The FTC alleges that Amazon did not run a straightforward Generalized Second Price (GSP) auction, where the winner pays the amount offered by the second-highest bidder (+ a nominal 1 cent). As y’all recall, such SBA are supposed to reveal your true value for showing the ad. There is no reason to bid less, because it would just reduce your odds of winning the auction without changing the price you’d pay. In contrast, the FTC got from internal documents, the price paid by advertisers “isn’t set by an actual bidder,” but is instead a “proxy 2nd price that we calculate’. The consequence?

Amazon’s surcharges have resulted in advertisers paying their own bid amount an increasingly high amount of the time. For Sponsored Products ads, the complaint alleges that the percentage of the time advertisers paid the amount of their bid increased from between 30% and 40% in 2021, to 70% in 2022, and to approximately 80% in 2024 as a result of the surcharges.

Amazon Ads called the lawsuit misguided:

The FTC’s claim fundamentally misunderstands how advertisers operate. Advertisers adjust bids based on real-world performance, not descriptions of auction mechanics. Even accepting the FTC’s flawed premise that advertisers do not adjust bids, we estimate they saved over $8 billion from 2021 to 2025 as a result of Amazon prioritizing ad relevancy over selecting ads on bid price alone.

From 2019 through 2024, the average cost-per-click for Amazon’s Sponsored Products search ads remained flat adjusted for inflation, while conversion rates grew 24% from 2021 to 2025. Advertisers paid the same and got more as we meaningfully improved ad relevancy and therefore performance.

Note that Amazon Ads uses a word absent from the FTC lawsuit: ad relevancy. The winning ad is not the highest bid, it’s the best combination of the advertiser’s bid and the estimated relevance of the ad to the consumer. The Amazon Ads response details how it deliberately made that shift over the past decade, arguing this is exactly what a responsible marketplace should do. Understanding the relevance argument is key for the court to decide on this case.

Relevance helps consumers, advertisers & Amazon

Back in the day, Amazon’s advertising ran on simple rules-based systems that leaned heavily on bid amount, the closest thing to a textbook second price auction. As the catalog grew, Amazon says that approach broke down: purely bid-driven ranking surfaced ads shoppers didn’t want, cluttering search results with irrelevant listings from whoever was willing to pay the most. Starting in 2014, Amazon built machine learning models to predict how likely a shopper was to actually want a given ad, weighing relevance alongside bid, and by 2019, the year I was hired, those models ran across all of Amazon’s Store advertising.

Now, I believe Amazon Ads’ story because I lived it for years. Our training focused on increasing the odds that ads helped instead of harmed consumer experience, the alpha and omega of Amazon’s flywheel. As I detailed in past Substacks, this focus took many forms, from the tactical (don’t add exclamation points to any ad !!! ) to the strategic (instead of targeting segments, talk about reaching relevant audiences).

Amazon’ reasoning: an irrelevant ad is worthless or worse to the shopper and Amazon’s search experience, and therefore worthless to the advertiser paying for it. Rewarding relevance instead of raw bid size is supposed to route ad dollars toward sellers whose products people actually want, which is also the group Amazon most wants to keep on the platform. Amazon cites a sharp shift in outcomes as evidence: the model does what it claims: average winning bids for Sponsored Products fell 50 percent from 2019 to 2025, and about 92 percent of placed ads did not go to the highest bidder. On Amazon’s account, the auction moved away from pure willingness-to-pay because the company decided that was the wrong thing to optimize. Routing ads to the shoppers most likely to respond helps consumers, advertisers and Amazon itself: because the company only gets paid for Sponsored Products and Sponsored Brands when consumers click on these ads, relevance improves Amazon’s revenue both in the short term and in the long term.

But how does Amazon estimate and share relevance?

But the devil is in the details. Amazon’s relevance score is not a single visible number. It’s a model that weighs, among other things, keyword and category match between the ad and the shopper’s search, and, critically, the advertiser’s own track record: historical click-through rate, and how often people who clicked went on to add the product to cart or buy it. Amazon does not publish the weighting, doesn’t let advertisers see a relevance score the way they see a bid or a budget. Some third-party sellers have described the entire mechanism as a black box they can influence only indirectly, through keyword targeting, listing quality, and reviews, without ever seeing how those levers actually move the number.

But of course, there is a reason why Amazon doesn’t open the black box, just as Google, OpenAI, and Anthropic won’t tell you exactly how their services reveal your offerings before your competitors’. Doing so would allow advertisers to game the system.

And even if you want full transparency, it’s a stretch to see fraud in this deliberate vagueness. A performance-history-weighted relevance score is not a hidden fee as the FTC implies, it is the mechanism Amazon told advertisers, in its Ad Console since 2018, that a bid represents a maximum you could be charged, not a promise you’ll always win on price alone. Every major digital ad platform, including Google’s and Meta’s auction systems, incorporates a version of predicted engagement into ad ranking for the same reason: an auction that ignores whether anyone will actually respond to the ad may be a ‘pure’ GSP but has worse outcomes. Track-record weighting also has a defensible upside the FTC’s complaint doesn’t engage with: research on cold-start problems in e-commerce search has found that using ad engagement as a quality signal can help platforms identify good new products faster.

How transparent should relevance be?

The complaint that the relevance formula is a black box is not going away just because it doesn’t prove deception. Advertisers reasonably want to know how much weight their historical CTR and conversion rate carry, because it changes strategy: if track record compounds, an established seller’s ads get cheaper and more prominent over time relative to an equally good product from a newcomer. Now THAT’s a competitive dynamic the FTC may want to look into for transparency and market-structure. It’s a different claim from the one the FTC actually filed, which is that Amazon lied about running a second price auction while secretly overriding it. Amazon’s defense is that it never hid the shift toward relevance, it built and marketed an entire ad product around it, and the industry-wide adoption of relevance-weighted ranking is the strongest evidence that this is a design choice, not a concealment scheme.

The marketing lesson, regardless of court outcome

For anyone teaching or building auction-based marketplaces, this case is a useful reminder that “second price auction” stopped being a complete description of any major ad platform’s ranking logic years ago. The real governance question is whether relevance models that lean on an advertiser’s own performance history need a level of disclosure and auditability that none of these platforms currently provide.

The first is Juozas Kaziukėnas’s tracker of six months of ChatGPT ads. The second is Katie Deighton’s Wall Street Journal report on Reddit’s moderators turning on the brands suddenly desperate to be liked there. In other words, should you buy the front door into AI (ads) or the side door?

Door one: pay OpenAI

OpenAI’s ad business has moved fast. February brought a first live test to US users. Resistance was not noticeable despite Anthropic’s SuperBowl ad mocking OpenAI’s ads. In April, I started with it giving my university-wide Klein lecture. By May there was a self-serve Ads Manager. Pricing dropped from a $60 CPM with $200,000 minimum commitments to open, CPC-based bidding with a conversion pixel. Targeting works on conversation context rather than keywords or audience segments, which is a different premise than fifteen years of Google and Meta trained us to expect. Ad load is climbing as more advertisers sign on, and the format has already stretched from a simple banner to Wayfair-style product carousels where OpenAI, not the advertiser, decides which SKUs to surface. In other words, OpenAI is adding all the functionalities I come to expect from an Amazon ad.

OpenAI is projecting $102 billion in ad revenue by 2030. Sam Altman has called ads “a last resort,” which is the kind of line that ages the way most last resorts do once the revenue math kicks in. Obviously, ads are a good deal for OpenAI

But are they for advertisers? Nobody is reporting good returns yet. Early advertiser chatter cited in the OpenAI tracker describes solid impressions and click-through rates paired with conversions that are, in one operator’s phrase, in the gutter. Twenty percent of ChatGPT conversations apparently carry shopping intent, which sounds like an enormous opportunity until you remember that intent sitting inside a conversational answer is not the same as intent sitting inside a search results page. The behavioral context is different, and the ad product hasn’t caught up. Remember Alexa’s problems demonstrating ad effectiveness despite heavy adoption?

This is a normal new-channel story. CPMs launch high to ration scarce inventory among the desperate and curious, infrastructure matures, self-serve arrives, and then everyone spends a year figuring out whether the channel actually converts or just looks good in vanity metrics. The only really new variable is that the ad sits inside a system whose core value proposition is that it gives you an honest, unbiased answer. That tension between “trustworthy assistant” and “media property with a revenue target” deserves more discussion than the CPM.

Door two: get into the main AI source: Reddit

If ChatGPT ads are the front door, Reddit is the side door for many brands. Reddit is now the most-cited site in AI answers according to visibility firm Parse. Job postings for “Reddit strategy manager” roles are appearing at brands that had never posted there. A cottage industry of Reddit-specific agencies has sprung up in the last year, some promising dozens of monthly brand mentions through accounts built to look like ordinary users, others charging thousands a month to suppress negative posts.

Of course, depending only on Reddit is not a good idea, as OpenAI can easily change sources. The recent past is a good illustration:

As Filipe Lins Duarte noted when posting this image yesterday:

Reddit has not become irrelevant because of this. Far from it. The platform has 500M+ Weekly Active Users. 4% of citations from Google AIO are still coming from Reddit (we track dozens of thousands of prompts daily and see this in our own data set).

Comeback. We've seen a huge drop in Reddit's share of ChatGPT citations in mid-June. In about 1 week, it dropped from 7% to <1%. In 2 months, it went back up to 5%. I believe something similar will happen here.

LLMs data sets. AI Engines use their own internal knowledge and datasets to help decide which brands get mentioned in AI Answers. Reddit has contributed extensively to that dataset already, and it still will. So, AI recommendations are still shaped (to some extent) by Reddit.

More importantly, Reddit’s moderators, mostly unpaid volunteers who run the platform’s actual social contract, are not having it. The Wall Street Journal’s reporting is full of subreddits banning promotional posts outright, moderators warning communities to watch for coordinated brand content, and users reacting to marketing intrusion with the specific outrage of people who feel their community space is being colonized. One line captures the mood: users don’t want to hear from the brand while they’re gossiping about the brand.

The brands that have actually built goodwill on Reddit did it by not marketing. Sonos sent a real employee, KeithFromSonos, in to solve support tickets and translate corporate announcements into plain language. Fidelity built subreddits around genuinely useful finance discussion. HelloFresh tried something similar with an employee persona named Katie, who spent months absorbing complaints and answering questions before quietly going dark this spring, an outcome that reads like a company deciding the ROI math on sustained authentic participation didn’t pan out either.

The structure connecting the doors

Line these doors up, and a familiar structure appears, just with new machinery. One channel is paid media with an auction, a pixel, and a still-unproven conversion story. The other is earned reputation, except the earning now happens inside communities that have gotten sharply better at detecting unwanted branding and punishing it in real time, and the stakes are higher because that reputation now gets scraped straight into what a chatbot tells the next million people who ask about your category.

Signaling theory explains why the Reddit story is the harder one to game. When a brand shows up with obvious promotional intent, that itself is the signal, and it’s a bad one: it tells the community the brand values reach over relationship. What Sonos and Fidelity signaled instead, through sustained low-cost, low-drama helpfulness, is exactly the costly signal that’s hard to fake and therefore worth believing. Reddit’s moderators aren’t reading brand strategy decks. They’re doing what humans have always done with signals of intent, they’re pattern-matching on behavior over time.

The CPM auction has no equivalent immune system yet. It will develop one, probably in the form of ad fatigue, blindness, or outright platform backlash if OpenAI pushes load too far, but for now it’s a market that rewards showing up with a budget rather than showing up with credibility.

What this means for where you spend

If you’re deciding where to put incremental effort in the next two quarters, my advice:

(1) Treat ChatGPT ads as an experimental line item with a real budget cap and a test-and-learn structure, the same discipline you’d apply to any first-year platform, and don’t expect it to carry conversion targets yet. In financial terms, it’s a ‘real option’: you want to experience how it develops.

(2) Treat Reddit, and AI-cited communities generally, as a long-horizon reputation investment that has to be staffed by people who can tolerate being ignored or mocked for months before anything resembling trust shows, and that cannot be delegated to an agency promising volume.

Both channels ultimately feed the same downstream outcome, which is what a chatbot says about your brand when a customer who has never heard of you asks for a recommendation. One door gets you there faster and cheaper today. The other gets you there in a way that’s much harder for a competitor to copy, and much harder for OpenAI, Google, or anyone else to sell you access to.

As my previous newsletters and Amazon Days substacks talked a lot about the cost cutting and culture, let’s focus this one on the top line:

$200.6 billion in quarterly revenue, up 20% year-over-year, the first time the company has crossed $200B in a single quarter.

Where does that revenue growth come from? AWS grew 37%, its fastest pace in nearly five years, and Amazon’s AI and chips businesses are each now running above $25 billion a year, with major AI labs among the customers.

And of course advertising, growing 26% to nearly $20 billion. On the Q2 call, CEO Andy Jassy tied that growth directly to agentic and conversational experiences: shoppers who click a Sponsored Prompt convert to a sale 48% more often, and spend 21% more on average, than those who don’t (Amazon Ads Q2 2026 recap).

As to AI shopping, Jassy said more than 350 million customers have used Alexa for Shopping in the last 12 months, with active users nearly doubling and interactions up over 5x year-over-year in Q2 alone (Amazon Stores Q2 2026 recap).

On the surface, this looks like a straightforward growth story. Underneath, it means something bigger on the future of agentic commerce.

Who’s browsing: human or agent?

ChatGPT Shopping is now live for every US user, recommending products and routing shoppers to merchant sites, with Etsy and over a million Shopify merchants already plugged in. Google’s AI Mode has launched agentic checkout, with Wayfair, Chewy, and Etsy among the first retailers on board. I am participating myself in a ‘ManyPrompts’ project developing and testing agents to buy…socks

This is exactly the “funnel collapse” theme we kept coming back to at the IJRM preconference: when a customer tells an AI “find me a blue top for a bridal shower, formal, under $80” and the agent does the searching, comparing, and often the buying, the ‘fat middle’ online consumer journey may no longer be there to slide into the DMs. As you recall, the low search cost on the internet changed the traditional funnel

into this diamond, allowing smaller brands (with great reviews and/or online presence) to expand the customer’s consideration set.

But now this window may be closing.

And Amazon itself is the clearest illustration of just how contested that window is. Even as Jassy touts agentic engagement on the earnings call, Amazon has spent the past year building a wall against everyone else’s agents. It has blocked 47 external AI bots from crawling its site and sued Perplexity over its Comet browser’s shopping agent, winning a temporary court order in March (CNBC). Tellingly, Amazon’s complaint argued that AI-generated traffic complicates its own ad business, since bot impressions have to be detected and filtered out before an advertiser can be billed; the exact governance and measurement problem I flag below.

So the funnel collapse isn’t a surrender to agents. Instead, it’s contested territory, and Amazon’s $19.8B ad business is a big part of why it’s fighting to keep the browsing on its own turf, on its own agent, rather than someone else’s.

Execution is catching up fast, too. Reports this month describe the first in-flight agentic optimization of a live programmatic campaign, with an AI agent adjusting bids and targeting between reporting cycles a human used to manage. That raises a governance question every CMO approving this kind of system needs to answer: what can the agent actually touch, and what’s the audit trail when it acts on your media budget without you in the loop.

Where this leaves fairness and trust

The other thread from the preconference was dynamic pricing and fairness. It is getting more, not less, urgent as agents enter the picture. Research shows that consumers don’t mind dynamic pricing that smooths over demand based on factors they can control: location (a Coke costs more in a fancy beach hotel bar than in the supermarket), timing (buying your plane tickets earlier). They do feel discriminated if they can’t control the factors, such as temperature (remember the coke vending machine on the Brazilian beach?), and of course personal data.

Regulators are moving on exactly this line between market-based and person-based pricing. State lawmakers introduced more than 40 bills across at least 24 states in 2026 to regulate personalized algorithmic pricing (Inside Privacy), and Maryland went furthest: Governor Wes Moore signed the Protection From Predatory Pricing Act in April, restricting AI-enabled pricing in food retail and delivery starting October 2026 (Consumer Finance Monitor). At the federal level, the FTC’s ongoing surveillance-pricing study and a House Oversight Committee investigation opened in March both signal the same scrutiny is coming nationally (Holland & Knight).

Some companies aren’t waiting to be told: Instacart halted its AI-powered item-price experiments after a Consumer Reports investigation into surveillance pricing (coverage here).

That distinction is about to get more complicated once an AI agent is the one doing the shopping. Whose data is the algorithm reading, and who’s accountable when a shopper never actually sees the price they would have gotten?

What I’m watching next

Amazon’s ad business will be one of the clearest windows into how this plays out, given how much of its growth is now AI-adjacent, and given Amazon’s own increasing use of agentic tools in its ad stack — even as it fights to keep third-party agents off its own site. If retail media dollars start following the agents rather than the humans, we’re going to need new metrics, because impressions and click-throughs assume a person is doing the looking. Guess this metric guy is not out of work....yet!

I spent a gorgeous morning yesterday at BU Questrom’s Platform Strategy conference, an annual gathering of economists and strategy researchers who study … platforms: how they compete, personalize, and increasingly, how they get disrupted or “sherlocked” by their own gatekeepers.

Four talks stood out for me, and here’s what I took away from each.

1. Does personalization actually help?

Our own (Northeastern U D’Amore-McKim marketing group) colleague Emil Palikot (with Susan Athey and coauthors) presented evidence from an ed-tech platform on the economics of algorithmic personalization. Results were nuanced:

• Gains are front-loaded. If you’re running an RCT (Randomized Controlled Trial) to measure the value of a recommendation algorithm, timing matters a lot. Steer clear of naturally low-engagement weeks (Diwali and Spring Break for articles in this college setting) because they can distort your read on the treatment effect.

• The benefit decays fast down the list. Comparing a better algorithm to a worse one, most of the advantage shows up at the very top of the feed. By the time you get to something like the 20th-ranked story, the gap between “good” and “bad” personalization has essentially disappeared.

• Who benefits most? In the Q&A, Hanna Halaburda asked whether students with less typical preferences might gain more from personalization. Emil’s answer: it’s less about how unusual your taste is and more about how much you seek out: users who actively look for more content show the largest measured benefits.

The answer to the title question: it depends. And it’s complicated: the value of personalization is concentrated at the top of the ranking and among your most engaged users. Next up: let’s analyze the embedded space for mismatch between high demand and lack of articles!

2. When does algorithm quality beat data headstart?

Sarit Markovich (with Hanna Halaburda and Yaron Yehezkel) tackled a classic platform-competition question: can a better algorithm allow the entrant overcome an incumbent’s data advantage? Claude over ChatGPT is the example du jour, as you know also from my Stacks.

The analytical (game theory) model contrasts the higher-quality algorithm of the entrant with the abundant data (thanks to more users) of the incumbent. Who wins the market? And how increases market share (has momentum) over time? It depends on how valuable said data is to the quality of the algorithm:

• Below the threshold, data isn’t that valuable, and the entrant can capture 50% of the market fast, and its share keeps growing over time.

• Above the threshold, the entrant gets essentially 0% market share at the start

So far, so ‘winner-takes-all’ as usual. But the authors add the evolution over time:

• Dynamically, even an entrant starting from zero share can eventually break in, because the incumbent’s existing data depreciates over time. That gives the entrant an option value: it can choose when to enter, biding the perfect time

• Near the threshold, the model produces three possible equilibria for the long term: entrant-take-all, incumbent-take-all, or a stable split market.

One counterintuitive result: more product differentiation is bad for the incumbent, because it makes it easier for the entrant to peel off users sitting at the far end of the preference spectrum (the “Hotelling line”). And if a platform tries to monetize its data directly, that acts like a hidden price increase on users, reducing their utility just as a price hike would. So now OpenAI should be reluctant to introduce ads, because Anthropic can use it against them. Sounds familiar?

The magnificent Marshall Van Alstyne asked a fun question: Grok’s market share has reportedly fallen from 17% to 8% … what would you tell them? Markovich’s answer: this is a quality problem, not a data problem. Indeed.

3. Why more competition doesn’t fix harmful content

Francesco Decarolis presented on platform competition and harmful content, with a mechanism that runs against conventional antitrust intuition: more competition doesn’t necessarily help, because platforms aren’t symmetric in who they attract.

The logic: large social media platforms disproportionately draw in “naïve” users: people who don’t fully weigh the costs of harmful content when choosing a platform. If a competition policy intervention pushes more sophisticated (”rational”) users toward a new entrant, the incumbent is left holding a user base that’s disproportionately naïve, so it has less incentive to clean up harmful content, not more. Meanwhile, the entrant’s rational users don’t necessarily benefit either, since their only alternative is an even worse incumbent.

Net effect: opening up competition can leave harmful content worse off on both platforms, just distributed differently.

4. Sherlocking: when the platform becomes your competitor

Benjamin Leyden kept all of us from lunch as we could smell the delicious flavors…but so made it worth our while.

“Sherlocking” is the strategy-lit term for when a platform owner enters a market that its own third-party developers created. Think Amazon Basics, or in this case Apple building its own version of a popular app category.

It’s a topic ripe for legislation (Leyden noted Elizabeth Warren has been a vocal critic) and he argued that public debate has run well ahead of the empirical literature on what happens as the net effect on demand, quality and prices.

The twist: developers often like such entry! When Apple enters a submarket, it validates that niche market, growing demand and investment.

The empirical findings, using a two-way fixed-effects difference-in-differences design on AppMonsta data:

• When Apple enters a submarket, pricing tends to bifurcate: apps shift from free to paid-download models, likely to differentiate from Apple’s free alternative

• Apple’s entry disproportionately pulls in the more price-sensitive users, pushing surviving third-party apps toward adding functionality and raising prices to compete on a different margin.

• The effect is strongest for developers who are (moderately) close to Apple’s offering but still meaningfully differentiated from it.

• Ratings actually rise for more differentiated apps post-entry, and fall for apps that look too similar to Apple’s own.

• Whether Apple’s app is “integrated” (the first thing you see) matters, as does the entering cohort.

The key question from the audience, including myself: what’s special about platforms? My own research ‘Who Benefits from Store Brand Entry’ showed consistent findings 2 decades ago:

1) The market bifurcates: high-end brands continue to innovate and raise prices, while low-end brands are too close to the store brand and suffer

2) The net effect include more consumer choice and differentiated assortment, but also lots of pressure on manufacturer margins

3) The physical display matters: is the store brand the first you see on shelf, and right next to a top brand or even imitating its packaging?

Takeaway across all four talks:

Platform competition rarely has a clean, one-directional answer. Personalization gains are concentrated, not uniform. Data advantages can be temporary. Competition policy can backfire on content quality. And even a platform “stealing” your product idea might expand your market rather than kill it. If there’s a common thread, it’s that the distribution of effects, across users, developers, and time, matters as much as the average effect itself.

Read more

Read more

If your numbers are good, you show…

Read more

Read more

Read more

Read more

Last year in this Substack, I asked whether retailers should buy advertising technology from Amazon. Today the story is bigger. Amazon is …

Read more

Read more

Read more

Benevolent Bandwidth

Read more

Read more

Read more

AI alignment is not possible if affected communities are brought in only after the model is built.

T…

Read more

You Are Already Managing Something

At Amazon, I started as an Individual Contributor (IC) but then moved into the management role, where my first task was to manage my first report through Performance Improvement (PIP)! Management skills can be taught, as decades of organizational theory and training have demonstrated. More recently, we witnessed the transition from tool-based AI (a model that does one specific thing when you tell it to) to agent-based AI, which takes a goal and figures out the steps. The difference is not technical. It is managerial.

A tool waits for instructions. An agent needs a manager.

Rumman Chowdhury, founder of Human Intelligence and one of the sharpest voices at the conference, put it plainly: if you do not give AI intention and motivation, it will sit idle.

Or worse, it will optimize for the wrong thing with great efficiency. The manager’s job includes setting direction, defining success, checking outputs, correcting drift. It does not disappear when you introduce an AI agent. It becomes the whole job.

Laura Burkhauser, CEO of Descript, described what she sees in her highest performers today. They hold lots of context, load it into an agent, and oversee what comes out. Instead of ‘just prompting’, that is delegation with accountability. Anyone who has managed a talented but junior employee (or PhD student) will recognize the skill set required: clarity on the goal, patience with the process, judgment about when to intervene.

The title on your business card does not matter. If you are working with AI agents, you are a manager. The question is whether you are a good one. The conference highlighted the 3 main ways you can be bad at AI management.

What Bad AI Management Looks Like

Read more

Read more

Read more