the log nobody can read. These are built into the architecture. They will emerge when supervision is removed. The question for brands is whether to address them before that point or after.

What the brief actually does now

A media brief, in traditional practice, is an instruction that assumes interpretation. The human planner reading it will exercise judgment. Weighting reach against precision. Reading the competitive context a client didn’t write down. Pushing back on a target CPA that conflicts with the campaign’s real objective. The brief is a starting point for a conversation.

An agent does not have that conversation. It reads the brief as a specification. “Drive consideration among 25-to-54-year-old Australians with a purchase intent index above 110” is, to a human planner, a starting point for a channel discussion. To a buying agent, it is an executable instruction. It will find the inventory that satisfies those parameters at the lowest price, or the most scale, or whatever optimisation function the platform defines, and it will buy it. At speed. At scale. Without checking back.

This is not a failure of the agent. It is working exactly as designed. The failure, where it occurs, is upstream. The brief was written for a human and handed to a machine.

When I was building commerce media at Afterpay, we kept an informal log of what briefs said versus what planners actually needed. They were rarely the same document. What closed the gap was two conversations and a spreadsheet. An agent doesn’t have those conversations. It has the document.

The fix is not subtle. A brief built for agentic execution needs constraints a machine can honour. “No more than 40% of total investment may optimise toward lower-funnel signals” is a machine-readable instruction. “Balance brand and performance” is not. The agent will find the metric that satisfies “performance” and follow it every time. Writing briefs this way looks less like creative direction and more like a procurement specification. That is not a downgrade. It is an accurate description of what a brief now is.

The brief is the one lever in this system that brands directly control. Use it as such.

The brand the agent cannot see

There is a specific problem the industry is not yet talking about clearly: agentic buying creates a structural bias against brand investment.

Agents optimise toward measurable signals. Long-term brand building does not produce measurable signals on the timeframe of a campaign flight. The evidence base for the value of brand investment (the work on long and short, the IPA Effectiveness Databank) is built on years and categories, not impressions and clicks. An agent briefed to optimise a campaign cannot see that evidence. It can only see what the measurement model returns. And measurement models are built for attribution, not for salience.

Run that forward over time. Brands running agentic buying without explicit brand investment parameters will find their agents systematically underinvesting in the media types that build long-term pricing power: premium contextual environments, high-attention formats, the inventory categories that create memory structures rather than conversion events. The agent is not wrong to deprioritise them. They are harder to justify in an optimisation function. That is exactly the problem.

The brands that avoid this encode the investment ratio into the brief as a hard constraint: no more than X% of this investment can optimise toward lower-funnel signals. That is something an agent can honour. Your brief probably doesn’t say this. The evidence for why it should has existed for decades. The only thing that changed is that ignoring it is now faster.

That is the measurement architecture problem. The agent cannot optimise toward what the measurement model cannot return. There is also a commercial architecture problem, and it predates the agent entirely.

Who is the agent working for

The principal media problem in human buying is structural: an agency that holds proprietary inventory operates within a commercial architecture that creates a genuine tension between its own economics and the client’s objectives. This series opened with that frame, the Publicis and Trade Desk dispute as a question about where commercial value accrues in the buying chain, not simply a transparency question. The FirmDecisions audits, the ANA’s work on principal media guidelines, the entire industry conversation about commercial disclosure. All of it is downstream of the same structural fact.

That structural tension does not resolve itself when buying is automated. An agent deployed within a holdco’s buying infrastructure operates under the same commercial architecture the humans operated under. The difference is that in the human version, the tension is visible and negotiable. In the agentic version, the architecture is encoded into the agent’s configuration. The log records what the agent chose. It does not always explain the constraints under which it was making choices.

The ANA’s March 2026 study quantified where the governance gap currently stands. Ninety percent of marketers question whether principal media recommendations are actually in their interest, up from seventy-nine percent in 2024. Only fifty-seven percent have written guidelines governing the relationship. (The ANA does not publish a prior-year figure for guidelines, so “flat” overstates what the data shows. What it shows is that near-universal concern has not translated into formal governance.) Use of principal media grew from forty-seven to fifty-eight percent of marketers in a single year. The governance gap is not narrowing as the practice matures. It is widening as adoption outpaces the rules.

Agentic buying does not resolve it. It inherits it.

A buying agent deployed by an agency operates within the commercial architecture that agency has built. Its access to inventory, its preference ordering among deal types, the data it draws on to make targeting decisions. All of these reflect choices made by the organisation that built and deployed the agent. A client does not automatically have visibility into those choices. The architecture does not require it. That is the governance gap, and it exists regardless of the intentions of the people running the system.

The audit trail AdCP’s architecture generates is, in principle, directly relevant here. Two questions determine whether it functions as accountability in practice: whether brands will have access to the log, and whether anyone on their side can interpret it at scale. Most client-agency agreements have not addressed the first. Almost no brand marketing organisations are currently equipped for the second.

There is a version of the future worth naming: the brand builds the agent, encoding its own brief, objective hierarchy, and principal media framework, and gives it to the agency to operate within the agency’s buying infrastructure. That is a model where the agency’s expertise and the brand’s objectives are explicitly aligned before the machine runs. Whether that is the conversation brands and agencies are having yet is a different question.

The governance question for brands is not “how do we audit the agent.” It is “what does our contract say about access to the agent’s reasoning, and do we have the capability to use it if we get it.”

What the fee is actually for

The total cost of agentic buying is not lower. It is differently structured.

When the agent does the execution (as it will), the labour line on the agency invoice will shrink. That is visible and negotiable. What replaces it is not. Data access fees, MCP licensing, agent infrastructure, the cost of operating within proprietary ecosystems. These costs are real. In the commercial structures taking shape around agentic infrastructure, the technology cost will be paid by the brand and make money for the platform. The agency remains the client relationship and the strategic interface. The brief will be optimised. The margin will be somewhere else in the chain. The early deployments in the US and UK already show this pattern. It is a design feature, not a prediction.

The traditional agency value proposition rested on three foundations: market expertise, trading relationships, and the labour of execution. Agentic buying displaces the third and changes the second. Agency value now concentrates in what goes into the agent, not what the agent does. Brief quality, objective hierarchy, principal media framework, audit access provisions. These are strategic inputs. The commercial model built around labour hours is the wrong wrapper for that kind of value.

Brands negotiating agency fees down while platform technology costs compound unchecked are optimising what is visible and ignoring what is substantial. The question to put to your agency is not what they are doing for the fee. It is what they are putting into the agent on your behalf, and whether that input is worth what the platform is charging to execute it.

What ANZ brands are not doing

There are no Australian or New Zealand brands that have made a public commitment to either of the two standards governing how buying agents operate. There are no public case studies of an ANZ CMO defining the governance framework for their agency’s agentic deployment.

This is not a criticism. The infrastructure has moved fast and the standards are young. PubMatic’s AgenticOS only went live in Australia in April 2026. Yahoo DSP launched agentic capability here in January. The platforms arrived before the governance conversation has had time to catch up.

But the governance gap is not just a timing problem. The Australian agency market is concentrated among global holdcos whose agentic infrastructure is being designed offshore, for different market conditions, and deployed here with limited local adaptation. There is no ANZ equivalent of the ISBA or ANA other then the AANA. The programmatic transparency conversation arrived late in this market; the agentic one is arriving later still. The smaller scale of ANZ media investment reduces the urgency of building bespoke governance when the global default is available and apparently working.

The consequence is that the governance gap in ANZ is not merely behind the global pace. It is behind it for reasons that will not self-correct when the platforms mature.

When selling agents begin negotiating at scale on behalf of consolidated inventory owners, and the proposed Seven West Media and SCA consolidation has the potential to concentrate that inventory substantially further, [verify merger status and completion date before publication], the brands whose buying agents are operating without clear objective hierarchies, principal media guidelines, or audit access provisions will be on the weaker side of that negotiation. Not because agent-to-agent buying disadvantages buyers by default. Because in a concentrated market, the selling side has a simpler mandate, and a selling agent with a clear brief will consistently outperform a buying agent without one. The governance gap is not an abstract risk. It is leverage in the hands of the counterparty.

The liability question is not settled

When a human trader makes a buying decision that damages a brand, placing an ad adjacent to content that generates a crisis, executing a buy that constitutes a breach of exclusivity, the liability chain runs through the agency’s master services agreement. There is a legal framework, however imperfect, for assigning accountability.

Read your current agency agreement. Find the clause that addresses what happens when an automated system, not a human trader, makes a buying decision that damages your brand. If that clause doesn’t exist, you are already in the liability gap, before you have authorised a single agentic campaign.

When an agent makes that decision, the liability chain runs through a different set of documents: the brief, the contract governing the agent’s deployment, the terms of the DSP platform on which the agent operates.

This chain is not currently clear, and the technical standards don’t close it. The IAB Tech Lab’s AAMP framework addresses agent credentialing and audit logging. AdCP’s architecture includes an authentication step (a confirmation key for every media buy) designed to prevent unauthorised transactions. These are technical foundations for accountability. They are not legal frameworks.

Before you authorise agentic deployment, the clause you need covers three things: liability assignment when an automated system makes a buying decision that damages the brand, not just when a human does; access rights to the agent’s decision log in machine-readable format; and a definition of “automated system” broad enough to cover architectures that don’t exist yet. Your legal team will push back on the last one. Push back harder.

Brief your legal and procurement teams before you brief your agency on agentic deployment. That is not a technology task. It is a contract task, and it has a deadline.

The log nobody can read

The ISBA/PwC programmatic supply chain study is instructive here, though not in the way it is usually cited. It found that fifteen pence in every pound of UK advertiser programmatic spend could not be attributed to any part of the supply chain. That study required forensic-level effort, tracing hundreds of millions of pounds through eight advertisers, twelve agencies, five trading desks, and ten DSPs, to surface a partial picture of what the system was doing. The information existed in the system. The audit rights existed. The problem was that nobody was equipped to use them systematically until PwC got involved.

The ANA’s Q4 2025 Programmatic Transparency Benchmark measured the same condition from the demand side. Advertisers who actively governed quality, using log-level transparency, curation, and disciplined execution, converted 56.7% of programmatic spend into benchmark-qualified impressions. Those who did not converted 37.5%. A nineteen-percentage-point performance gap between the brands that can read what the data says and those that cannot. Not fraud. Not hidden. Just capability.

That is the baseline agentic buying replaces. The log it generates is more complete, more granular, and more machine-readable than anything the ISBA study tried to reconstruct after the fact. The performance gap between brands that can read it and those that cannot will be larger, not smaller.

If you were handed your agent’s decision log tomorrow (ten million entries, each with its reasoning), who on your team would open it? What would they look for? What would constitute a finding that changed something? If you don’t have answers to those questions, the log is not an accountability tool. It is a record of decisions you cannot review, attached to a liability you cannot dispute.

Transparency without legibility is a different kind of opacity. The industry has been here before. This time the paperwork is better. The question is whether anyone reads it.

The thing that does not change

Everything in this piece is about governance: brief design, objective hierarchy, principal media guidelines, audit access, fee structure, investment ratios. None of it is the interesting thing.

The interesting thing is that all of it requires brands to have clarity about what they are actually trying to achieve. Not a CPA target. Not a reach goal. Not a brand safety score. A business objective, clearly enough defined that a machine can pursue it without producing an outcome the brand didn’t want.

Most brands do not have that clarity today. Human buying survives the ambiguity because humans adapt, interpret, and exercise judgment. Agentic buying does not. The ambiguity becomes instructions. The instructions become a campaign.

So: before your agency introduces agentic buying on your behalf, ask for the brief template they plan to use. Ask what constraints are hardcoded versus optimisable. Ask what the agent’s objective hierarchy looks like and who wrote it. Ask what the contract says about log access. If they cannot answer those questions yet, that is useful information. The agent will be deployed when they can. The contract should be ready before that.

The agent is not going to fix the ambiguity. It is going to reveal it. At scale. On every impression. With a log.

Part Four looks at what this means for agencies, the value chain being compressed, the holdco race to own the decisioning layer, and the structural question of what an agency is actually for when the buying is automated.

Something happened in December 2025 that most of the industry read as a technical milestone. It was not. It was a political one.

Magnite, one of the world’s largest supply-side platforms, ran the first live agent-to-agent media transaction using a new protocol called AdCP. A piece of software acting on behalf of a buyer negotiated directly with a piece of software acting on behalf of a publisher group. LG Ad Solutions, Warner Bros. Discovery, MiQ. Real inventory, real money, no human making the buy decision.

The industry covered it as a technical proof-of-concept. What it actually was: the first live test of who gets to write the rules for a buying process that could eventually replace the entire stack of interfaces, platforms, and intermediaries sitting between a brand’s objective and a media impression.

The first piece in this series framed the Publicis and Trade Desk dispute as a power question, not a transparency one. The same logic applies here, one layer down. The companies that write the standard governing how agents communicate when buying media will define who sits at the centre of every agentic transaction and who sits at the edge. That is the decisioning fight by another name.

Understanding why requires understanding one variable.

The one variable

The stack is splitting along first-party data at scale.

Entities that have it (user behaviour, purchase history, payments data) are building agent-compatible walls. Proprietary MCPs, on their own terms, with full visibility into how agents query them and what signals they surface. Entities that don't have it are competing within open standards, because open standards are the only infrastructure they can access. Publicis acquiring LiveRamp for $2.167 billion this week is what crossing that line looks like.

There is a middle tier worth naming. Publishers with genuinely valuable first-party data (property intent, financial behaviour, health signals) but without the engineering capacity to build proprietary agent infrastructure. They sit in the open-standards camp by necessity, but they are potentially its biggest beneficiaries. AdCP gives them direct access to buying agents without needing an SSP or DSP to intermediate. That asymmetry is underreported.

Every product announcement, every protocol launch, every governance body decision in this space maps back to that divide.

The Publicis acquisition of LiveRamp, announced this week for $2.167 billion, is the most concrete proof yet. Most coverage is framing it as an agentic play. That is not wrong. It is incomplete. Publicis is buying their way to the right side of the data sovereignty line. The agents are the output. The data is the moat.

Understanding which camp any given player occupies tells you more about how the stack restructures than anything the trade press is currently writing.

The protocol fight

Two standards are competing to govern agentic media buying. Neither has won. Both are live in market.

AdCP, the Ad Context Protocol, launched in October 2025 by a coalition of more than 20 organisations including Yahoo, PubMatic, and Scope3. Clean-sheet design, built from the ground up on the assumption that both buyer and seller are AI agents. Under AdCP, a publisher declares its inventory and deal conditions at a standard web address. A buyer agent reads that file, connects to the seller agent, requests available inventory against a brief, and confirms a buy. At every step, the buyer agent logs its reasoning. That log is the audit trail. In a market in the middle of one of the most public transparency disputes in programmatic history, a machine-readable record of every buying decision is not a neutral technical feature. It is a commercial and political one.

AAMP, the IAB Tech Lab’s competing standard, takes the opposite approach. Rather than rebuilding from scratch, it wraps the existing programmatic stack (OpenRTB, the Deals API, OpenDirect) with agent-compatible connectors. The case for it is simple: fifteen years of deal infrastructure, publisher relationships, and commercial agreements do not get retired. They get a new front door. The argument against is that grafting LLM reasoning onto OpenRTB parameters leaves too much behind. A buying agent briefed to find inventory aligned with a sustainability narrative at a specific CPM range cannot fully express that brief within a schema built for speed and price. AdCP was designed to carry the full brief. OpenRTB was not.

Before taking a position: the organisations most actively building AdCP are supply-side players. PubMatic. Magnite. Scope3. The standard that disintermediates the demand side is being written by the supply side. That is not a coincidence. It is a commercial interest. The same logic applies to AAMP. The IAB Tech Lab’s member base is heavily buy-side, the agencies and DSPs whose existing infrastructure AAMP is designed to preserve. Both standards are written by parties with a stake in their adoption. Anyone evaluating which to build on should factor in who benefits most if it wins.

Where I stand

AAMP will win the near term. Not because it is better designed, but because it asks less of buyers.

Large buy-side organisations will not rebuild integrations from scratch. They will add an agentic interface to what they already have and call it done. AdCP is the right architecture for the long term. The clean-sheet design, the audit trail, the ability to carry a full brief without compression are the infrastructure of a genuinely different buying system.

AdCP’s Curation Protocol, planned for Q2 2026, is the most likely bridge between them. Whether it ships on schedule is the one variable to watch in the protocol fight. If it does, there is a path to convergence on AdCP’s architecture. If it slips, AAMP calcifies as the market default.

Watch the schedule. Not the announcements.

The more consequential implication of the Curation Protocol is structural. Curation, the packaging of inventory against specific signals or brand criteria, currently happens inside SSP infrastructure. You need the platform to do it. When the Curation Protocol ships, any agency will be able to build a curation agent that packages inventory across publishers without touching SSP infrastructure. The function locked inside specific platforms becomes available to anyone who can build the agent. Where value gets created in the supply chain moves with it.

The wall nobody is talking about

The traditional model assumes inventory flows from publisher to SSP to exchange to DSP to agency to brand. In an agentic stack, that chain is under pressure at every joint simultaneously.

On the demand side, Adam Roodman (Yahoo’s head of DSP) said something in early 2026 that the industry has not fully processed: “We don’t care if you don’t use our UX anymore.” Yahoo is repositioning itself not as a platform you log into, but as infrastructure your agents plug into. The seat is no longer the point. The data and the signal are. On the supply side, PubMatic’s AgenticOS had run more than 30 fully autonomous campaigns with more than 1,000 direct publisher deals by May 2026, built deliberately on the supply side. They are not waiting for a DSP to arrive with an agent. They are building infrastructure for buying agents to come directly to inventory.

Meanwhile, every major walled garden has launched its own MCP server. Google. Meta. TikTok at TikTok World on 13 May 2026, with explicit intent to let AI agents plan, launch, and optimise campaigns without manual intervention. This looks like the wall coming down. It is not.

Every one of these is a campaign management interface, not an auction access point. An AI agent using TikTok’s MCP can adjust bids and shift budgets inside TikTok’s platform. It cannot negotiate TikTok inventory against a brief using AdCP or AAMP. The auction remains closed. The walled garden remains intact. What has changed is the front door.

The reason each platform is building its own MCP rather than plugging into an open standard was stated plainly by adtech expert Shirley Marschall in coverage of the TikTok launch: “If you route through someone else’s MCP, you lose visibility into how agents are querying you, what they’re comparing you against, what’s driving conversion. In an agentic world, that’s your most valuable signal, which you don’t want to hand to a third party.”

This is the data sovereignty thesis playing out in real time. The walled gardens are not opening. They are building agent-compatible walls. An external buying agent running on open standards today cannot reach Facebook inventory, YouTube inventory, or TikTok inventory through that agent. The framing of this as platforms “opening up” to agentic buying is the coverage getting it wrong.

There is one platform that sits outside this logic entirely.

When the surface is a conversation and the buyer is a machine, what does a brand actually buy?

Amazon: The exception

Every platform described above sits on one side of the data sovereignty line and protects it accordingly. Amazon does not.

Amazon sits simultaneously inside the walled garden and within the open stack. First-party purchase data at a scale no pure-play media platform can replicate. An open DSP that participates in the broader programmatic ecosystem. And Rufus.

Rufus is Amazon’s consumer-facing AI shopping agent. It sits at the point where consumers discover and decide. It is already integrated into Amazon’s monetisation and reporting layer. Advertisers can currently see Rufus prompt data inside Amazon’s own reporting. They cannot yet bid on it.

That gap, between visible and biddable, is the most important unresolved question in retail media. When Amazon opens Rufus to advertiser demand, they will have built something no other platform can replicate: an AI-native advertising surface inside a walled garden, at the point of purchase, with transaction data closing the attribution loop. Meta has user data but not purchase data. TikTok Shop is building toward this but is years behind at scale. Google has search intent, but AI search is disrupting it from within.

They are blocking nearly 50 AI crawlers from amazon.com via robots.txt while opening the MCP door. One door open, one door closed. This is not inconsistency. It is a deliberate access architecture: let agents manage campaigns inside our platform, on our terms, while controlling precisely what those agents can observe.

Amazon is not rushing to become the bridge between open-web agentic infrastructure and walled-garden inventory. But they are the only player structurally positioned to build that bridge if they chose to. Their moat is the retail shelf and the purchase data, not the advertising interface. Opening the interface does not undermine the moat. That asymmetry gives them a freedom of movement that Google, Meta, and TikTok do not have.

Whether they use it is the most interesting open question in the stack.

The question worth sitting with

The standard that governs agentic buying is being written now, in specifications and governance bodies that most practitioners in this industry are not reading. The companies that are reading them are shaping the rules: how agents transact, what data they have access to, what audit rights buyers retain, and what the cost of access looks like.

At every layer, the same question. Who is writing the rules? And what is your contingency if the layer you’re building on restructures without you?

The contingency is not a product decision. It is a position decision. Know which side of the data sovereignty line you’re on. Build your strategy from that reality rather than from the infrastructure you’ve always used.

Subscribe now

Part Three examines what this means for brands. Specifically: the agent will do exactly what you told it to do. What you told it to do may be the problem.

On 16 October 2025, a media buy cleared without a human making the call.

LG Ads inventory. Real money. An AI agent received the brief, broke it into tasks, checked brand safety, negotiated terms, trafficked the creative. The buy settled. Adam Broitman, watching for McKinsey, described what he’d just seen as “like watching the first banner ad being served.” It was the day after AgenticAdvertising.org was founded. One transaction. One protocol. One day.

This series is about what that transaction means. Not as an AI story. As an infrastructure story.

The Publicis and Trade Desk dispute settled one argument: the fight in programmatic media is not about transparency. It is about who controls the system that decides where money flows. I wrote about that fight a few months ago on this newsletter. The system being built underneath that dispute (the protocols, the agents, the standards) is what this series covers. Who makes buying decisions, how those decisions get made, and who is accountable when they go wrong. Not at some theoretical future point. Now. Live campaigns are running on it in Australia today.

The industry has mostly framed this as an AI story. Smarter optimisation. Faster creative. Better targeting. That framing captures some of what is happening and misses the point of most of it. What is actually changing is not how media gets bought. It is who, or what, is making the buy decision. That is a materially different kind of shift.

Three things that look similar but aren’t

The confusion in this market is understandable. For two years, “AI in advertising” has covered everything from a language model writing ad copy to a bidding algorithm chasing a CPA target. They are different things. Conflating them has made it harder to think clearly about any of them, and the conflation is costing people clarity they need right now.

Generative AI applied to advertising is when a language model produces a headline, a product image, or a six-second cutdown. A human reviews it. A human approves the buy. The machine made the asset. The human made the decision. Useful, fast, already embedded in every major platform. Not what this series is about.

Traditional programmatic is when a bid request goes out, a DSP bidding model returns a price, and an auction clears in roughly 100 milliseconds. No language model in the path, no reasoning about context, no decision carried forward to the next impression. The machine executes the rule a human set, at a speed a human cannot match. This is how the vast majority of digital impressions clear today, and has been for fifteen years.

Agentic buying is different. An LLM-backed agent receives a campaign brief in natural language. It breaks that brief into structured sub-tasks: find inventory, check brand safety, negotiate terms, traffic creative, monitor pacing. It calls tools to execute each step. It can push back on a media plan and decline a placement it calculates as off-brief. It carries context across a campaign flight rather than treating each impression as an isolated event. And it writes a transaction log at every step.

That log is the part the industry is not talking about enough. Every major dispute in programmatic advertising (the MFA inventory problem, principal media concerns, rebate opacity) comes back to the same root question: what actually happened, and who can prove it? Manual buying processes generate patchy, inconsistent records. Agent-to-agent infrastructure generates a complete, timestamped, machine-readable log at every step: which placements were considered, what was rejected and why, what price was paid, which creative ran. In a market where the central dispute has always been about what gets recorded and what doesn’t, that is not a neutral feature. It is leverage.

The question of who holds that leverage is not yet settled. The draft framework governing agentic buying — the ARTF, published by the IAB Tech Lab for public comment — specifies that all agent inputs and outputs are mediated and audited by the “host”: the platform operating the infrastructure. Not the buyer. Not the agency. The platform. The final version of the ARTF has not been released. But if the draft holds, whoever operates the infrastructure owns the log — and owns the evidentiary record of how client money moved.

How fast this moved

That October transaction, one day and one protocol, was followed five weeks later by the IAB Tech Lab putting its Agentic Real-Time Framework out for public comment.

In January 2026, PubMatic launched AgenticOS at CES with more than twenty agents running on NVIDIA-accelerated infrastructure. Yahoo DSP launched its “Yours, Mine, Ours” framework on 7 January, live in Australia and Singapore from day one. Amazon opened its MCP server (the specification governing how AI agents call external tools and data sources) to 36 countries. In March, Google deepened Gemini’s integration into DV360. In April, The Trade Desk launched Koa Agents. PubMatic confirmed live agentic campaigns running in Australia. An IAB survey conducted through early 2026 found two-thirds of US ad buyers planned to increase their focus on agentic buying across the year.

From the founding of a governance body to live campaigns running in Australia: under six months.

The advertising industry has a well-documented habit of announcing infrastructure changes that take years to show up in actual buying behaviour. Header bidding. Attention metrics. The addressable TV promise. All real. All slower than the announcements suggested. The difference here is the money is already moving. That is not the same kind of announcement.

What’s real and what isn’t

Most of what is live today is assisted workflow automation, not fully autonomous market-making. Agents are doing real work: trafficking, pacing anomaly detection, post-buy log ingestion, invoice reconciliation. Valuable work. It is also the connective tissue between functions, not the strategic decisions at the top of the stack.

The credentialing campaigns are real transactions. Butler/Till’s autonomous CTV buy in December 2025. NBCUniversal’s first AI-agent-led programmatic-guaranteed deal. Real money, real agents, real inventory. Also small, closely supervised, and better understood as proof-of-concept than proof-of-scale.

The holdco “agentic operating system” announcements should be read as positioning for the next pitch cycle, not a description of how the majority of impressions clear today. That is not a criticism. It is an accurate description of where the industry is.

One thing is worth stating plainly: the framework governing these transactions has not been finalised. Public comment on the ARTF closed in January 2026. No ratified version has been released. Live campaigns are running on a draft. The infrastructure is ahead of the standard that is supposed to govern it, and that matters because the standard is where the rules on accountability, log ownership, and buyer rights will eventually sit. Those rules are still being written, by the same platforms already running campaigns under the draft.

The trajectory: the next twelve months close the gap between assisted and autonomous on specific use cases. Programmatic-guaranteed CTV. Always-on measurement-to-activation loops. Genuine end-to-end autonomy across a diversified media plan is a 2027 problem at the earliest. But the infrastructure being built now will determine who is positioned when that autonomy arrives and who is not.

What this means in Australia and New Zealand

The global briefings tend to frame ANZ as a follower market. On agentic buying specifically, the reality is more complicated.

PubMatic’s AgenticOS went live in Australia in April 2026, ahead of most of EMEA outside France and the Netherlands. Yahoo DSP launched agentic capability in Australia and Singapore at general availability, not as a delayed regional rollout. Amazon’s MCP server is open in Australia. IAB Australia formally adopted the IAB Tech Lab roadmap and committed its Standards Council to local pilot work.

The supply side is consolidating faster than the buy side is adapting. The Seven West Media and SCA merger in January 2026 concentrated national TV, BVOD, radio, and audio into fewer counterparties. When agents start negotiating at scale, the leverage dynamics in a concentrated market are different from the fragmented open-web environment most agentic infrastructure is being designed for. The global platforms are building for the US open web. Australian buyers are going to run that infrastructure against a fundamentally different market structure.

Local operations built for specific market conditions (commerce media networks, regional trading relationships, market-specific infrastructure) have found themselves absorbed into global AI restructures with timelines and logic set from platforms headquartered elsewhere. The decision-making process is not hostile to local markets. It is not designed to see them. That distinction matters: the outcome is structural, not personal, and it will keep repeating as long as the infrastructure being built is optimised for global open-web conditions rather than concentrated local ones. The risk for ANZ operators is not that global platforms dislike them. It is that they are not legible to the system making the call.

A regulatory thread the global conversation is missing: automated decision-making transparency obligations under the Privacy Act land in December 2026. An ACMA AI-disclosure rule for commercial radio takes effect on 1 July. These are local constraints the platforms building agentic infrastructure are not designing for. That thread comes back in the final piece in this series.

What ANZ does not yet have: a domestic DSP, SSP, or major broadcaster with a public commitment to either of the two standards governing how agent-to-agent buying works. That gap matters more than it might appear. The risk is not being left behind by the technology. The risk is becoming a reseller of decisions made by global platforms rather than an architect of local ones.

That risk is the same one sitting at the centre of the global picture, just with less runway to respond to it.

What the system is actually for

Every platform building agentic buying infrastructure is making the same argument: the machine will make better decisions than the human. Faster, more consistent, less subject to attention limits and cognitive load. That argument holds in narrow execution contexts. It breaks down at the level of strategy.

The buying agent optimises toward the objective it was given. That objective was written by a human, inside a contract negotiated by humans, based on a brief built on measurement systems designed by humans. If any of those inputs carry the wrong incentives (misaligned principal-media relationships, flawed attribution, a brief that confuses execution metrics for business outcomes), the agent executes that wrongness faster and at greater scale than a human would.

Start with the baseline agents are inheriting.

The ANA released a study in March 2026. Ninety percent of marketers’ top concern was whether the media being recommended to them was actually in their interest. Only 57% had written guidelines on it. That is the state of human-governed buying before agents arrive. Agentic buying does not solve that problem. It inherits it, and runs it at machine speed.

Agents amplify the system they are given. They do not correct it.

The infrastructure for agentic buying is being built now, by platforms with their own commercial interests, on standards being written by consortiums representing specific parts of the industry. The system the agents will operate inside is being designed without most of the brands and agencies who will eventually depend on it having a seat in the room.

Right now, the box is drawn by comprehension as much as by commercial interest. The platforms, the infrastructure providers, the standards bodies, and parts of the holdco world understand what agentic infrastructure is and how it works. Most independent agencies and brands do not. That gap — not just the commercial one, but the comprehension one — is what determines who has a say in how the system gets built.

That is not an AI question. It is the question this newsletter has been asking since the beginning: who draws the box, and what happens to everyone else when the machine starts working inside it.

This series is one answer to that question.

Subscribe now

Part Two gets into the infrastructure itself — the two competing standards governing how agent-to-agent buying works, what is actually live in market, and the largest unresolved gap in the stack.

That buyer is not unusual. They are the norm. Part One mapped the market that keeps them there: the duopoly, the challengers, and the payments layer. This edition names the five structural reasons the market is built to obscure performance. Then it shows what to do about it anyway. The central argument is this: the payments networks arriving in ANZ are not additional ad products competing for the media budget. They are independent measurement infrastructure. And they are about to audit everything the grocery networks have been reporting.

I know this from the inside. I led the launch of The Wallet Advantage at Afterpay Ads in November 2025, making it the first payments-native commerce media product formally positioned in the ANZ market. The project ended before it reached full commercial scale. That experience informs the argument in this edition.

Five reasons the market is built this way

The gap between what gets reported and what is true is not an accident

The gap between reported ROAS and incremental ROAS is concrete. A mid-tier FMCG brand running a standard sponsored-product campaign across a major ANZ grocery network will typically see ROAS reported at 3 to 4x. A geo holdout on a comparable campaign, using Circana scan data as the measurement currency, routinely returns 1.4 to 1.8x incrementally. That gap is not a measurement error. It is the intended output of a system optimised to report its best number. Five structural problems make that system work the way it does.

1. Pricing without a benchmark

The most common retail media valuation methodology, a fixed percentage of GMV, is, in Sonder’s words, “a blunt instrument that has wreaked untold havoc for retail media teams.” Networks built on GMV benchmarks set revenue expectations that actual advertiser demand does not meet. The typical response is more inventory at lower rates, which degrades what makes retail media worth buying in the first place.

Sonder’s 2025 Global Market Report finds that 60 percent of owned media businesses globally have no rate card. Buyers negotiate blind against a number that does not exist. Separately, 57 percent are giving owned media to partners at no cost, surrendering $48 million in media value per Retail Aggregator business per year. Not earning it. Giving it away. Less than 30 percent of the $573 billion global owned media opportunity is currently being captured. Over-valuation and under-valuation are both symptoms of the same root cause: there is no independent pricing benchmark. The number the network quotes you is not anchored to anything you can check.

2. Measurement graded by the seller

Every major retail media network in ANZ measures its own campaigns. The attribution methodology is defined by the network. The reporting is produced by the network. The result is delivered to the brand by the network. The conflict of interest is structural, not incidental.

A network interpreting an ambiguous methodology choice, whether to use a 7-day or 14-day attribution window, whether to count view-throughs alongside clicks, has a commercially rational incentive to favour the interpretation that produces a higher ROAS. That is not corruption. It is incentive structure. The ACCC documented the downstream consequences: brands unable to independently verify attribution claims, investment decisions made on ROAS figures with no external validation. The Coles 360 and Circana partnership is the strongest response any ANZ network has made toward accountability. It is not sufficient and it is not universal.

3. The incrementality gap

Problem 2 is about who controls measurement. This one is about what that control conceals. Closed-loop attribution proves correlation. Incrementality measurement proves causation. The retail media industry has sold the former as the latter for five years. Incrementality testing requires a control group, a comparison of purchase rates between exposed and unexposed customers, and an honest accounting of what would have happened without the campaign. None of these steps are in the standard retail media reporting suite. Operationalising them at scale reduces reported ROAS by an estimated 30 to 70 percent depending on the category and network.

The global market has moved. TransUnion and eMarketer research from July 2025 found 52 percent of brands and agencies globally are now using incrementality testing. Google cut its Conversion Lift minimum threshold from approximately $100,000 to $5,000 in November 2025. The online testing tooling exists and is now accessible at meaningful scale. The harder problem for FMCG and grocery categories is offline: capturing in-store sales requires integrating loyalty or point-of-sale data, a step most ANZ campaigns have not taken. The technical barrier is falling. The integration work is the remaining gap.

There is a second dimension to this problem that the industry has not named clearly. April 2026 research from In-Store Marketplace, based on interviews with brands, agencies, and RMN executives across the US and UK, found that in-store media has suffered from what it calls “digital envy”: the assumption that it must replicate one-to-one, closed-loop digital attribution to be credible. The result is a channel evaluated against a framework it was never built for, while the methodologies that actually work for physical retail — matched-market tests and pre vs. post analyses — are treated as second-rate rather than fit-for-purpose. The same research identified four competing scorecards operating simultaneously across agencies, merchants, RMNs, and brand shopper marketing teams. When those scorecards do not align, measurement gridlock occurs regardless of technical capability. The incrementality gap in ANZ is partly a testing gap. It is also a stakeholder alignment problem.

4. When media and distribution merge

The ACCC inquiry and AFGC submissions describe the environment in institutional language: brands presented with media investment requirements as a component of stocking and promotional negotiations. The plain version is this. It is a cost of distribution rendered in media language. The ROAS reported on a bundled investment is not a media performance number. It is a commercial cost with a denominator attached.

The brands in this position face a collective action problem. The first to resist publicly pays the full cost alone, losing shelf priority, promotional placement, and commercial goodwill, while the benefit of establishing a precedent is shared across every brand in the category. No individual brand’s incentive calculation supports being first. The result is an industry where the measurement question is asked in an environment where the buyer knows the answer may affect stocking decisions that extend far beyond media performance. The ACCC inquiry has created a public record. How networks respond to that record will shape the next phase of commercial negotiations.

5. The franken-campaign

Each of the four preceding problems compounds in the multi-network campaign. The incrementality gap is present in each network separately and invisible in the aggregate. Self-reported attribution from five networks with five different methodology frameworks produces five clean numbers that add up to nothing coherent. Budget coercion at the individual network level means the media plan may never have been built on genuine ROI. Inflated valuations stacked across networks inflate the total investment against performance that nobody is measuring in a common currency. The Franken-campaign is not a fifth, separate problem. It is what the first four look like when they are running simultaneously.

A brand running across Cartology, Coles 360, CommBank Connect, DoorDash, and Amazon Ads has no single number that describes overall performance. Audience overlap is unmeasured. Incremental reach from each additional network is unknown. Reaching the full grocery delivery audience in ANZ requires buying both DoorDash Ads (Woolworths and Aldi shoppers) and Uber Ads (Coles shoppers), two incompatible measurement frameworks for what looks like one audience segment in a media plan. Every new launch, Myer with 5.1 million Myer One members, Virgin Australia on Velocity data, adds another incompatible framework to an already incoherent landscape.

Amazon Ads is the exception worth noting. It is the only retail media network in that list that ships measurement infrastructure as standard. Amazon Marketing Cloud gives advertisers the tooling to interrogate fragmentation, run custom audience analysis, and conduct incrementality measurement within Amazon’s signal environment. The fragmentation problem is still real. But Amazon’s approach to it is structurally different from the grocery networks that produce a single clean ROAS number and stop there.

The Franken-campaign will not be solved by the networks. It will be solved, if it is solved, by the independent infrastructure layer: whole-wallet attribution from payments networks, Circana validation extended across the market, or a regulatory reporting standard above a revenue threshold.

Subscribe now

What to do with what you now know

Five moves for buying in a market built against you

The data assets are real. The audience advantages are structural. None of that is in dispute. What is in dispute is whether the measurement honestly reflects them. A structural advantage and a system that accurately accounts for it are not the same thing. They have been treated as the same thing for five years. That is the problem. Here is what to do about it.

1. Audit for incrementality before you scale

The right test design depends on your category. For FMCG and grocery, where most purchases happen in-store, run a geo holdout with independent sales data as the measurement currency. Select matched geographic markets, suppress the campaign in the control group, and use Circana or Nielsen scan data, or retailer loyalty card data, to compare actual purchase rates. This approach captures offline sales by design and produces a number that reflects what happened at the register, not what was visible in a digital attribution window. Meta’s GeoLift tool is open source and built specifically for this design. For e-commerce and digitally-purchased products, Google Conversion Lift, minimum now $5,000, gives a methodologically sound starting point using randomised user assignment rather than last-touch attribution. Note the structural limit here: Google is measuring the effectiveness of Google campaigns. The conflict-of-interest problem does not disappear just because the methodology is sounder than last-touch. Independent sales data is the correct measurement currency for in-store categories. For those, the geo holdout is the standard, not a supplement.

The test will tell you three things: whether your ROAS holds up incrementally, whether you’re paying for customers who were going to buy anyway, and whether you’re cannibalising organic purchase while paying for attribution credit. None of those answers are comfortable. All of them are more useful than continuing without knowing.

2. Demand independent measurement as a contract condition

Networks confident in their attribution methodology have nothing to fear from independent validation. Make external measurement a procurement condition, not an optional upgrade. If a network refuses, that refusal is itself a signal. The Circana standard established by Coles 360 is the current ANZ benchmark for grocery. Price the risk of self-reported attribution into your investment decision accordingly.

3. Benchmark your rate cards against independent data

Sonder’s 2026 report publishes the first credible bottom-up commercial potential benchmarks for ANZ by network type: major grocery networks at $102M, omnichannel at $61M, telco at $49M. These are not revenue figures. They are bottom-up estimates of what each network type is capable of earning if its owned media is properly valued and activated, derived from a $12 billion rate pool across 2,500+ CPMs. I should note that Sonder’s commercial model involves selling benchmarking and valuation services to retail networks. Their methodology is transparent and the underlying data is independently derived; the commercial interest is relevant context, not a disqualifying conflict. If the network you’re buying quotes materially above these benchmarks for comparable inventory, you have a negotiating data point. If it quotes below, you have a signal about either rate strategy or inventory quality.

4. Solve the Franken-campaign before adding a network

Establish a single measurement framework for the networks you already buy before expanding your footprint.

Amazon Marketing Cloud deserves more than a passing mention here. AMC is the most measurement-mature clean room available to ANZ marketers today. It runs custom SQL across Amazon’s full signal set, supports incrementality analysis as a native capability, and allows advertisers to bring their own first-party data in for matched-audience measurement across both Amazon and non-Amazon exposure. The conflict of interest is still present and worth naming: AMC measures Amazon campaigns with particular precision, and Amazon has a commercial interest in that result. That context belongs in the room. But the tooling is built to a standard that the ANZ grocery networks have not matched, and brands running Amazon Ads without AMC are paying for measurement capability they are not using. The constraint is not access. It is the data engineering investment required to connect it. Build that capability before you need it to resolve a disputed ROAS number, not after.

Snowflake Data Clean Rooms and the Criteo-Amperity architecture are also operational in ANZ for brands managing cross-network measurement outside Amazon’s environment. Every new network you add without solving the measurement fragmentation deepens the incoherence. Add networks after you can measure them, not before.

5. Build your measurement architecture for the payments networks now

The Wallet Advantage at Afterpay Ads was the first payments-native commerce media product formally positioned in the ANZ market, launching in November 2025. It is no longer operational. Mastercard Commerce Media will follow in 2026. Visa Audiences is already active in ANZ programmatic. The brands that have built incrementality testing and clean-room measurement infrastructure into their buying practice will be positioned to use payments data as an independent measurement arbiter. The brands that have not will be buying on self-reported attribution from networks that a third-party data source now contradicts. Build the infrastructure before the contradicting data arrives, not after.

Three things need to be in place before payments networks can serve as an independent measurement arbiter. First, establish an incrementality baseline now. Visa Audiences is active in ANZ programmatic today and can serve as that baseline source: a number you own, measured against a methodology you control, before Mastercard Commerce Media is fully operational and before a contradicting result arrives in a briefing room. Second, connect clean-room access. Snowflake Data Clean Rooms and the Criteo-Amperity architecture are both live in ANZ. Connect your CRM or loyalty data to one of them before you need to cross-reference a disputed ROAS number, not after. Third, agree internally on what “independent measurement” means before a contradicting number arrives. This sounds procedural. It is not. If procurement, media, finance, ecommerce, and category teams do not agree on what counts as independent evidence, every contradicting result becomes a political argument. The network has one number. The agency has another. The category buyer has a third. Finance does not trust any of them. The brands that define the process first will use payments data as leverage. The brands that do not will dispute the result instead..

The measurement framework we were building at Afterpay Ads tracked incremental lift across GPV, total sales, AOV, share of wallet, LTV, and new vs. returning buyers. That is not a media metric stack. It is a business performance stack. We were close. The project ended before it launched. First-mover advantage in payments commerce media requires institutional backing to compound. The data asset and the positioning exist. The execution capacity does not. The brands that build measurement infrastructure now are entering a relationship with payments networks that are still developing their commercial scale. That is the right moment to establish a baseline and set the terms. The moment that follows it is not.

Retail Media Won The Lower Funnel. The Question Is Whether It Deserved To.

The ANZ retail media market in 2026 is at the end of its first chapter. The category was built on the promise of closed-loop measurement. The second chapter will be built on whether those loops actually closed.

The dominant grocery networks in ANZ hold genuine data assets: longitudinal household purchase history, in-store screens, and supplier relationships built over decades. Those distribution moats are real and durable. The measurement accountability gap is equally real. A distribution moat and an independent account of whether the media worked are not the same thing. The networks that close that gap first will hold the stronger position. The ACCC has applied the regulatory pressure. The payments networks will apply the data pressure. The first major brand to run a geo holdout, publish the result, and reallocate accordingly will apply the social pressure. The sequence is not speculative. It is in motion.

No grocery network is exempt from this dynamic. The payments networks arriving in ANZ are not additional options for a media plan. They are whole-wallet attribution infrastructure that will make every grocery network’s closed-loop claims look like a subset. When Mastercard Commerce Media is fully operational in ANZ, CommBank Connect’s positioning does not gradually evolve. It flips. From an interesting ANZ experiment to a local node in a globally scaled measurement architecture that sees every transaction across the wallet.

The networks that solve incrementality first will own the second chapter. Everything else is distribution.

The structural problems across these two editions are not the problems of a mature market encountering its limits. They are the problems of a young market that has grown faster than its accountability mechanisms. That gap will close through industry discipline or regulatory intervention. By the end of 2026, at least one major ANZ brand will publish the results of a geo holdout against a grocery network campaign. When that happens, the market’s relationship with self-reported attribution will not recover. The brands waiting for the industry to solve this on its own behalf are waiting for an incentive structure that does not exist. The work of building measurement accountability into retail media investment is buyer-side work. It always was. That category buyer who knows the number but cannot say it in the meeting? The payments networks will say it for them.

I spent the last year of my career building the product that was supposed to say it first. The gap between what brands think is working and what is actually working is real. We measured it. The project ended. The gap did not.

If you have run a geo holdout against a retail media campaign in ANZ and have results, I want to hear what you found. Real numbers from real campaigns are the fastest way to move this conversation forward. The more buyers share, the harder it becomes for the market to stay comfortable with self-reported attribution.

Subscribe now

Systems That Decide publishes on the infrastructure of decision-making in marketing and advertising. If this edition was useful, share it with someone who is still treating the seller's number as the real number.

The Argument.

Retail media sold the lower funnel on one promise: we can show you who bought after seeing your ad. For five years, brands believed it. Nobody asked whether the ad was the reason.

That is not a small oversight. It is the structural fault line underneath a $5 billion ANZ market. Closed-loop attribution proves a customer bought. Incrementality measurement proves the ad caused it. The industry has spent five years selling the first as if it were the second, and it has had no commercial incentive to stop.

Networks have no reason to introduce incrementality testing at scale because the answer consistently reduces apparent ROAS. The advertiser buying a reported 4x ROAS campaign discovers they were running a 1.8x incremental ROAS campaign after proper testing. The gap between those two numbers is the gap between measurement that looks good for the network and measurement that actually helps a brand make decisions. That gap is not incidental. It is the product.

The closed loop closes. The attribution story is irresistible. The incrementality question goes unasked.

The data assets underneath this problem are real, and I want to be precise about that before making the critique. Cartology’s 12 million active Everyday Rewards members, Coles 360’s 10.1 million Flybuys members, CommBank’s 17 million customers whose transaction data crosses every merchant category, and the 7.2 million active script patients in the Terry White Chemmart database: these are structural data advantages that traditional digital advertising could not touch. The brand buying a grocery retail media campaign reaches customers with verified purchase history in the category. That advantage is real. The measurement architecture built on top of it is not yet fit for the claims being made.

What is specific to ANZ in 2026 is that the buyer is more trapped here than in any comparable market. In the US, a brand that discovers its ROAS is inflated can reallocate across a dozen networks. In ANZ, with two grocery networks both under regulatory scrutiny, the brand that discovers its ROAS is inflated has to negotiate. With the same network that produced the number it now doubts, under the same ranging dynamic that shaped the budget to begin with. The incrementality argument is not new. The tightness of the trap in ANZ is.

The networks that solve this gap first will own the next decade of brand investment. The ones that do not will face a reckoning from three directions. The ACCC has created a paper trail on coercive investment practices. The payments networks are arriving with whole-wallet data that makes the grocery attribution claim look like a partial view. And a growing cohort of brand buyers is starting to make the incrementality question a procurement condition, not a polite briefing-room inquiry. That third pressure is the quietest one now. It will not stay quiet.

Part One maps the market as it actually stands: who holds what, what is arriving, and why the arriving event changes what closed-loop measurement means for every player in the stack.

The Market As It Actually Stands

A field assessment of every major ANZ player

The ANZ market organises into five layers. Two players hold structural duopoly in grocery. A second tier of challengers holds specific data advantages the duopoly cannot replicate. A payments layer is arriving that changes the measurement stakes for every player above it. An infrastructure layer quietly runs the pipes beneath all of it. What follows is a plain account of where each player sits, what they hold, and what they face.

Tier One: The Grocery Duopoly

Cartology holds the largest data asset in ANZ retail media: 12 million active Everyday Rewards members, longitudinal household purchase history, in-store screens, and catalogue relationships no other network can match. It also carries the market’s most acute accountability exposure. The ACCC documented what practitioners have known privately: retail media investment at Cartology is not always a freely made marketing decision. Supplier ranging negotiations and media spend have been bundled in ways the ACCC has now put on the public record. Cartology grades its own homework with no independent external validator for its attribution methodology.

Coles 360 has 10.1 million Flybuys members and made one strategically significant decision that separates it from Cartology: the Circana measurement partnership. Independent third-party validation is a materially different commercial proposition. The gap Coles 360 has not yet closed is extending that partnership to incrementality testing. Validated attribution and validated incrementality are different things. Coles 360 has the foundation to own the measurement credibility argument in ANZ grocery. It has not yet built the house.

Tier Two: The Challengers

Chemist Warehouse built the original pharmacy retail media model to $638 million in annual advertising revenue by 2022, then watched the market it created compete with it. Revenue declined by approximately $150 million between 2022 and 2024. The ACCC named it alongside Cartology and Coles 360 for supplier transparency concerns. The company that built near-monopoly conditions is now defending market share in the category it invented.

LocalEyes (Metcash) aggregates 6,500 independent stores into real regional reach the duopoly cannot replicate: 25 times screen growth in 15 months. The critical path is incrementality measurement. Without it, LocalEyes will always be a trial allocation. The moment it can produce credible incremental reporting, its positioning as the independent alternative becomes commercially actionable. That is the story to watch.

DoorDash crossed $1 billion global ads revenue and holds exclusive grocery delivery relationships with both Aldi and Woolworths in ANZ. The Aldi angle is the most underreported: Aldi’s value-oriented, fiercely loyal customer base is inaccessible through Cartology or Coles 360. DoorDash now holds it. Coles runs its grocery delivery exclusive through Uber Eats, which means Coles shoppers on delivery are accessible through Uber Ads, not DoorDash. The two major grocery networks have split the delivery audience across two competing platforms with incompatible measurement frameworks.

CommBank Connect is, in my view, the most globally significant ANZ retail media story the local industry has undervalued. Seventeen million customers. Transaction data across every merchant category. And, critically, a Mastercard partnership. Mastercard Commerce Media launched globally in October 2025 and named ANZ as a 2026 expansion market. When Mastercard arrives formally, CommBank Connect stops being an interesting ANZ experiment and becomes a local node in globally scaled commerce data infrastructure. The local industry has not made this connection explicitly. It is the most important strategic development in the category.

60% of Owned Media Businesses Have No Rate Card. This Is What That Looks Like.

Underneath the payments layer, the existing market has a structural pricing problem the industry has not fixed. Across global owned media businesses, the numbers from Sonder’s 2026 research are precise about how broken the valuation architecture actually is.

The ANZ market is not an outlier in these figures. It is a subset of them. Networks without rate cards, handing inventory to suppliers at no cost, operating on GMV-percentage valuations that no buyer actually meets: this is the existing dysfunction. The payments networks arriving with whole-wallet data are not arriving into a healthy market. They are arriving into one that cannot currently price its own assets.

The Network That Sees Your Whole Wallet Is or Was Already Here

I should disclose my position here. I led the launch of The Wallet Advantage at Afterpay Ads in November 2025, making it the first payments-native commerce media product formally positioned in the ANZ market. The pitch was straightforward: Afterpay sees every BNPL transaction across fashion, beauty, home, and electronics, which are categories the grocery networks cannot access, whilst also seeing purchases across the grocery networks themselves. A brand buying Afterpay Ads inventory reaches a younger, lifestyle-oriented customer base with verified cross-merchant transaction history. The grocery attribution claim covers a fraction of a customer’s wallet. The Wallet Advantage said: we see a different more broader slice, and we can prove it.

First-mover advantage in retail media requires institutional backing to compound. Positioning and data assets are necessary but not sufficient. Without the commercial resource to execute at scale, a first-mover position becomes a proof of concept for whoever arrives with deeper backing. That is the structural dynamic the institutional entrants, Mastercard and CommBank, now step into.

The players who can capitalise are Mastercard and CommBank. Mastercard Commerce Media launched globally in October 2025 with ANZ named for 2026 expansion. Visa Audiences is already active in ANZ programmatic. These are not startup experiments. They are institutional entrants with global infrastructure, and they are not arriving into a category they have to explain. Afterpay already explained it.

The direction of travel sharpened further on April 27, 2026. PayPal launched PayPal Ads ID in the US: a deterministic identity layer built on 400 million verified accounts and 25 billion annual transactions, linking ad exposure to actual purchase. Rokt was named as launch partner. PayPal Ads is not yet operational in ANZ. That is directional evidence, not a market event here, but when a payments network with that transaction footprint repositions as attribution infrastructure, it changes what closed-loop measurement means for every other player in the stack.

The grocery network that can see ten percent of a customer’s wallet will eventually have to account for itself against the network that sees all of it.

The market above has five structural problems the industry has not solved. Part Two names them in order of how they compound, and gives practitioners a specific framework for acting inside a market that is designed to obscure performance. If you are buying retail media in ANZ and have not yet run an incrementality test, Part Two gives you a precise framework for where to start. The question worth sitting with before you get there: what is your current network asking you to take on faith?

Next Edition:

Part Two: Five Problems. Five Moves.

What the market hasn’t fixed and what you should do about it.

---

Not a little. Not on the margins.

Your marketing mix model (MMM) and your multi-touch attribution system (MTA) are measuring the same spend. Under current privacy-constrained conditions, simulation work puts the correlation between their channel-effectiveness rankings at r=0.02. If that number means nothing to you: zero is no relationship, 1.0 is perfect agreement. These two systems score 0.02. That’s not a rounding error. That’s noise.

This sits at the centre of marketing measurement in 2026. Not that we can’t measure. We measure obsessively. But the dashboards contradict each other, the platforms reporting the numbers have a financial interest in making them look good, and the methods we’ve relied on for a decade are collapsing under the weight of privacy regulation they were never designed to survive.

You are drowning in dashboards and starving for truth.

This edition is about why that happened, what’s actually broken, and what it looks like when you fix it.

---

The signal collapsed faster than anyone planned for.

The attribution systems most marketing teams built their infrastructure on were designed for a world of persistent cookies and cross-site tracking. That world is gone.

iOS’s App Tracking Transparency runs a 74% opt-out rate and the third-party cookie is functionally dead in most browsers. Modelling of the current privacy landscape puts average cross-channel data completeness at around 46%. Less than half.

You’re trying to do precision attribution with a dataset that’s missing more than half its observations. The models don’t tell you that. They just give you numbers.

---

The platforms grade their own homework.

Here’s the structural problem sitting underneath the data quality problem.

The companies you pay to run your media are the same companies that tell you how well your media performed. That’s the conflict. It’s structural, not personal.

The numbers they report are systematically higher than what independent measurement finds.

Real-world meta-analysis of walled garden platforms consistently finds systematic over-attribution in the range of 17–35% when platform-reported ROAS (return on ad spend, meaning the revenue your campaigns generate per dollar invested) is compared against independently verified incremental returns. The walled gardens sit at the higher end of that range; Connected TV, with cleaner exposure data, tends to sit lower. The direction is consistent. The magnitude is material.

This isn’t fraud. It’s the natural output of a system where the incentives aren’t aligned with yours. Attribution windows, model assumptions, the definition of a conversion: these are methodology choices. In a system where the measurer and the media owner are the same entity, those choices will trend in a predictable direction.

When you see a 4.2x ROAS in Ads Manager (meaning the platform claims you earned $4.20 back for every $1 spent), you’re looking at a number produced by a system where the incentives point toward a high number. That’s not a small caveat. That’s the whole game.

---

Two systems, zero agreement.

The r=0.02 correlation between MMM and MTA outputs on channel effectiveness rankings is the clearest expression of this. Two independent measurement systems, analysing the same spend, producing essentially random agreement on which channels are working. That finding comes from simulation modelling, but it’s consistent with what any practitioner who has run both methods on the same client will tell you off the record.

Both systems are trying to answer the same question: which channels are driving outcomes? If they were measuring reality, they should agree. They don’t.

They disagree because they have fundamentally different failure modes.

MTA can only measure what it can track (which is increasingly little) and it’s structurally biased toward conversion-stage channels that leave visible digital signals. The last click, the final retargeting ad, the search term. Awareness spend, offline media, brand activity: invisible.

MMM can see the broader picture, but it’s a statistical model that requires a lot of data and a lot of judgment. The judgment lives in the assumptions baked into the model: what it’s told to believe before it looks at the data. A poorly built MMM can tell you almost anything you want to hear.

Used alone, each method is dangerous. Used together, they contradict each other.

Retargeting is the cleanest illustration. Across hundreds of incrementality experiments, retargeting repeatedly claims more credit than it deserves. The attribution system credits the ad. The incrementality test often shows the conversion was going to happen anyway. Same spend, fundamentally different verdicts.

---

The iceberg.

Most attribution captures what’s visible above the waterline: direct response, trackable conversions, last-click events.

Below the waterline: brand awareness, earned media, word-of-mouth, the halo effect of being recognised before a customer is even thinking about buying. Everything that makes your paid performance channels work better because your brand already exists in the customer’s mind.

Research suggests a significant share of marketing’s actual effect, conservatively more than a quarter, sits in channels that even sophisticated measurement frameworks struggle to attribute: social, affiliate, influencer, awareness video. The activity is happening. The impact is real. The number in your dashboard is zero.

This is why budget optimisation based purely on reported attribution systematically underfunds brand and awareness activity. You’re cutting the stuff you can’t see. You’re rewarding the stuff that’s easy to measure, not necessarily the stuff that’s working.

---

What good looks like.

The answer emerging from the serious measurement work (academic, practitioner, platform-agnostic) is triangulation.

Not a single source of truth. Three methods with different failure modes, used together to converge toward something closer to reality.

Marketing mix modelling for strategic allocation. MMM sees the whole picture, including offline media and brand effects, but at low resolution.

Multi-touch attribution for tactical optimisation within channels, where data completeness is sufficient. High resolution, narrow field of view.

Incrementality testing, specifically geo-lift experiments, as the causal check on both. The only way to know if your MMM or your MTA is telling the truth is to run controlled experiments that isolate the actual incremental effect of turning spend on or off.

None of these methods is sufficient alone. Each anchors the others.

---

Where you are on the journey.

A useful way to locate yourself is the five-level measurement maturity ladder that’s become common currency in the MMM space, popularised by Mutinex. Most organisations, when they’re honest about it, land on Level 1 or 2.

Level 1: Platform-reported metrics. What the platforms tell you, taken at face value.

Level 2: Multi-touch attribution. A more sophisticated view of the customer journey, still dependent on trackable signals.

Level 3: Experimentation. Incrementality tests that give you causal evidence for specific questions.

Level 4: Marketing mix modelling. A strategic-level model integrating all channels, validated against experiments.

Level 5: Measurement-Ready Organisation. Data is governed, decision-grade, and the measurement infrastructure drives actual budget decisions.

The tools to reach Levels 4 and 5 are more accessible than they’ve ever been. The barrier isn’t technical.

Research across 197 senior marketing leaders puts it plainly: the confidence gap in measurement is “as much human and organisational as it is technical.” The data is there. The question is what counts: who decides, based on what, and whether the measurement actually changes the decision.

A governance question. Not a data question.

---

Not all MMMs are equal.

The generation of marketing mix models most companies are running is a regression model fit to historical data. It produces a coefficient for each channel: essentially a score representing how strongly that channel appears to drive outcomes. The channel with the highest score gets more budget.

The problem: a model fit to historical data tells you what worked in the past, at past spend levels, in past market conditions. It doesn’t tell you what happens if you double spend in a channel, if the market shifts, or if you enter a new geography.

The better approach, now accessible through open-source tooling and specialist vendors, is a Bayesian MMM. Bayesian just means the model deals in probability and uncertainty rather than false precision: it updates as new evidence comes in, and it tells you how confident it is rather than pretending it knows. It models diminishing returns (the point where spending more in a channel stops paying back at the same rate) and carryover effects (the way ad exposure keeps working even after a campaign ends). Crucially, it’s calibrated against real experiment results, so its outputs are grounded in what you’ve already tested, not just assumed.

The practical difference: a standard model tells you a channel returned 3.2x and presents that as fact. A good Bayesian MMM tells you the likely return sits between 2.4x and 4.1x, with 70% confidence, and here’s where that confidence gets thinner. A range, not a point. Honest, not precise-but-wrong.

This matters enormously for the CFO conversation. “Spend another $2M in brand” is a hard sell.

“Our model gives us strong confidence in a positive return over 18 months, calibrated against geo experiments, with a range we can defend” is a different conversation.

It’s still uncertain.

But it’s honestly uncertain.

The open-source options: Meta Robyn, Google Meridian. Commercial: Recast, which builds causal structure directly into its models and integrates geo-lift testing; Mutinex, an Australian-built platform now working with brands including Optus, ING, SEEK, Asahi and Domino’s. The local market has more mature tooling than most teams realise.

The Iconic is the clearest local proof. In February 2024, while Shein and Temu were extracting nearly $3 billion from Australian fashion retail and traditional competitors were exiting the market, The Iconic shifted investment toward brand. The ‘Got You Looking’ masterbrand campaign drove a 34% increase in unprompted brand awareness over 12 months. ANZ NMV (net merchandise value, meaning the total value of goods sold through the platform), which had declined 12.6% in Q4 2023, grew 9.4% in Q4 2024. GFG’s CEO credited the brand reset directly. That’s the commercial argument MMM is built to make. In this case, the numbers moved with it.

---

The geo experiment problem.

Incrementality testing sounds clean in theory. Turn spend on in some markets. Turn it off in others. Compare.

In practice, it’s harder in Australia than almost anywhere else.

Australia has eight states and territories, several of them very small. Major metro populations are dense and concentrated. Classical randomisation (split markets cleanly, assign to treatment and control) doesn’t produce reliable results at the scale most Australian brands can actually run.

The practical fix is synthetic control methods: tools like GeoLift and CausalImpact that construct a statistical comparison market from a combination of real ones, rather than requiring clean geographic separation. They were designed for exactly this problem. They don’t eliminate uncertainty, but they handle small markets far better than a standard template applied without thought.

Methodology choice matters as much as budget. The template will give you results. It won’t tell you when those results are wrong.

---

Once teams accept that measurement is a systems problem, infrastructure questions follow.

Clean rooms: right tool, wrong expectations.

Data clean rooms let two parties analyse shared data without either side seeing the raw records. For retail media specifically, matching your sales data with a retailer’s ad exposure data, a clean room is often the only privacy-compliant way to close the measurement loop. If retail media is a meaningful share of your spend, you need a clean room strategy.

If it isn’t, you probably don’t.

The fragmentation problem is real: a Google clean room doesn’t interoperate with a Meta clean room, and the skills required to run meaningful analyses are significant. Clean rooms are useful infrastructure. They are not, on their own, a measurement strategy.

---

Where to start.

If you’re at Level 1 or 2 on the maturity ladder, the first move isn’t to build an MMM.

It’s to run one incrementality test on your biggest platform spend.

Pick your largest channel. Run a geo holdout or conversion lift study. Compare the platform’s claimed ROAS to the incremental ROAS the test reveals. You will almost certainly find a gap. That difference between what the platform says it delivered and what it actually caused becomes your internal case for better measurement. It’s also the number that changes the conversation with your CFO.

From there: build the data foundation. Clean, consistent, at least two years of weekly spend and outcome data. Then commission or build an MMM. Then run geo experiments to calibrate it. Then treat the maturity ladder as a governance document, not a research exercise.

The whole stack doesn’t have to arrive at once. But it has to start somewhere.

Start with the test.

---

Measurement is the operating system.

Measurement is not a reporting function. It is the system by which marketing capital is allocated.

If the system produces numbers that are wrong, because of platform conflict of interest, signal collapse, a model built on wrong assumptions, or organisational inertia, the allocation decisions that follow are wrong. Systematically. At scale.

The CFO conversation is changing. It used to be enough to point at the dashboard. It isn’t anymore. The question is shifting from “what did the platform report?” to “what did we actually cause?” That question has no good answer if your measurement is borrowed from the platform running the media.

The teams that get ahead of that question, with methodology the CFO can scrutinise and grounded in real experiments, will find the budget conversation gets easier. The ones still working toward it will find it gets harder every cycle until they do.

The window is now. Not when the CFO asks. Before.

That’s not a bad thing. The old measurement was convenient. The new measurement might actually be true.

Next edition: retail media. The same measurement problem, compressed into a walled garden where the retailer controls the shelf, the audience data, and the attribution. The conflict of interest is structural. The opportunity is real. More soon.

---

Systems That Decide publishes on the infrastructure of decision-making in modern business. If this edition was useful, share it with someone who should be asking harder questions about their attribution stack.

In a downturn, that question doesn’t just shape the budget conversation. It determines who wins it.

Late last year something quiet happened across a number of Australian media budgets. Experimentation dried up. Not dramatically. No announcements, no strategy pivots. Just a gradual tightening of the discretionary end of the plan. Test budgets got folded back into proven channels. New platform trials got deferred. Anything that couldn’t show a clear payback story inside a quarter got harder to defend internally.

It wasn’t panic. It was the quiet application of that CFO question. And how a business answers it under pressure reveals whether its marketing system is actually built for decisioning, or just optimised for execution when conditions are good.

The second-order story is the interesting one

When geopolitical tension rises and consumer confidence falls, the first-order story is obvious. Budgets tighten. Discretionary categories get hit. Brands pull back. That story will get written extensively and it’s largely correct.

But the second-order story is the one that actually determines where money moves and who benefits. When macro pressure rises, most businesses start by asking five questions:

Where can we slow spend without immediately losing volume?

What channels have provable payback we can defend to the CFO?

What is protecting our core customer base right now?

What can wait until conditions improve?

And what are we spending on because it’s always been on the plan, rather than because it’s earning its place?

That last question is the one that changes things. In easy markets, budget allocation persists through inertia rather than evidence. When the pressure comes on, inertia stops being a good enough reason.

What actually happens to budgets in a downturn

The pattern I’ve seen play out in ANZ is reasonably consistent regardless of the specific cause of the pressure.

Experimentation falls first. Test budgets, new platform trials, incremental channel expansion. Anything justified on a learning agenda rather than a performance agenda becomes hard to defend when the CFO is asking questions. This is rational in the short term and often damaging in the medium term, because the brands that stop learning during downturns tend to be behind when growth returns.

Broad brand investment gets scrutinised harder. Not necessarily cut, but questioned more carefully. Most brand investment is justified through reach, frequency and awareness metrics that don’t translate easily into the language finance departments speak. When budgets are comfortable, that gap is tolerated. When they’re not, it creates real internal tension.

CFO-friendly channels gain favour. Channels with closed-loop proof, clear attribution and short payback windows become easier to defend internally even when their actual incrementality is questionable. This is the measurement-as-control argument from Edition 3 playing out in real time. The channel that controls the proof controls the budget conversation, and that matters most when budgets are under pressure.

Retail media often benefits. Partly because of the closed-loop proof story. Partly because in a market where consumer confidence is falling, being present at the point of purchase feels more defensible than being present earlier in the funnel. And partly because the major retailers in ANZ have done a very effective job of making their media networks feel like a cost of doing business for FMCG brands rather than a discretionary investment.

The counterintuitive case for brand in a downturn

The assumption most marketers fall back on when pressure comes is that downturns mean performance. Cut the brand, protect the conversion. Hold your existing customers, stop chasing new ones. Optimise the bottom of the funnel and wait for conditions to improve.

That logic is understandable. It is also, in certain situations, exactly wrong.

Earlier in my career I watched this play out at Vodafone. The business was facing a structural consideration problem. Telstra and Optus customers weren’t switching not because they’d evaluated Vodafone and rejected it, but because they weren’t even thinking about it. The brand wasn’t in the consideration set at all.

Here’s the part that makes this more than a brand vs performance story. The dashboard wasn’t showing a consideration problem. Conversion rates for the people already considering Vodafone looked reasonable. The measurement environment was reporting the business as healthier than it actually was, because it could only see the customers already in the funnel. The ones who’d mentally written the brand off were invisible to it.

The business was slowly losing the market and the metrics were hiding it. By the time the consideration gap was large enough to show up in market share data, it was expensive to fix.

The strategy that worked involved product innovation to give people a reason to reconsider, pricing that removed the rational barrier, and brand messaging that created enough presence and distinctiveness that Vodafone became a genuine option in the mind of someone who’d written it off.

But the reason performance media had dominated the budget conversation wasn’t because it was solving the right problem. It was because it came with proof. The CFO could see conversion data. Nobody could measure consideration loss until it was already too late.

That’s the same structural problem this series has been arguing about from Edition 1. Whoever controls the proof controls the budget conversation. And when the proof only reflects part of the picture, the budget conversation leads you somewhere you didn’t intend to go.

Why average performance misleads you most when you need it least

Warren Buffett said: “Only when the tide goes out do you discover who’s been swimming naked.” In media, the tide is the dashboard. And right now, it’s going out.

In a normal market, average ROAS guides allocation reasonably well. But average performance and marginal performance are different things. Average tells you how all your spend performed. Marginal tells you what the next dollar will actually return. In a growing market those two numbers are close enough that the distinction doesn’t matter much. In a stressed market they can be very far apart.

When consumer confidence falls, demand doesn’t fall evenly. It falls first at the margin. The consumers who were marginally interested stop converting. The transactions that were happening because conditions were good stop happening. The average result in your dashboard still looks relatively healthy because it’s being held up by your core customers doing what they were always going to do.

Macro shocks don’t just reduce budgets. They expose where those budgets were riding the wrong part of the curve.

A channel that looked efficient because it was capturing easy demand in a growing market will look much less efficient when that demand softens. But the dashboard won’t tell you that immediately. The average is still being calculated across a mix of resilient and deteriorating demand, and it’s flattering you.

This is exactly why incrementality measurement matters more, not less, in uncertain markets. Incremental performance tells you where genuine demand still exists and where you’re just harvesting the residual of a better market. The brands that navigate downturns well are the ones that could read that distinction and reallocate toward genuine demand quickly.

That requires decisioning capability, not just execution efficiency.

The ANZ-specific pressure points

Australia has some specific characteristics that shape how macro pressure flows through to media budgets.

Consumer sentiment moves quickly here. The relationship between headline economic news and actual spending behaviour is tighter in Australia than in larger markets, partly because of mortgage exposure concentration and partly because the retail sector is concentrated enough that sentiment shifts show up in sales data fast.

Imported inflation flows through the cost base quickly for categories exposed to freight, fuel and overseas supply chains. That creates margin pressure that directly reduces marketing budgets before consumer demand has even softened.

The concentration of retail creates a specific dynamic worth understanding clearly. Woolworths and Coles become more powerful in a downturn, not less. When discretionary retail is under pressure and FMCG brands need to protect volume, the argument for retail media investment becomes harder to resist internally. The proof is easy to present, the supply chain relationship makes it feel essential, and the CFO conversation becomes straightforward.

But this is exactly the dynamic Edition 3 was describing. Channels accumulate budget not always because their incrementality is best in class, but because their proof is easiest to defend. In a downturn, that dynamic accelerates. FMCG brands can find themselves paying an increasing share of a shrinking marketing budget to access their own customers through a retailer’s media environment, at the worst possible moment commercially.

That isn’t an argument against retail media. It’s an argument for knowing whether it’s actually working incrementally before the market tightens and you no longer have the budget to find out.

What the stressed market reveals

If you’re a CMO heading into a budget conversation right now, here’s the question worth asking before you decide what to cut: do you know which parts of your current spend are driving genuine incremental growth, and which parts are simply claiming credit for demand that would have existed anyway?

Because that distinction is about to become expensive.

In easy markets, budget allocation can look smarter than it really is. Growth covers a multitude of poor decisions. Channels claiming too much credit still look fine when the underlying demand is rising. Measurement gaps that would be exposed in a flat or declining market stay hidden when everything is growing.

When conditions tighten, the system gets tested. Inertia stops being a strategy. The CFO questions that were politely deferred get asked with more urgency.

The brands and agencies that come through downturns in the best position aren’t the ones that optimised the old plan harder. They’re the ones that could ask, quickly and credibly: where does the next dollar actually deserve to go given what we know right now?

Not where did it go last quarter. Not where does the plan say it should go. Where does it actually deserve to go.

Growth is a generous cover story. It hides weak decisions, flawed measurement and misallocated budgets for years at a time. A stressed market takes that cover away. What’s left is the system underneath. Build one worth exposing.

The measurement gap doesn’t close itself. Edition 5 is about the tools the market is actually reaching for: MMM, incrementality testing, clean rooms. And the question worth asking clearly: do they actually work, or do they just sound like they do?

Never a decisioning partner.

That distinction sounds simple. The commercial implications are not.

Two months ago, I was part of a team that learned exactly what it costs to understand that too late.

The argument behind that story is worth understanding first. Because the Afterpay Ads experience is not just a personal anecdote. It is a specific expression of a structural problem playing out across the market right now.

The problem is the difference between attribution and incrementality. And much of the industry still behaves as if the two are interchangeable.

Why incrementality breaks the old model

For most of the programmatic era, attribution was the language of performance.

It was also, in retrospect, a very convenient language for the ecosystem.

Attribution answers the question: who gets credit? It works by distributing credit for a sale or conversion across the channels a consumer encountered. Last click, first click, linear, time decay, data-driven. The specific model varies. The underlying limitation does not.

Attribution tells you how to divide up the credit for something that already happened. It does not tell you whether the media caused it.

Incrementality asks the harder question: what actually caused this?

That sounds like a subtle distinction. It isn’t. It is the difference between a system designed to allocate credit within an existing budget and a system designed to determine whether that budget is worth spending at all.

Attribution is an execution tool. It tells you who performed within the campaign you already ran.

Incrementality is a decisioning tool. It tells you whether the campaign should exist in the form it does, and where the next dollar should go instead.

That shift, from attribution to incrementality as the dominant measurement standard, is the single most important change happening in media measurement right now. And it has profound implications for who holds power in the market.

The inconvenient truth about incrementality

Genuine incrementality measurement is hard. It requires controlled experiments, holdout groups where part of the target audience is deliberately not shown the ad, clean test-and-control design, and enough volume to produce meaningful results. It takes time, rigour and a willingness to accept that some campaigns will fail the test.

Most businesses in media find this difficult. Not because they are incompetent. Because the incentive structure makes it complicated.

Running a rigorous incrementality test that might show a channel is not driving incremental growth creates real commercial tension for the platform running the media, the agency recommending it, and the measurement partner being paid to evaluate it. The path of least resistance is measurement that looks credible enough to reassure the client without asking questions that might reduce the budget.

The result is a measurement environment built for a different era. Attribution made sense when it was the best available tool, when data was fragmented, when independent verification was expensive and slow. The industry didn’t build it badly. It built it for the conditions that existed at the time.

The conditions have changed. And the gap between where measurement is and where the market is heading is where the next round of competitive advantage gets built or lost.

Real incrementality requires independence. The framework has to sit outside the platform running the media, designed by someone whose commercial interest is aligned with the truth of the outcome rather than the continuation of spend.

That is harder to build. It is also increasingly the only version sophisticated marketers are willing to accept.

We learned this the hard way

Two months ago I was part of a team building exactly that kind of product at Afterpay Ads.

We had incrementality measurement. We could show advertisers whether their campaigns were driving genuine lift, new sales that would not have happened without the media, or simply taking credit for purchases that would have happened anyway. For a retail media business built on the promise of closed-loop proof, that mattered enormously.

The problem was that the methodology was not ours. It belonged to our DSP partner, the platform we used to buy and deliver media programmatically. We were using someone else’s framework to answer the most important question our advertisers were asking.

Everyone on the team knew this was a structural problem. Because if the methodology lives inside your DSP partner’s platform, you do not control what gets measured, how lift gets defined, or what counts as incremental. You are renting the proof. And a business that rents its proof is not a decisioning partner. It is a media vendor with a good story.

So we started building our own.

What we were actually trying to build

The framework we were developing was not built on ad-serving signals or attributed clicks. It was built on payments data. And what made it different was not just that it could show campaign performance. It could connect media exposure to business movement.

A brand running on Afterpay Ads would have been able to understand incremental lift across gross payment volume, total sales, average order value, share of wallet within category, lifetime value of customers acquired through the campaign, the split between new and returning buyers, where incremental demand was actually being created geographically, and how customer behaviour changed over time, from first-time buyers becoming habituated customers through to super users and churn.

That is not a media metric stack. It is a business performance stack. The difference matters.

Most measurement environments in market are still telling brands how their campaigns performed. We were trying to build something that told them how their business moved.

No DSP, no walled garden, and no attribution vendor was going to tell a brand that its media investment was converting first-time buyers into habituated customers at a measurably higher rate than a true control group. That insight only exists if you sit on the transaction data. And even then, it only has commercial credibility if the methodology is independent of the platform running the media.

Payments data creates a structural advantage, but it does not remove the need for rigorous design. Transaction visibility is powerful, but it still needs independent methodology to separate causation from correlation.

We were close. And then the project ended before it launched. The irony of the timing was not lost on me.

Because the force that brought the project to a close is the same force accelerating the shift from attribution to incrementality across the industry. When machines are doing the optimisation, the only meaningful human question left is whether the outcome is real.

The window to own your measurement layer is real. And it is shorter than most people in this industry seem to think.

The payments data opportunity the market still has not fully captured

What we were trying to build at Afterpay Ads represents an underdeveloped opportunity that still exists in this market.

Think about what payments data actually gives you. Real transaction outcomes. Not modelled, not inferred from a pixel firing on a confirmation page, not reconstructed from platform logs. Actual purchases, with actual values, connected to identifiable consumer behaviour over time.

That is a fundamentally more honest foundation for answering the question: did this campaign actually work? And more importantly: did it drive revenue that would not have happened otherwise, with customers who are now more valuable to the business than they were before?

This matters beyond one business for a simple reason. As payments companies, retail networks and commerce platforms move deeper into media, the businesses sitting on real transaction data have a structural measurement advantage over those that do not.

That advantage has not yet been fully exploited in ANZ. But the direction of travel is clear. The companies moving closest to the transaction are also moving closest to the most defensible incrementality proof. And defensible proof, as I argued in Edition 2, is the foundation of decisioning power.

In a stable market that matters. In a volatile one it becomes a real competitive advantage: the difference between reallocating budget before the dashboard catches up and reallocating it after.

Who benefits from the gap

Right now, in many Australian businesses, there is still a large gap between the incrementality measurement marketers say they want and the measurement they actually use to make decisions.

That gap is not neutral. Someone benefits from it.

In the absence of rigorous incrementality proof, budget allocation defaults to attributed performance. That tends to favour whoever is closest to the conversion, strongest at telling the performance story, and easiest to defend in a CFO conversation. Channels accumulate budget not always because their incrementality is best in class, but because their evidence is easiest to present as credible.

What retail media built was genuinely innovative. Closed-loop proof was a meaningful step forward from the open web measurement environment that preceded it. Cartology and Coles 360 didn’t just build media businesses. They built proof engines, systems that connect media exposure to a verified purchase transaction inside their own retail environments. That proof became a decisioning input. Brands bought more retail media not just because it performed, but because it could demonstrate performance in a way that strengthened the budget-allocation conversation, on the retailer’s terms.

That is the measurement-as-control-layer argument made visible. The retailer that controls the proof has disproportionate influence over the budget conversation.

The question the market is now asking is what comes next. As incrementality becomes the standard, every channel will need to demonstrate not just that a purchase happened, but that the media caused it to happen when it otherwise wouldn’t have. The brands and channels that get ahead of that question will be in a structurally stronger position than the ones waiting for it to be asked of them.

In a growing market, the measurement gap is expensive but survivable. In a flat or declining market, it becomes dangerous. The brands that close it before conditions tighten are the ones with genuine options when the CFO starts asking harder questions.

What this means for the market

The shift from attribution to incrementality does not just change how campaigns get evaluated. It changes who gets to be in the room when budgets are set.

A channel that can show genuine incremental outcomes, independently verified and connected to real commercial results, earns a different kind of budget conversation than one that can only show attributed performance inside its own reporting environment.

And this is not just a retail media phenomenon. You can already see the same logic across ANZ. MOVE 2.0 bringing mobility data into OOH planning. VOZ Streaming embedding incremental reach proof into BVOD trading. CRA Audio ID creating unified frequency measurement across competing radio networks.

Each is an attempt to answer the same question: what does this channel actually add? Not in a way that justifies spend after the fact, but in a way that is credible enough to influence allocation before the next budget is set.

The channels that answer that question most convincingly will attract the most budget. Not because of reach, targeting or format alone. Because they have made the incremental case more credibly than the alternatives.

The window is shorter than you think

AI-driven media buying is abstracting execution so thoroughly that the only meaningful question left is whether the outcomes are real.

When a machine is doing the optimisation, the human question becomes: optimising toward what? And is that outcome actually incremental, or is it simply the most efficient way to claim credit for demand that already existed?

The businesses that can answer that question, with proprietary data, independent methodology and genuine commercial proof, will hold the decisioning power in an AI-driven media market.

The ones still renting their proof will find themselves in a familiar position. Owning the asset. Not controlling the system that measures it.

Most of the industry is still asking who gets credit. The market is about to start asking something harder.

In easy markets, weak measurement gets hidden by growth. When conditions tighten, the gap between what brands think is working and what is actually working becomes expensive. That’s Edition 4.