The grudge is about a specific kind of leader: the one who takes copy a writer spent two weeks on, kills it in final review, and replaces it with something worse because they like it better.

No data or concrete argument as to why, either; just personal preference and taste wearing an executive title.

What made me want to ask Kayte about it is that I now watch the same leaders paste whatever ChatGPT hands them straight onto the homepage. They prompted it, so it sounds like them, so it must be good.

It’s the same instinct with a different tool and leads to the same underperforming page on a KPI they don’t personally own.

I had hoped she would agree with me. She mostly didn’t, and her reasoning softened my stance.

Taste always wins, so stop bringing taste to the fight

Her answer was to zoom out.

“Think about a book you like. Think about a style of writing you like, a style of music you like. Taste is personal, taste is individual, taste is unique.”

Her read is that when a CEO says “I don’t like that,” they usually aren’t dismissing the thinking behind the work. They’re reporting a preference, and their preference wins because it’s their company (or at least they have more stake in the company’s success).

Then she added the part that’s been rattling around in my head for a few days.

“If you did the research, you know what’s right and what’s correct, you fight for it. You give them a reason... At the end of the day, your boss or your boss’s boss is gonna have the final say, and if their taste is what wins, their taste is what wins.”

She compared it to product defending a roadmap decision. You don’t get to say “trust me,” but you can always say “here’s why.”

On the recording I told her I thought that was too forgiving to leaders. I still think that.

But she’s right about the mechanism. Most writers who lose that argument lost it because they showed up with taste to a taste fight. Evidence is the only thing that reliably beats preference in a room where someone else signs the checks.

Red pen experience is why she can trust the agents

Kayte started as a freelancer and kept a 96% repeat business rate. Her explanation was almost boring: ship it done, and ship it early. Client wanted it Friday, they got it Thursday, as close to publishable as she could make it.

She also used to print her drafts, sit down with a red pen, and read them out loud.

“Would somebody actually say this? Like, does it sound good or does it make sense? Because those are two different things.”

That’s the fundamental underneath everything else she said.

What that skill buys her now

She’s running an n8n build with (her words) far too many agents stacked on top of each other, and she has Claude publishing to Webflow so she never copy-pastes into a CMS again.

“I get to be creative and strategic and make all these decisions, but I don’t have to worry about publishing on Webflow... Claude can do that. Like, are you kidding me?”

She can hand that off safely because she can still tell when a draft is bad. Take away the red pen years and the agents produce confident garbage she has no way to catch.

She said it plainly about her 13-year-old: you write the paper yourself, then you get the calculator, then GPT can help.

Worth knowing where she started, by the way. When her CMO first demoed ChatGPT to her team, she watched it write a story and shut the laptop.

“Robots are running the world. They’re gonna destroy my career. This is the worst thing I’ve ever seen.”

The confession I keep coming back to

The best moment of the episode was Kayte calling herself out.

“I built the engine, but what I need to do is, like, put the keys in and let it run, as opposed to putting the keys in every time myself.”

The scheduling, the final loop that would make it autonomous, she hasn’t gotten to. Not because she doesn’t trust the tools (she clearly does), but because time.

I think that describes most people who believe they’ve automated something. There’s a difference between a human step you chose to keep and a human step you never got around to removing. The first one is judgment whereas the second one is a to-do list item wearing a costume.

Audit your own stack for that this week. Every place you manually kick something off, ask which kind it is.

Two more things worth stealing

Ten minutes, three newsletters

She challenged her team to block ten minutes every few days to skim AI news. Not an hour because that just leads to rabbit holes. But ten quality minutes to skim is something anyone can do.

“If I go down a rabbit hole, I get so sucked in and it gets so far away from the one thing that I’m trying to learn.”

It’s the only sustainable version of keeping up I’ve heard anyone describe.

Her answer on what our kids should do

I didn’t expect this from an AI-forward content leader. She thinks the trades are where a lot of this rights itself. Welders, machinists, electricians, plumbers, all short on people while our generation went to college and got jobs in tech.

I’ve been telling my kids to practice washing robots (spoiler: hers is a better answer).

The full conversation is up now. If you take one thing from it, take the reason-not-preference framing into your next review cycle.

And go follow Kayte on LinkedIn, where she’s finally started making videos and is much better at them than she thinks.

Chat soon,

Nathan

But nobody from your company was invited.

There’s no calendar link, no Gong recording, and there’s no entry in the CRM. A buyer opened a chat window on their browser, described their problem in plain language, and asked what they should do about it. The answer they got back will decide whether you’re in the deal before you know the deal exists.

There’s only one part of your company that did make it into that room: the content the user’s model read.

That’s the argument I want to make today (with receipts) for why SEO/AEO should be the backbone of your entire go-to-market function. And if you run sales, CS, product, or the P&L, the middle section of this piece is written for you specifically.

The room you’re not in

Twice as many buyers named generative AI or conversational search as their most meaningful information source than named any other source. Ahead of vendor websites, product experts, and even sales reps.

Inside those AI tools, buyers are doing the work your funnel was built to capture: they compare vendors, research products, and build the internal business case, all before any vendor contact.

G2’s March 2026 survey of 1,076 B2B software buyers puts the shift in starker terms. 51% now start their research in an AI chatbot more often than in Google, up from 29% a year earlier. 69% chose a different vendor than they originally planned because of an AI chatbot’s guidance. A third bought from a vendor they had never heard of before the chatbot mentioned them.

That means nearly seven out of ten deals changed direction based on a conversation no vendor attended.

Stop calling it “content marketing”

When a buyer asks ChatGPT “which content platforms work for a two-person marketing team,” the answer is assembled from whatever machine-readable material exists about your category.

• Your comparison pages

• Your documentation

• Your pricing transparency

• Third-party reviews

The structured, crawlable, extractable answer layer that SEO/AEO work produces.

That means your SEO/AEO library has taken over jobs that used to belong to other departments:

It runs the first sales call

The buyer’s opening conversation about your category happens with a model, and your content is your only rep in the room. The material that business case gets built from is your answer layer or your competitor’s.

It builds the shortlist

Buyer research has shown for years that most of the journey happens before the first form fill, and that the day-one shortlist barely changes afterward. AI compressed this further.

The shortlist now forms in a chat window, from cited sources, and shortlists are shrinking (roughly 2.5 vendors on average, down from 3.2, per Apollo’s 2026 data). Fewer seats at the table, and the seats are assigned before you know the meeting exists.

It does the vetting

G2 found 85% of buyers think more highly of a vendor when an AI chatbot cites it in a recommendation. Citation has become a trust signal, the way a first-page Google ranking was in 2015, except the buyer never sees a results page with ten alternatives (they only see an answer).

If a single library of structured content is running your first call, building your shortlist, and carrying your credibility, calling it “content marketing” is an org-chart misnomer.

It’s actually become the substrate of the whole Go-to-Market motion.

What this means for each team

1. For sales

The buyer arriving on your calendar has already compared you, priced you, and drafted the internal case, with AI assistance. Gartner’s late-2025 survey of 645 B2B buyers found 69% of buyers use sales reps specifically to validate what the AI told them.

Your job on the first call has changed from “educate” to “confirm or correct the model’s summary of you.”

Two implications:

• First, you should know what the models say about your category (ask them, it takes ten minutes).

• Second, every objection-handling answer, competitive breakdown, and pricing explanation you wish buyers understood belongs in the public answer layer, because that’s where the pre-call briefing comes from.

The content team can only write it if you feed them the objections.

2. For outbound and SDRs

Cold outreach lands differently when the model has already introduced you. That 85% citation-trust number is your air cover. And the inverse is brutal: pitching a buyer whose AI-generated shortlist doesn’t include you means arguing with their research, not adding to it.

The highest-leverage thing an outbound team can do this year is route what prospects actually ask (the real questions from real replies) back into the answer layer.

3. For customer success and support

The same structured answers that earn citations are your knowledge base.

One canonical explanation of how your product handles X serves the prospect asking Perplexity, the customer asking your help widget, and the support agent (human or otherwise) resolving a ticket.

Fragment those and you get the enterprise classic: the website says one thing, the docs say another, and the model quotes whichever it found first. Coherence in the answer layer is a retention function rather than simply an acquisition one.

4. For product

The questions buyers ask answer engines are the cleanest market research you’ll ever get, because people interrogate a chatbot with an honesty and bluntness they never bring to a discovery call.

The query and citation data coming out of AEO work (what’s asked, what’s cited, where you’re absent) is a roadmap input. And there’s a distribution wrinkle worth knowing: Moz’s February 2026 analysis of nearly 40,000 queries found 88% of Google AI Mode citations come from pages outside the organic top ten, and an academic study published in December 2025 found 37% of AI-cited domains don’t appear in traditional search results at all.

The models read deeper than the SERP ever rewarded. Your documentation, changelogs, and technical pages are now discovery surfaces.

5. For the exec team and RevOps

Traditional attribution captures roughly 27% of the buyer journey; the rest happens in channels you can’t track. And AI search referrals, small as they look in the traffic report, convert absurdly well: Ahrefs measured it on their own site and found AI search visitors were 0.5% of traffic but drove 12.1% of signups, roughly a 23x conversion edge over classic organic.

(This is just one company’s data, so hold it loosely, but the direction matches every study since.)

So the surface that influences most of the journey is the one your dashboard sees least, and the trickle it does send you is the highest-intent traffic you have.

Meanwhile Forrester reports companies seeing 10 to 40% declines in website traffic as research migrates into answer engines. If you’re managing GTM by the metrics that made sense in 2022, you’re basically watching the scoreboard of a game the buyer stopped playing.

i.e. It’s time to adapt to the rules of a new game.

Why SEO/AEO specifically, and not just “do content”

SEO/AEO is the only content discipline that produces infrastructure instead of moments.

A LinkedIn post is a moment, a webinar is a moment, and a paid ad is a moment with a timer set. They’re meaningful moments, no doubt, and they matter (I’m literally building a media brand), but they decay in days and they aren’t queryable.

SEO/AEO work produces something structurally different: a permanent, machine-readable library of answers, organized around the questions your market actually asks, that compounds instead of decaying.

And here’s the part I care about most as a systems person: that library is the same asset your internal AI runs on. The structured answers that earn citations in ChatGPT are the structured answers your own workflows draw from when they draft a follow-up email, brief a rep before a call, or resolve a ticket.

Build the answer layer once and it serves two audiences:

• The models your buyers ask

• The systems your team runs

That’s why this belongs at the backbone. Every pipe in the GTM factory (outbound, inbound, ABM, events, case studies, support) draws from the same reservoir of structured answers.

SEO/AEO is the discipline of building and maintaining the reservoir, and everything else is plumbing that assumes the water exists.

An honest admission

Three things I’d want you to weigh against everything above.

First, the obvious one.

I sell this. Content engines and answer-layer builds are my business, so I have a direct interest in you believing this argument. The numbers are real and linked above, but so is my bias.

Second, AEO measurement is genuinely murky right now.

Engines disagree with each other (cross-engine citation overlap runs somewhere between 6 and 16% by one large analysis), the same engine gives different answers to different personas, and most of the tooling is young. E

ven the studies I linked disagree on magnitude (overlap estimates between AI citations and top-ten rankings range from 17% to 38% depending on who measured and when).

Anyone selling you a precise “AI visibility score” is selling directional confidence more than exact measurement. Directionally, you can absolutely see whether you’re present or absent in the answers that matter. Precisely? Not yet.

Don’t get me wrong, there’s value in directional metrics, but you leadership team needs to understand the difference before making major strategy shifts.

Third, humans haven’t left the process.

Forrester’s own analysts stress that buyers distrust AI answers enough to validate them with people, and Gartner’s data shows reps being used exactly that way. The answer layer shouldn’t replace your sales team, but it should decide what your sales team has to confirm or spend the whole call un-teaching.

What to do with this on Monday

Ask ChatGPT, Claude, and Perplexity the five questions a buyer in your category would ask. Screenshot the answers and put them in front of your sales, CS, and product leads.

Then ask one question: is this how we’d want the first sales conversation to go?

If the answer is no, you now know exactly why the content team’s SEO/AEO work belongs in the revenue conversation, funded like infrastructure instead of decoration.

In my piece on Anthropic’s watermark, I used an example of two writers. Writer A prompts one sentence and publishes whatever comes out. Writer B writes every word themselves, then pastes the piece into Claude to catch a comma they always drop. My claim was that both walk away with the same watermark.

That example was wrong.

Michal Viskup pointed it out first, then Shane Andrews and Benedict Evans made the same point. The watermark lives in the generation layer. Claude embeds it by statistically nudging word choices as it produces text. If the model only fixes a comma or swaps a semicolon without generating new words, there’s (most likely) nothing to carry the mark. Anthropic’s own blog post on August 14 (two days after I published) confirmed the mechanics.

So the proofreading example doesn’t hold. I pushed it too far, and I’m not editing the original post to hide that. The comments are still there (some are meaner than others, though Michal, Shane, and Benedict kept it classy, for which I really appreciated).

But here’s what I keep coming back to: the example was wrong and the core argument stands.

There are two ways to use AI when you write:

• In the first, you do the thinking. You bring the strategy, the transcript, the client story, the position you’ll defend in the comments, and you use the model to structure and sharpen what came out of your head.

• In the second, you hand the thinking over. One sentence in, 1,200 words out, publish under your name.

Those two users are not the same, and the watermark treats them identically.

Take an actual workflow I use every week. I record a conversation where I asked every question. The transcript is my thinking, captured verbatim. Then, I hand it to Claude and have it help me structure the text-based outputs from that conversation with re-prompting and editing along the way. The output carries the watermark just the same as the output of “write me a blog post on leadership in marketing.” One of those documents contains a person’s lived experience, while the other contains nothing.

The mark can’t tell them apart.

The mark measures contact rather than contribution. That was the core of the original piece, and three smart corrections later, it still stands.

What I’d actually want is a watermark for the process instead of the output. Something that could see what went into the context window.

• Did you feed it a transcript?

• Did you write out your own perspective first?

• Did the published version go through editing rounds that made it different from what the model handed back?

That’s the difference between using a tool and being replaced by one, and it’s exactly the thing the current implementation is structurally blind to.

One more thing worth saying. Michael, Shane, and Benedict corrected me without a shred of ego or meanness. That’s rarer than it should be on LinkedIn, and it’s the kind of conversation this technology actually needs. More of that, please.

The full addendum episode is up in the video above. Original post stays as-is, errors on my sleeve.

Talk soon,

Nathan

Every model released on or after August 2, everywhere Claude is offered, worldwide.

My first reaction was to try to work out what they had to gain from it.

I sat with that for a while and couldn’t come up with anything.

• It doesn’t make the product better.

• It doesn’t make the output more useful.

• It gives every competitor an obvious line of attack (or so I thought at the time)

• And it hands ammunition to the people who have spent three years arguing that anything a machine touched is inherently worth less (if not completely worthless)

If you were designing this as a business decision, it’s hard to see the upside. Then I read further and it clicked: this isn’t a business decision at all.

The EU AI Act’s Article 50 transparency obligations came into effect on August 2. They require providers of generative AI to make synthetic output machine-detectable. The penalty for non-compliance runs to 15 million euros or 3 percent of worldwide annual turnover, whichever is bigger. Around 190 organisations signed the accompanying Code of Practice before the deadline, including OpenAI, Google, Microsoft, Amazon, Meta, Mistral and Cohere.

So the loudest response I saw online, some version of “fine, I’ll switch models,” has nowhere to go. There is, or at least there soon won’t be, unmarked providers. That door closed by regulation before most people noticed it was open.

Which changes the question from “What was Anthropic thinking” to “What was the EU thinking, and whether the rule does what it was written to do?”

An example: The two writers

Here is the scenario that made me want to write this, and I’d encourage you to sit with it for a second because I think it’s the whole argument:

Writer A opens Claude and types one sentence: “write me a 1,200-word blog post about why B2B marketing teams should adopt AI.” Twenty seconds later there’s a post. They skim it, change the title, and publish it under their name. They didn’t have a single thought on their own (they simply rented the thinking for their $20/month subscription).

^^ Spoiler: I think we can all agree this is a gross misuse of the world’s best technology, though I’d argue it says nothing about the technology itself.

Writer B spends two hours on a piece. Their argument, their client story, the numbers from a project they actually ran, the position they’ll defend in the comments. Then they paste it into Claude and ask it to catch typos, because they always drop the same comma and they know it. Plus, they already pay for the tool, so they canceled their Grammarly subscription two months ago to consolidate tooling. Claude fixes four things, and they publish.

Both documents now carry the same mark.

Anthropic is upfront about this, to their credit. Their documentation says the mark signals that Claude processed the text, not that Claude authored it. Axios flagged the same thing this week: a comms team that uses Claude to clean up or format a human-drafted press release will stamp that release with an AI signature.

The mark measures contact, not contribution, and it was never built to.

That’s a reasonable engineering limitation. It becomes a problem the moment anyone treats the output as a verdict, which is exactly what will happen because people reach for the measurable thing in the room and this will inevitably be the only measurable thing in the room.

The exemption Anthropic can’t honour

I assumed, before I read the regulation, that business writing had been swept up in a rule aimed at deepfakes. That isn’t what happened, and the truth is something I find even more interesting.

Article 50 splits AI rules by role.

Paragraph 2 governs providers like Anthropic, and it’s clear that text counts: providers of AI systems “generating synthetic audio, image, video or text content” have to mark their outputs in a machine-readable format. Text sits right there beside video which was deliberate wording.

Paragraph 4 governs deployers, meaning the people who publish, and that’s where deepfakes live. It applies to systems that generate or manipulate “image, audio or video content constituting a deep fake.” Text is handled separately, and only when it’s “published with the purpose of informing the public on matters of public interest.”

Almost nothing in B2B marketing meets that bar.

So far, so reasonable. Then I reread the last sentence of paragraph 2, and I’ve been thinking about it since.

The marking obligation, the one Anthropic is complying with, ends like this:

“This obligation shall not apply to the extent the AI systems perform an assistive function for standard editing or do not substantially alter the input data provided by the deployer or the semantics thereof.”

By the same plain text of the paragraph that created the requirement, writer B is exempt. Someone who writes their own piece and asks a model to catch typos is using an assistive function for standard editing. The semantics aren’t substantially altered, and the regulation says the marking obligation does not apply.

And yet, the mark gets applied anyway.

The EU also wrote a second exemption into the deployer obligation in paragraph 4, for text that “has undergone a process of human review or editorial control and where a natural or legal person holds editorial responsibility for the publication.” Two separate carve-outs, in two separate paragraphs, both protecting the same person:

• the human who did the thinking

• and the human who took responsibility for the result

The drafters saw this coming enough to write the distinction into the law twice.

I don’t think Anthropic is ignoring it out of bad faith. I think a statistical watermark applied at the model level is structurally incapable of honouring it.

The model doesn’t know whether the 1,200 words in the context window are yours or its own. It can’t distinguish “fix my typos” from “write this for me,” because by the time it’s generating tokens, both look identical from the inside. The only implementation available is to mark everything, so everything gets marked.

Which leaves us somewhere genuinely odd.

The regulation grants Writer B an exemption. The technology built to satisfy that regulation can’t deliver it. And whoever eventually runs a detector over Writer B’s post will see a mark that, by the letter of the law that required it, shouldn’t have been there at all.

Paragraph 2 also obliges providers to make their marking “effective, interoperable, robust and reliable as far as this is technically feasible.” I keep coming back to that last clause. The honest reading of what shipped last week is that marking everything is what’s technically feasible right now, and the precision the law asked for isn’t.

The part nobody can check

The stranger detail is that the mark is live and the detector isn’t.

Anthropic has said a text detection API is coming, and an engineer on the Claude Code team confirmed it publicly. But it hasn’t shipped, so right now every piece of Claude output carries a signal that nobody outside Anthropic can read, with no published technical documentation, which means nobody can independently verify what it survives or how often it produces a false positive.

And there’s a real bind underneath that. One critic quoted in The New Stack put it bluntly: if Anthropic releases the detector publicly, they defeat their own watermark, because people will find reliable removal strategies by testing against it until the signal disappears.

But the other option is worse. The code of practice Anthropic signed obliges providers to support third-party detection. Keep the detector gated, or charge for access, and nobody outside the company can check whether the watermark does what Anthropic says it does. Announcing an API doesn’t fix that, it only makes verification billable.

In other words, if they give open access, it becomes an evasion tool. Keep it closed and the transparency measure can’t be verified by anyone but the company doing the marking.

So the tool is either unverifiable or self-defeating (it’s an impossible pick one since you only get to pick one).

Meanwhile an Anthropic engineer has already conceded the obvious: it’s not perfect, you can edit it, it’s a first step. Paraphrasing degrades the signal which means Writer A, the one who typed a sentence and published the output, can run the text through a second model and come out clean. Writer B, who wrote every word and asked for a spellcheck, gets marked and stays marked.

The mechanism catches the honest and misses the motivated. That’s the inverse of a working enforcement system, and it’s the thing that actually bothers me about all of this.

Why I’m not going to change how I work

There’s a version of me that reads all this and quietly stops running things through Claude before they ship. Simply keep the tool away from anything that gets published because contact is now (apparently) a liability.

I understand the instinct and I think it’s the wrong call.

I spent years being paid to write, and the value was never in the sentences. It was in the strategy; in having a position and knowing which part of a customer story actually mattered; in deciding what was true. I have read plenty of entirely human-written B2B content that contained nothing at all: generic, keyword-stuffed, written by a freelancer who had never used the product.

No machine was involved, and it was still worthless because there was no thought in it to deliver.

And I’ve read machine-assisted pieces that were genuinely good, because a person with real experience had a real stance and used the tool to structure it faster than they could alone.

For art, whether a machine was involved is a real question, and I’m not going to pretend to settle it. A novel is the thing itself, and the sentences are the product. For business writing, I think the question is close to meaningless: the prose is the delivery mechanism for the thinking, and the thinking is where the value lives.

The watermark only works as a “gotcha” if we’ve all agreed that being caught using AI to structure thoughts is bad. And I frankly don’t believe it is.

If a marked document is just a document that a machine touched at some point, it’s a fact about production rather than a judgment about quality. Every book on your shelf went through a printing press and nobody thinks less of the argument for it.

An honest admission

I‘m publishing this on Substack, which three weeks ago partnered with Pangram to let readers scan any post over 100 words for AI involvement. CEO Chris Best coined a term for the thing it hunts: Claudefishing, the mismatch when a reader unwittingly invests attention in something with no human thought behind it.

To his credit, Best is careful about the distinction.

He says outright that Substack isn’t against people using AI to assist their work. And Substack’s tool has something Anthropic’s watermark structurally can’t: a “How I make this” statement, where a writer declares their process in their own words.

That’s the difference between provenance and a verdict. One asks the writer to explain while the other hands a reader a number.

So I’m making an argument on a platform that has already picked a side, and I want to be straightforward about my own position rather than pretend I’m neutral. I use these tools every day, and I build systems around them for clients. I have an obvious interest in the answer being “this doesn’t matter much.”

My problem is more that there’s no way to watermark the context and prompting (i.e. the human thought and strategy) that goes into creating the watermarked output. And, as a result, people start treating Writer A and Writer B in the same category.

I fully believe they are two distinct types of users, and that there’s nothing wrong with how Writer B is operating the tool.

The thing I genuinely don’t know is how detection plays out, and neither does anyone writing confidently about it this week. The API doesn’t exist, and the documentation isn’t published. UCLA and UC San Diego both switched off their AI detectors in 2024 and 2025 after deciding the false positive rates were too dangerous to keep using, and those were classifiers rather than watermarks, but the institutional appetite for a clean number has not gone anywhere.

The most likely near-term outcome is a lot of confident claims built on a tool none of us can inspect.

Which is, when you think about it, the same complaint people have been making about AI writing all along.

It didn’t have adult content on it, and there was nothing inappropriate on the page at all.

It was just a clean little SEO audit: here’s the traffic this site gets, here’s its backlink profile, here’s how it ranks. Somebody types the site name into Google, lands on a marketing tool’s audit page, bounces immediately, and a session gets logged.

Millions of visits from this type of traffic, which makes for a beautiful chart and a great screenshot for the board deck.

But I have to imagine it has zero buyers.

I’ve written about this example before, and I think about that example a lot, because the people who built it weren’t stupid. They were doing exactly what content marketing rewarded them for doing. Volume was the scoreboard, and they won the game they were playing.

I just also believe they were playing the wrong game.

And I bring it up because it’s the cleanest version of a mistake I made myself, at a much larger scale (and got paid well to make).

The part where I’m the guy in the story

At Copy.ai, we ranked for “Instagram caption generator.” That one term brought in somewhere between forty and fifty thousand organic clicks a month.

There were dozens of free tools like it, which made sense when we were a product-led growth, self-serve startup.

Then the company moved from self-serve to enterprise B2B, and every one of those visitors became irrelevant overnight. Nobody searching for an Instagram caption generator was ever going to sit through an enterprise procurement cycle.

So I killed it.

Content responsible for about 140,000 monthly visits, gone. Traffic dropped from 350,000 to 210,000.

But enterprise pipeline went from zero to multi-millions in ARR.

That story has become a thing people quote back to me, usually as a bravery story. It isn’t. It’s a correction story.

The strategy was correct right up until the moment the company changed who it was selling to, and then it was actively harmful, and I was the last person to want to admit that.

What I actually learned there had nothing to do with traffic volume. It was this:

The funnel stages weren’t the problem. The source material was.

TOFU, MOFU, BOFU, and what actually shifted

Quick refresher, because I want us on the same page.

Top of funnel is the current state of things in your field. Somebody wants to understand their problem. High level, informational, higher search volume, nowhere near a purchase.

Middle of funnel is where things are headed. Expert POV and thought leadership. The reader is drafting a blueprint and deciding whether you’re a credible source to build it with.

Bottom of funnel is why you and why now. Comparisons, case studies, proof, the stuff that lets a champion justify a line item to their CFO.

None of that changed with AI. But what has changed (or should change) is what those stages have to be made of.

Top of funnel used to be an effort problem. You found a high-volume informational keyword, read the top-ranking pages, and out-thoroughed them. You went deeper in depth and usefulness. And that worked because producing genuinely complete informational content was slow and expensive, which meant effort was a real moat.

But effort at an informational level isn’t a moat anymore.

I can prompt Claude for a competent article on content marketing strategy and have it in ninety seconds. So can your competitor. So can the freelancer they hired who has never used your product.

Which leaves exactly one thing that still differentiates informational content: the material a model cannot generate because it doesn’t have access to it.

• Your podcast

• Your webinar transcripts

• Your product data

• Your sales calls

• Your actual opinion

Middle of funnel now feeds top of funnel

This is the structural change, and it’s the reason I recorded this week’s episode.

Your thought leadership shouldn’t sit in its own lane, produced by its own person, on its own calendar, measured by its own vanity metrics. It should be the raw input for your informational content.

• The quote from your podcast guest goes into the TOFU article.

• The proprietary data point from your product goes into the TOFU article.

• The position you took on where the industry is heading (the thing only you can say, backed by something you actually lived) goes into the TOFU article.

Without that, your top of funnel content is a slightly-better-formatted version of what everyone else can generate for free.

And the input for MOFU can also be your customer and sales call transcripts.

Three hundred calls is far too much for one person to sit through without over-indexing on whichever one they happened to listen to last. AI is genuinely, unglamorously excellent at this: read all of them, find the threads that repeat, hand a human the list.

Then a human decides what to write.

The same is true at the bottom:

• What objections keep surfacing

• What shipped this month that unblocks a stalled deal

• What your competitors quietly added to their changelog

All of it already exists inside your company and almost none of it is being routed anywhere useful.

Why most teams can’t do this

It comes down to how most teams are organized.

One person owns SEO, and one person owns thought leadership. Then someone else, usually a product marketer, owns sales enablement. That leaves you with three calendars, three sets of inputs, three editors, and three tones of voice.

Then the buyer shows up (not in order, because nobody moves through a funnel in order) and reads three things that don’t sound like they came from the same company.

But if you want consistency across the funnel, the funnel has to be fed from the same sources.

The honest admission

I don’t have this fully wired yet for my own business.

I have the pieces. The podcast produces transcripts. The blog has 400+ posts on it. The Brand Brain exists and my clients run on it. But my own sales call transcripts are not yet feeding my own content calendar, and I’m sitting here telling you to do the thing I’m two weeks behind on.

Pipes before chocolate applies to me too, and I’m currently making chocolate.

In other words, it’s not something that happens by accident. It’s not sexy work (it’s literally digital plumbing), but it’s the most impactful and something I’ll be setting up this week.

What’s in the episode

I walked through the full breakdown on this week’s Barely Shipping: what each stage is, what should feed it, and the specific source material most teams are sitting on and never using.

There’s a diagram on the blog if you want to see it laid out: https://systemsledgrowth.ai/blog/tofu-mofu-bofu/

If you only do one thing this week: pull your last thirty sales calls and find the three questions that keep repeating. That’s your next three articles, and you didn’t open a keyword tool to find them.

Reply and tell me what came up. I read all (most) of these.

Nathan

Systems-Led Growth

I’d just finished recording with Brian Sowards, a go-to-market engineer who builds agentic systems for operators inside fast-moving companies. Somewhere in our conversation he mentioned that a team of three opened $100 million in channel-source pipeline. When I first read that, I assumed it was a typo. It wasn’t. And the reason it wasn’t is the whole point of this episode.

Stop aiming AI at the cheap work

When most teams got real AI, they went hunting for the most repetitive, low-value tasks to knock off the list. Brian’s take on that is the cleanest reframe I’ve heard all year: there are no low-level tasks.

(i.e. nobody’s biggest problem is publishing fifty blogs they already wrote).

For that team of three, the expensive work was activating an entire distribution network in a way that set up real conversations. The relationship-driven slog that everyone knows moves pipeline and that big teams never get organized enough to actually run.

AI has removed the logistics tax that kept teams from doing their best work in the first place. That’s how three people do what a department can’t while still building those real human relationships (remember those, anybody?).

We got into the proof from a few angles: the Aviatrix turnaround (80% less marketing effort, $10M in new pipeline in a single month), why complex enterprise sales turned out to be the real unlock instead of simple ones, and Brian’s line that I think is the entire anti-SaaS argument in one sentence: take the most important thing in your business, and there isn’t software for it. So you build it.

The part I can’t stop thinking about

Here’s the thing, though. The number that hooked me isn’t what stuck with me after the call ended.

Near the end, Brian said something I haven’t been able to put down:

Everyone is overwhelmed right now.

Not just the junior folks. The VPs, the directors, the C-suite, the board. He thinks the whole “996” thing is underselling how much people are actually working, and worse, nobody knows what “enough” looks like anymore because no one is willing to say “this is the bar, you cleared it.” He admitted he works more now that he has AI agents than he did before.

So do I, if I’m honest.

Unfortunately, a real transformation isn’t something you coast through on a tidy schedule. It’s the stuff you do early, late, and on the weekends.

I’m not going to wrap that in hype, because the hype is exactly what I started this podcast to avoid. The leverage is real and it’s also genuinely hard right now. Both things are true.

But Brian landed it somewhere I didn’t expect, and I’ll just pass it along the way he said it:

The answer he’s found isn’t waiting inside your org to be handed to you. It’s in deciding for yourself what’s worth charting toward, and then building the personal system that gets you there. The org-wide AI still has big gaps, but the personal AI assistant is here today, and it compounds the moment you start.

That’s a message worth more than a pipeline number, and it’s why this episode is one of my favorites so far.

In this episode

• Why there are no low-level tasks, and where AI’s real leverage actually lives

• Aviatrix: 80% less marketing effort and $10M in new pipeline in one month

• Moxie Power: activating a relationship-driven channel to $100M with a team of three

• The software bifurcation, and why you build the tools that are core to your business instead of buying them

• Why marketing is in the prime seat in the AI era after years as the punching bag

• The honest culture conversation: overwhelm, “996,” and why everyone’s lost on context

• Why Brian’s clients walk away owning their AI stack instead of renting it

Listen or watch

The full episode is above. You can also watch it here:

https://systemsledgrowth.ai/podcast/there-are-no-low-level-tasks/

This newsletter is free, and it’s staying that way. I’m sharing what I build and what I learn in the open, every week, as the ground keeps shifting under all of us.

If someone you know runs a lean team and is tired of AI takes that don’t survive contact with a real number, forward this their way.

Talk soon, Nathan

Systems-Led Growth

About an hour and a half in, the kids needed a break. I was standing in waist-deep water looking at the bigger sets rolling in further out. Not huge or dangerous or anything like that, but bigger than anything I was comfortable riding.

I asked the instructor if he’d take me out there while the kids rested, and in a classic surfer-dude fashion, he said “Yeah, let’s do it.”

So we paddled out (while I started to regret asking in the first place).

I tried to catch a few waves and ate it every time. And these weren’t graceful wipeouts, either. They were the kind where you’re embarrassed to pop your head back out of the water because then you’ll have to see who saw.

After watching me fail on three or four waves, the instructor said something that carried more weight with me than he probably intended:

“You keep falling because you' keep looking back at the wave. If you want to stop falling, you need to look where you want to go.”

Then he explained it a bit more: when you’re lying on your stomach and you feel the wave building behind you, the instinct is to look back at it. Especially when they’re bigger than what you’re used to.

The wave looks massive from that angle, it’s scary, and your brain wants to track the thing that feels like a threat.

But if you look back, you fall backwards. If you look down, you fall forwards. You will always go in the direction you’re looking.

His instruction for the next wave was simple. Get in position. Feel the wave start to take you. Puff out your chest and look up.

Stand up with your eyes fixed on where you want to go and don’t look back.

And I’m not just trying to make a great story when I tell you that I caught, rode, and celebrated the next wave (the biggest one I’ve ever ridden).

I’ve been thinking about that a lot lately.

The past couple of months have been a transition for me. Some days I have total clarity on where I’m headed.

Other days, if I’m being honest, I’m looking back at what feels like a tidal wave.

I’m not the only one. The job market in B2B right now is brutal.

People with real talent and real track records are getting laid off, or stuck in roles that are changing underneath them faster than they can (or care to) adapt. Smart, experienced people are watching everything they know change all around them, and they’re questioning what we’re all questioning: what’s next?

But here’s what I keep seeing: the people who are struggling the most are the ones who can’t stop looking back:

• Back at the title they used to have

• Back at the team they used to lead

• Back at the way the industry worked eighteen months ago.

I get it. The wave is scary from that angle. It looks like it’s going to crush you.

But, as I’m still learning, you can’t ride it facing backwards.

I finally have some clarity on where I’m going, and I’ll be sharing more about that next week. But this newsletter isn’t about my plan. It’s about one piece of advice from a surf instructor in Tofino that I keep coming back to whenever things feel uncertain.

Look where you want to go because you’ll go where you’re looking.

More next week.

Nathan

I killed 140,000 monthly visitors on purpose. Traffic at Copy.ai went from 350k to 210k in 2024/2025, but enterprise pipeline went from zero to multi-millions.

In my view, the numbers that went down were the right ones to cut.

That experience forced me to confront something the industry doesn’t often say out loud: most B2B content was already bad before AI showed up. But Romana Kuts, founder of SaaStorm, hopped on LinkedIn and said that verbatim.

When I read that post, I knew I had to reach out to see if she’d chat with me on Barely Shipping. I thought it was a long shot because of how impressive her resume is (plus, it’d be my first official guest). One highlight was taking a climate tech SaaS client from near-zero search visibility to €3M in pipeline in 12 months.

But, to my surprise, she agreed to hop on a call and tell me her thoughts on SEO, AEO, pipeline over pageviews, and how she’d advise entry-level employees to build experience with AI.

Here are four things from our conversation that stuck out to me:

1. You need to audit your SEO fundamentals before worrying about AEO

Romana was doing solid SEO for clients before LLMs existed. When AI search showed up, those same clients started appearing in AI-generated answers without her doing anything extra.

Zero additional work.

She credits this as being because LLMs pull from the same sources Google rewards (accurate, clear, helpful content from credible domains). The fundamentals haven’t changed at all but taking lazy shortcuts have simply stopped working.

The audit takes one afternoon:

• Pull your top 20 pages

• For each one, ask three questions:

  • Does it answer the query directly?

  • Does it have a proper FAQ section?

  • Is the H1 the actual question the reader typed?

If the answer to any of those is no, fix those before you add a single new tactic.

The agencies charging $2,000/month as an “AEO add-on” to their existing SEO retainer?

In her view, that’s largely a rebrand on basics they should have been delivering since 2019.

2. Add one human-made asset to your next three blog posts.

Here’s what Romana started doing with clients: Loom videos embedded directly under the H1. A 3-minute recording from a subject matter expert answering the core question of the article.

She’s not doing this because she has some insider information that the algorithms are rewarding Loom videos specifically. But she (correctly) believes that these are the types of elements that keep humans on the page when they see it.

Before LLMs, we got away with walls of text because the algorithm didn’t care. Now, the content that was always missing the human layer is getting outranked by content that has it. Video was always needed, but most people were just ignoring it because they could.

3. Commit to 90 days before you judge the system.

Romana’s climate tech client got to €3M because the client trusted the process long enough for it to compound. For anyone in content marketing, that’s rarer than it sounds.

Most CEO’s agree to a 3-month retainer and start asking for results in week two.

Her system had four layers:

• Topics defined with subject matter experts.

• Content briefs through Romana, then through the client’s content manager and SME for approval.

• Then back to Romana for SEO compliance.

• Finally back to the writer, and on to the client.

Same process for three months before results appeared (measured in pipeline instead of traffic).

If you’re building a content system right now, write down the process before you start. Define the review stages, name the people in each stage, set a 90-day check-in date and don’t evaluate ROI before it.

The compounding doesn’t start until the system has run long enough to have something to compound on.

4. Before your next content brief, make someone on your team demo the product.

Romana had a client whose marketing manager had never logged into the product. You read that right… two years into the job and this manager didn’t even have their logins.

Every piece of content that team produced had a quality ceiling built into it from the first sentence. You can’t write accurate, specific, helpful content about something you’ve never used.

And being accurate, specific, and helpful is the entire job now.

This is the pre-brief ritual: before anyone writes anything, the person writing it opens the product and spends 20 minutes using it. If they don’t have a login, get them one before the brief is written.

The content that ranks (and that LLMs cite) is the content that sounds like it was written by someone who knows the product from the inside. Because it was.

Bonus: How to Get Experience When You’re Brand New

The most common question I get from people early in their careers right now is some version of: “How do I compete with people who have 10 years of experience, when AI is changing everything anyway?”

Romana answered this better than I could have.

She got sick for a few days, came back to LinkedIn, and half her feed had apparently mastered Claude, built insane systems, and shipped more in a week than she had in a month.

But her reaction wasn’t panic.

It was: “We’re all figuring this out in real time. The stream is moving fast enough that experience only gives you pattern recognition, not a permanent lead.”

That’s the most honest thing anyone has said about this moment. And it has a practical implication: the advantage right now goes to whoever ships fastest, not whoever has been in the field longest.

Here’s what that looks like in practice.

1. Get a $20 Claude subscription and build something for yourself

Something you care about. Your own blog or a workflow that automates something you do manually. Maybe a simple content brief template for the niche you want to work in. You learn more from one broken workflow than from three certifications.

2. Find one person doing the work you want to do and ask them a specific question

Ask a specific question: “I’m trying to build a content brief workflow in Claude and I can’t get the tone consistency right. How do you handle that?” Romana mentors on GrowthMentor, and she’s a great resource for anyone who needs it.

Most practitioners at this level will answer a direct, specific question because it’s interesting (remember, vague outreach often gets ignored).

3. Publish what you’re building, even when it’s broken

This is the one that separates people who grow fast from people who don’t. The instinct is to wait until it’s perfect. The right move is to document it while it’s in progress. “I tried to build X. Here’s what broke. Here’s what I changed.” That kind of content builds an audience of people who are trying to do the same thing.

It also forces you to think clearly about what you’re building, which makes you better at building it.

4. Skip the certifications

Romana didn’t say this, but I will. You can’t certify judgment. The people selling “AI marketing certifications” right now are going to look like the people who sold “social media marketing certifications” in 2012.

There’s no piece of paper that can give you actual experience. You need to get the reps in.

The 22-year-old who ships something real in the next 90 days has a better shot than the 32-year-old who’s waiting to feel ready. The field is moving fast enough that the gap between “just starting” and “legitimate practitioner” has never been shorter.

Final Thoughts

I spent three years at Copy.ai watching smart people confuse traffic for traction. We had 350,000 monthly visitors and an enterprise pipeline that was effectively zero. The content was technically correct, reasonably well-written, and almost entirely useless as a business asset.

Killing 140,000 of those visitors was the right call. It was also uncomfortable in a way that’s hard to explain until you’ve done it.

What Romana and I kept coming back to in this conversation is that AI didn’t change the definition of good content. Accurate. Clear. Helpful. That’s what it always was.

What changed is that the shortcuts that let you fake those three things without doing the underlying work have stopped working. The algorithm got smarter, the reader got more options, and the companies still producing content the old way are going to feel it in their pipeline before they feel it in their rankings.

The pipes matter more than the chocolate (and they always have).

But as someone who’s done their fair share of ghostwriting, here’s how it actually worked at most B2B companies: a marketing coordinator interviewed the VP or CEO for 20 minutes. Took notes. Went back to their desk and wrote a draft from memory. The draft went through four rounds of review. Each round sanded off another sharp edge because legal was allergic to strong opinions and brand wanted everything to sound “professional.”

By the time it published, it sounded like a press release written by committee. Twelve likes from employees who were told to engage. And then it died a slow and boring death.

Three months later someone asked, “Whatever happened to our thought leadership initiative?”

And the cycle started again.

The problem was never that the executive didn’t have interesting things to say. They usually did. The problem was that the production model required them to write (which most executives don’t have time for) and the approval model stripped out everything interesting.

AI changed the production model, but the thought leader’s job stayed the same.

The Quality Spectrum

Here’s how I think about thought leadership quality now, ranked from best to worst:

Best: Real conversation → AI structures it → human edits the prose.

1. The thought leader talks for 45 minutes. That conversation becomes a transcript.

2. AI extracts the arguments, structures them into a narrative, drafts the article in the speaker’s voice.

3. The thought leader (or an editor who knows their voice) reviews it, adds a personal detail, sharpens a sentence, cuts what doesn’t land.

4. The output has the substance of the original conversation and the polish of editorial attention.

Good: Real conversation → AI structures it → published without editing.

Same process, minus the editorial pass.

The prose might be less refined. A sentence might be awkward. A transition might feel mechanical. But the core insight is preserved because it came from the person who actually lived it. The substance is real even if the craft could be better.

Worst: Ghostwriter interviews exec → writes from memory → four rounds of review strip out every edge.

The prose might be beautiful and the structure might be clean. But the substance has been diluted to the point where the content could have been written by anyone.

The sharp opinions that made the executive interesting in the first place got removed because someone in the review chain was uncomfortable with them.

Here’s where things get uncomfy for a lot of writers: that second tier (AI-structured, unedited) contains more real insight than the third tier (ghostwritten, polished to death).

The prose is rougher, but the thinking is sharper. And the thinking is what the reader came for.

Why This Works: The Source Material Changed

The old model started with a 20-minute interview and a ghostwriter’s memory. The new model starts with a full transcript of the actual person saying what they actually think, in their actual words.

That’s a fundamentally different input.

A transcript captures the specific story about the deal that almost fell apart, the specific number from last quarter’s pipeline, the specific opinion about why the industry is wrong about a particular trend.

A ghostwriter captures a summary of a summary, filtered through their own understanding (or misunderstanding) of what the exec meant.

When AI structures content from a transcript, it’s working with the real material. When a ghostwriter drafts from memory, they’re reconstructing a version of the material.

The gap between those two inputs shows up in every output.

How to Build This System

The system has three phases: before the conversation, the conversation itself, and after.

1. Before: research the person.

Pull their recent LinkedIn posts, any published content, their company’s recent news. Generate tailored questions based on this research. These should be specific enough that the person feels prepared for, not ambushed by, a generic “tell me about your role” prompt.

“I noticed your team recently expanded into healthcare. What surprised you about selling to compliance-heavy buyers?” is a better question than “What’s your go-to-market strategy?”

The prep takes a workflow five minutes. It used to take a human six hours plus four expensive lattes. And the result is a better conversation because the questions are sharper.

2. During: have the actual conversation.

This is the irreplaceable human part: two people talking.

The best thought leadership comes from conversations where someone says something they haven’t said before, an insight that emerges from the dynamic of the discussion. The interviewer’s job is to listen, follow threads, and ask the question that makes the other person pause before answering.

Record everything.

Video if possible (for clips), audio at minimum (for transcription). This sounds obvious, but I’ve seen companies host excellent conversations and realize afterward that nobody hit record.

3. After: run the transcript through the repurposing workflow.

This is where one conversation becomes ten assets:

1. Full-length article in the speaker’s voice

2. LinkedIn post (shorter, punchier, hook in the first two lines)

3. Newsletter draft for your email audience

4. YouTube description and show notes

5. Quote cards (3-5 per conversation, each a standalone opinion)

6. Sales talking points (specific insights formatted for reps)

7. Social clips (2-3 best moments marked for video extraction)

8. Content library entries (claims, data points, customer language, tagged and stored)

9. Landing page (for high-value conversations)

10. Follow-up sequence for attendees or subscribers

A 45-minute conversation, ten assets, and one workflow. Human review on each asset before it goes live is the ideal state.

The thought leader spent 45 minutes talking about what they know, but they never wrote a single word.

• The system did the production.

• The human did the quality review.

And the content sounds like a person because it started as a person talking.

Where It Breaks

I’ll be honest about two problems I haven’t fully solved.

1. Quality variance across assets.

The full-length article needs the most editorial work because long-form writing is where tone and voice are hardest for AI to get right. LinkedIn posts and quote cards tend to be closer to publishable because they’re short and structurally simpler. Sales talking points are often excellent because they’re extractive (pulling out specific claims) rather than generative (creating new prose).

I’ve learned to budget my review time accordingly. The article gets 30-45 minutes. The LinkedIn posts get 10 minutes total. The quote cards get a glance. The sales talking points get a quick accuracy check.

2. Input quality determines output quality.

The system multiplies whatever goes in.

If the person has sharp, specific opinions and real stories from their work, all ten assets are strong because the raw material is strong. If they speak in generalities and corporate platitudes, all ten assets are weak because you can’t extract specificity from vagueness.

The system doesn’t fix a thought leader without thoughts, and it’s spectacular at distributing boring content really efficiently.

This means the highest-impact investment in the whole system is choosing who to put on the mic and preparing the right questions.

• People who write with specificity on LinkedIn tend to speak with specificity in conversation.

• People who share real numbers publicly tend to do the same on the mic.

• People who speak in buzzwords publicly will give you buzzwords in the recording.

Choose your inputs carefully, and the right system handles the rest.

The Wrong Debate

The industry keeps arguing about “AI content vs. human content.” That framing misses the point entirely.

The real comparison is AI from the source vs. a ghostwriter who never lived it.

When the AI’s input is a real transcript of a real person sharing real opinions, the output carries the weight of that person’s experience. When a ghostwriter’s input is a 20-minute interview filtered through their own interpretation, the output carries the weight of a telephone game.

The question was never “human or AI.” The question is: did the real person’s thinking make it into the final product?

If the answer is yes, the method doesn’t matter nearly as much as people want it to.

If the answer is no, all the beautiful prose in the world won’t save it.

Nathan Thompson is the author of Pipes Before Chocolate. He builds AI-native content systems for B2B companies with lean GTM teams.

Earlier this year, I set to work on revising it for 2026 and, today, I’ve decided to pre-release it here on SubStack.

It’s free. Here’s why.

For the past three and a half years, I’ve been helping run lean a marketing team across multiple B2B SaaS properties. Content, SEO, demand gen, sales enablement, AEO.

Many, many hats, all made possible with AI.

During that time I built systems that let me do the work of a department: a content engine that produces five ICP-focused articles per day, outbound workflows that generate personalized sequences for 50-75 accounts per week, inbound processing that responds to leads in under two minutes instead of the industry average of 47 hours.

I deliberately killed pages driving tens of thousands of visits because they attracted the wrong people. Traffic went from 350k to 210k. Pipeline went from zero to millions for our enterprise audience.

Every system I built, I documented.

And the result is Pipes Before Chocolate: 15 chapters covering how to build AI-augmented go-to-market systems with a skeleton crew. Content, sales outbound, inbound processing, thought leadership, ABM, events, case studies.

Plus four appendices including a 30-day build plan.

The book’s core argument: the companies that will pull ahead in the next 12-18 months are the ones who lay the right digital plumbing for their entire content base.

Again I’m calling this a pre-release.

The full book is a Google Doc PDF right now, not a polished final product. I wanted to get it into the hands of my SubStack and LinkedIn community first, before the official launch, because the people who’ve been following along deserve early access.

Grab it here:

https://drive.google.com/file/d/1-94K6uqX0XkV68gM21jziq8qhN39acdA/view?usp=sharing

A couple of things about what’s inside:

1) Every system I describe, I’ve built.

Every number I cite from my own experience is real. Where something didn’t work, I say so. Each tactical chapter ends with an honest admission about what’s still broken.

2) This isn’t a beginner’s guide to AI.

I’m assuming you’ve used Claude or ChatGPT at least a handful of times. It’s also not a book about prompting. It’s about what happens after the prompt, when you need to connect the output to everything else your business is trying to do.

Over the next few weeks, I’ll be publishing the sharpest ideas from each chapter as standalone posts here on SubStack.

Consider those the highlight reel for the book, which is the full playbook.

If you read it and build something from it, I want to hear about it. Reply to this email or find me on LinkedIn. That’s not a marketing line. The best part of writing this was hearing from people who actually implemented the systems and got results.

More soon.

Nathan

P.S. Did I use AI to generate this book?

Yes and no.

Like the first edition, all of the thinking (the part that counts) was created through transcripts of me speaking to myself while walking my dog, Zoomer.

AI then took those transcripts, restructured everything, and created the first draft. Over the past 6 weeks, I’ve read and edited each chapter.

Yes, you’ll likely find a few AI stock phrases in there. Also, yes, I’m OK with that.

The goal of this book wasn’t to show you that I can pour blood, sweat, and tears over a typewriter.

The goal is to show you that I’ve poured blood, sweat, and tears into AI systems-thinking over the past three and a half years, and to share the results of what I’ve learned to date.

Not even a little.

The notion that marketing and sales should operate as one connected system, with shared data, structured content, and automated handoffs between functions, has been floating around B2B for at least fifteen years.

It’s gone by many names, like:

• Marketing Operations

• Revenue Operations

• Demand Generation Infrastructure

• Growth Engineering.

If you’ve been in B2B long enough, you’ve sat through a conference talk or hired a consultant who told you to break down your silos, connect your data, and build cross-functional workflows.

They were right. Every single one of them was right.

So why didn’t every company actually ever pull it off?

Why was the theory spot on, but implementation could never cross over the finish line?

What Systems-Led Growth Costs

I’ve seen what it actually took to build connected go-to-market systems before AI. A dedicated ops team, usually three to five people.

Six to twelve months of implementation.

$200k to $500k in tooling and integration costs.

A marketing automation platform talking to a CRM talking to a data enrichment provider talking to a content management system talking to a BI tool, and every connection between them required either a native integration that was never quite flexible enough or custom API work that required engineering resources marketing never had priority access to.

The RevOps movement of the early 2020s tried to solve this by creating a dedicated function: a team whose entire job was building the connective tissue between systems.

For the companies that could afford it, it worked.

But those were overwhelmingly enterprise companies with dedicated ops headcount, mature data practices, and the budget to stitch together a dozen specialized tools.

And now that money isn’t free, many companies are realizing their processes + tech stacks were so bloated, they needed to lay off many of those early 2020 hires.

For everyone else, connected operations remained a dusty slide deck from the offsite and a Jira ticket that never made it to the top of the backlog.

What Changed

AI didn’t change the strategy. Connected systems where every input compounds across every function has been correct for fifteen years. What AI changed is the implementation layer.

Three things, specifically.

First, AI collapses the logic layer.

Before AI, connecting a sales call transcript to a personalized follow-up required a rules engine: if the prospect mentions X pain point, attach Y case study, reference Z value prop. Building that meant mapping every possible path, coding the logic, and maintaining it as your messaging evolved.

AI replaces the brittle rules engine with a flexible reasoning layer. You give it your value props, your case studies, your ICP definitions, and it maps the right components to the right context.

Second, AI handles transformation work that used to require specialized roles.

Turning a call transcript into a blog post, a set of talking points, and a follow-up email used to require three people (or one very overworked person doing all three badly… I’ve been that guy).

AI handles the conversion while preserving context across each output. The human still decides what to transform and reviews the results. But the assembly work that used to eat 80% of the time compresses to near-zero.

Third, AI makes the data layer queryable without engineering.

The structured content library I describe in the book would have required a database engineer, a front-end developer, and an API specialist to build five years ago.

Today, a solo operator can set up Supabase table, populate it with tagged content and strategic context, and have every workflow query it in real time. The barrier between “I know what data I need” and “my system can access it” dropped from months of engineering to an afternoon of setup.

What Didn’t Change

These three shifts don’t make the work easy. The gap between “possible” and “easy” is where most of the disappointment in AI lives.

Building connected systems still requires deep strategic thinking about your ICP, your positioning, and your content quality standards. It still requires knowing what to automate and what to keep human. It still requires maintenance, because data drifts and workflows need tuning and content goes stale.

None of that changed.

What changed is that you don’t need an enterprise ops team to do it anymore. One to three people who understand the strategy and have the patience to build the infrastructure can now construct systems that would have required a department and a year of integration work three years ago.

The playbook that used to require a team of five and a $300k tool stack can now be run by a team of one with a $500/month infrastructure budget.

The Two Kinds of Readers

If you’ve been in RevOps or marketing ops for years, you’ll read the book and think “I’ve heard this before.” You’re right.

The difference is that the system you’ve been trying to build with a six-person team and a six-figure budget can now be built by a skeleton crew. If you’ve been preaching connected systems for years only to be told the company can’t afford it, this might be the most satisfying “I told you so” of your career.

If you’ve never built these systems, if you’re a marketer or founder who inherited a pile of disconnected tools and a mandate to make pipeline happen, the advantage you have is that you’re starting from scratch.

You can build the connected system the ops people always wanted, using tools that didn’t exist when they were fighting for budget.

The destination is the same place it’s always been. The only thing that’s changed is that now you can actually get there.

The book comes out April 16th. It’s free, and I’ll be releasing it here on Substack.

• Blog posts in a CMS

• Case studies in a Google Drive folder

• One-pagers in someone’s Canva account

• Sales decks in three different versions across four reps’ desktops

• An eBook from 2023 that marketing spent six weeks on and nobody has opened since launch month.

This is the raw material your AI workflows will pull from. And if you feed a system disorganized, outdated, inconsistent content, that’s exactly what it produces on the other end.

Good in, good out. Garbage in, garbage out.

Before you build a single workflow, you need to know what you actually have, what’s worth keeping, and how to structure it so every department can find it.

Think of it this way: if you hired a writer with a photographic memory, someone who could recall any piece of content your company has ever produced, what would you actually hand them?

Would you really drop an entire Google Drive in their lap? Hopefully not. Hopefully, you’d curate and give them the materials that represent how your company talks, what your customers say, and what your sales team needs on a Tuesday afternoon when a prospect asks for proof you’ve done this before.

That’s the content audit, and rather than being a content audit as an SEO exercise, it’s really a content audit as a systems exercise.

What goes into the library

The goal of this exercise is selecting what earns a spot in the system that feeds your workflows. Six categories matter.

1. Case studies

These are the connective tissue between awareness and decision. A prospect reads your blog post and thinks interesting. They read a case study from their industry and think these people understand my problem.

Every case study should be tagged by industry, company size, pain point addressed, outcome delivered, and the products or features involved.

If it isn’t tagged in some logically way, then it doesn’t really exist to the system (at least on practically speaking).

2. Blog posts and long-form content

Not all of them. Only the ones that reflect your current positioning, address real buyer questions, and still hold up. A post from 18 months ago about a feature you’ve since sunset doesn’t belong. A post that explains the problem your product solves, in the language your buyers use, does.

Be ruthless here.

Most companies keep content alive long past its usefulness because deleting feels wasteful.

3. One-pagers and sales collateral

Per use case, per persona, per vertical. These are the assets your reps reach for mid-deal. If they exist in a format a workflow can pull from, structured and tagged, the system can auto-generate personalized versions per account.

But if they exist as a PDF buried in Slack, they’re functionally invisible.

4. Customer quotes and language

This is the most underrated category. Pull direct quotes from case study interviews, G2 reviews, support tickets, sales call transcripts. Tag them by theme, by pain point, by persona. When a workflow needs to generate a follow-up email or a landing page, these quotes give it language that sounds like your buyers, not like your marketing team.

5. Guides, eBooks, and gated content

Only include what’s still relevant and what represents your current thinking. A guide you wrote two product versions ago will teach the system the wrong things. If it’s outdated, archive it. If it’s current, tag it by topic, funnel stage, and ICP segment.

6. Internal playbooks and process docs

How your team talks about the product, handles objections, positions against competitors. These are essential inputs for workflows that generate outbound messaging, battlecards, or call prep docs.

The audit process

Step 1: Inventory: Pull every content asset into one spreadsheet. URL or file location, title, content type, date created, last updated. This takes a few hours, and yes, it’s tedious.

Do it anyway.

Step 2: Filter for relevance: Go line by line. Ask two questions: does this still reflect how we talk about our product? Would I hand this to a new hire and say “this is how we do things here”?

If the answer to either is no, it goes in the archive pile, not the library.

Step 3: Tag what survives:

Every piece of content that makes it through gets tagged across at least four dimensions: content type (case study, blog, one-pager), funnel stage (awareness, consideration, decision), ICP segment (who is this for), and topic cluster (what problem does it address).

• If you serve multiple verticals, add industry.

• If you have named competitors, add competitive context.

Step 4: Identify the gaps: Once everything is tagged, the holes become obvious. You have twelve blog posts about awareness-stage topics and zero case studies in your fastest-growing vertical. You have a competitor battlecard for Company A but nothing for Company B, who shows up in 60% of your deals. You have plenty of content for marketing managers but nothing for the CFO who signs the check.

These gaps become your content roadmap. A list of assets your system actually needs to function rather than a list of TOFU keywords. Both matter, but the former matters more at this stage.

Step 5: Structure for retrieval: A content library is only useful if people and workflows can find things in it. This means a searchable layer: a database, a tagged CMS, an internal tool where anyone on the team can search “enterprise case study, healthcare, ROI data” and get results in seconds.

If a sales rep has to message you on Slack to ask “do we have anything for financial services,” the library isn’t working yet.

What changes when the library is built

The library is the foundation that makes everything else possible.

A sales call transcript flows through a workflow that matches the prospect’s pain points to your tagged case studies and generates a follow-up email with the right proof attached. That workflow only works if the case studies are tagged and stored somewhere the system can reach.

An ABM landing page gets auto-generated for a target account with their industry-specific case study, relevant customer quotes, and a tailored value prop. That page only works if those assets exist in a structured format.

A new rep joins and searches “objection handling, enterprise, security concerns” and gets thirty customer quotes, two case studies, and a battlecard. That search only works if someone ran the audit, did the tagging, and built the retrieval layer.

Every tactical system in your go-to-market depends on this library existing and being maintained. The workflows are the pipes, and the content library is what flows through them.

The honest part

This audit is not glamorous work (nobody, to my knowledge) has ever gotten promoted for tagging case studies. It takes a full day, maybe two, and the output is a spreadsheet and a tagging system, not a dashboard you can show your CEO.

But every company I’ve seen struggle with AI workflows has the same root problem: the inputs aren’t structured. They skip straight to building the workflow and wonder why the outputs sound generic. The system can only pull from what you give it.

Run the audit first. Then tag it. Then structure it. Then make it searchable.

And then build the workflows connected to that.

The pipes work better when you know what’s flowing through them.

This is adapted from Pipes Before the Chocolate, a book about building AI-augmented go-to-market systems. The content library is the foundation. The book covers what you build on top of it.

The second version of the book is coming out April 16th, right here on Substack.

The majority of teams are stuck on Level 1. They know it, they can feel it, but they just don’t have the language for what Level 2 and Level 3 look like.

Here’s the framework.

Level 1: Chat (a task)

You open ChatGPT or Claude, write a prompt, and get an output.

• Summarize this sales call.

• Draft this blog post.

• Rewrite this email.

Each request is standalone, and often disconnected from any centralized docs on an orgs brand voice, tone, or messaging. But you still get a quick-win result that’s “good enough”, so you close the tab, and start from scratch when the next task comes along.

This is where the majority of companies are. 96% of B2B marketers report using AI in their roles.¹ 95% of B2B organizations use AI-powered applications.² Those numbers sound impressive until you realize that for most of them, “using AI” means asking it to do one thing at a time from a bunch of individuals within the same org.

The productivity gain is real but limited. You might save 20 minutes on a first draft or get a decent outline faster, but each task exists in isolation.

Level 1 is “using AI.” It’s where everyone should start, but it’s not where anyone should stay.

Level 2: Workflows (a process)

You chain tasks together so the output of one becomes the input for the next.

For example, a sales call transcript flows through a system that extracts pain points, maps them to your value propositions, generates a personalized follow-up email, creates a custom one-pager, and tags recurring themes for the content team.

One input, multiple outputs, and all of them are connected to a bigger picture.

This is where the economics change. Production time collapses, and quality stays high because the system enforces structure. Plus, every output feeds the next workflow instead of disappearing into a Google Doc nobody opens.

The shift from Level 1 to Level 2 is the most important jump most teams will make. It’s also where I’ve spent the last three years of my career.

Building workflows at Copy.ai in early 2023 was the moment this clicked for me: the distinction between asking AI to do one thing and building a process where each step informs the next.

That distinction is the philosophical core of everything I’ve been building since.

But it’s not just anecdotal evidence; the research backs this up. AI-augmented roles show a 37% productivity improvement, compared to 12% from traditional automation alone.³

The difference is whether you’re using AI as a tool or as part of a larger system.

Level 3: Agentic AI (an autonomous system)

Agents that make decisions, take actions, and operate with minimal human oversight within defined parameters.

An agent that monitors your CRM, pulls account research, generates personalized outreach, and flags high-priority accounts for human review. Or an agent that watches your content performance, identifies decaying pages, and drafts refresh briefs before you even know there’s a problem.

This is where the industry is headed (but my fear is that it’s headed there too quickly… more on that in a moment). Over 40% of enterprise applications are expected to embed task-specific AI agents by 2026.⁴ Gartner predicts that by 2028, 33% of enterprise software will include agentic capabilities.⁵ McKinsey reports that nearly 50% of companies with over $5 billion in revenue are already scaling agentic AI in at least one function.⁶

At Copy.ai, we’re soon releasing an agentic version of our workflows to make them more accessible, not to mention more powerful, than what we’ve spent the last three years building.

But here’s the part most people skip: you can’t jump to Level 3 without building Level 2 first.

You Can’t (or Shouldn’t) Skip Levels

This is the insight that saves teams months of frustration.

Companies that jump straight to agentic AI without building workflow infrastructure first end up with what I call autonomous chaos.

• An agent that makes decisions without structured data to pull from only makes bad decisions faster.

• An agent that generates content without a defined brand voice generates off-brand content at scale.

• An agent that personalizes outreach without mapped value propositions personalizes the wrong message.

30% of generative AI projects are expected to be discontinued after pilot stages, mainly due to data quality issues and unclear business value.⁷ But it’s not always a technology failure. In fact, it’s more often a sequencing failure.

People tried to build Level 3 automation on a Level 1 foundation.

The build order matters:

Level 1 → Level 2: Map your processes. Document your inputs and outputs. Build the workflows that connect them.

This is where most of the value lives for teams.

Level 2 → Level 3: Once your workflows are producing reliable, structured outputs, you can start adding decision-making layers. Agents that monitor, flag, and draft within the guardrails your workflows define.

The unsexy truth is that Level 2 is where teams should spend most of their time right now. A well-built workflow system with human oversight will outperform a poorly scoped agentic system every time.

The Question to Ask Yourself

Where is your team right now?

If you’re still in Level 1 (individual prompts, standalone tasks, no connection between outputs) the move isn’t to buy an agent platform. The move is to pick one process (content production, sales follow-up, inbound processing) and build the workflow that connects it end to end.

Yes, it feels like homework. But that’s a good thing. If there were ever a time to go back to school and master the basics, it’s now.

If you’re in Level 2 (workflows running, outputs structured, processes documented) then you can start thinking about which workflows are stable enough to add a decision-making layer.

If you’re trying to jump to Level 3 and wondering why it feels chaotic, the answer is almost always: go back and build Level 2 properly.

The Worst-Case Scenario

Here’s what you don’t want: half your org stuck on Level 1 doing random acts of AI, while a handful of power users quietly build their own Level 3 agents without telling anyone. The Level 1 people are pulling brand voice from memory while the Level 3 people are pulling it from whatever docs they happened to have saved locally.

Marketing’s agent says one thing. Sales’ agent says another. The customer sees both.

This is the fragmentation problem, and it gets worse the more capable your team is. The power users aren’t doing anything wrong (they’re doing what smart people do: building tools to move faster). But without a central source of truth for everyone to pull from, every tool encodes a slightly different version of your brand.

This only serves to scale inconsistency.

The fix is neither slowing down the power users nor speeding up the beginners. It’s building the layer underneath both of them: the structured, codified knowledge base that any tool, at any level, pulls from so that everything comes out on brand regardless of who built it or how.

I’ve spent the last three and a half years building that layer across multiple B2B SaaS properties as a one-person growth team. The systems, the workflows, the architecture underneath the AI. I wrote all of it down in the second edition of my book, Pipes Before the Chocolate.

The updated version comes out April 16. It’s totally free, and I’ll be releasing it here on SubStack, so if you’re already subscribed, you’ll get it in your inbox.

More soon.

Notes

[1] Demand Gen Report, “2026 B2B Trends Research Report,” March 2026.

[2] Content Marketing Institute, “B2B Content and Marketing Trends: Insights for 2026,” December 2025.

[3] MedhaCloud, “AI Adoption Statistics 2026,” citing McKinsey research. 37% vs. 12% productivity improvement.

[4] AppVerticals, “AI Chatbot Adoption Statistics,” February 2026, citing Gartner.

[5] Gartner, enterprise software agentic capability forecast, cited in Kissflow, 2026.

[6] McKinsey, cited in MedhaCloud, “AI Adoption Statistics 2026.” Nearly 50% of $5B+ companies scaling agentic AI.

[7] Gartner, cited in multiple 2025-2026 AI adoption analyses. 30% of GenAI projects discontinued after pilot.

I stopped writing here because I wasn’t sure what I wanted this to be. For a while I’d been running a few experiments (like Signals, a weekly AI news roundup that was, honestly, an experiment in using AI to curate AI news... meta, I know).

The system worked beautifully. I built a workflow that pulled the week’s biggest stories, summarized them, organized them by theme, and packaged them into a clean newsletter. It was fast. It was consistent. And it was a genuinely fun thing to build. But I didn’t love editing it, and didn’t feel much pride when I re-read it.

And if I’m being totally transparent with you, I’m not sure anyone needed it.

Anyone can get an AI news summary now. ChatGPT will do it in 30 seconds, and Perplexity will give you the links. There are a dozen newsletters doing the same thing with bigger teams and better sources. The experiment proved the system could work, but it didn’t prove the system was worth running.

So I stopped. And then I got busy doing the thing I actually wanted to write about all along.

Here’s what the last five months have looked like:

I’ve been rebuilding the internal marketing systems at Fullcast from the ground up.

SEO, content, AEO, thought leadership, sales enablement. All of it. All running through connected workflows designed for a skeleton crew (which, if you know me, you know is how I operate). The mandate was the same one I’ve had for the last three years: build systems that let one person do the work of a department.

The results have been, frankly, something I’m quite proud of.

1. Organic traffic on relevant keywords is higher than it’s ever been on the Fullcast site. I want to be specific about that word “relevant” because I learned the hard way at Copy.ai what happens when you optimize for the wrong traffic. These aren’t students Googling “sentence rewriter.” These are revenue operations leaders searching for solutions to problems we actually solve.

2. AI mentions have more than doubled since I started the AEO framework. More importantly, we’ve had prospects get on calls and tell us they booked the demo because they found us through ChatGPT. Those are my favorite Slack messages to read in the morning, and they’re happening more frequently.

3. Copy.ai is still going strong. The systems I built there continue to compound, which is the whole point of systems-led growth. You build the infrastructure once, and it keeps producing. Relevant traffic is improving. The content engine is still running. The workflows I built in 2024 and 2025 are still generating pipeline in 2026. And it’s not because I’m any kind of genius, by the way. It’s because the system was designed to compound, and compounding is patient.

4. The thought leadership workflows have been one of the most satisfying builds. I’ve connected them directly to call transcripts and interview recordings, which means the content we publish is built from real conversations with real people about real problems. This signals genuine E-E-A-T to both Google and the answer engines, but more importantly, it means the content is actually good. It’s not some AI-generated summary of what everyone else has already said, but actual insights from practitioners, structured by a system, and reviewed by a human who knows whether it’s worth publishing.

5. I’ve also built more internal tools than I can count at this point. Outline generators, content brief builders, competitive analysis workflows, deal scoring calculators, sales prep tools, account research automations. Some of them are great, and some of them taught me what not to build next time (I’ve started calling those “educational investments” instead of “failures” because it sounds better at dinner parties and quarterly reviews).

But here’s the thing I didn’t see coming: I’m now working more directly with customers to help them nail down their own processes. Consulting, essentially. Helping marketing teams and founders figure out what their workflows should look like before they try to automate anything.

I’m getting to work with out incredible Solutions Consultant team, and it’s been remarkably educational on a whole other side of this world of AI: change management.

Turns out the hardest part of building AI systems isn’t the AI but getting people to clearly define what their process actually is.

You’d be amazed how many teams say “we need to automate our content production” and then can’t tell you the steps their current content production follows. You can’t automate what you can’t describe (and you definitely can’t hand it to an agent).

If that sounds familiar, it’s basically the thesis of Chapter 3 of the book I’m rewriting for this year. But more on that another time.

---

So that’s what I’ve been doing instead of writing newsletters. Building systems. Testing them. Breaking them. Fixing them. Helping other people build their own.

And what I realized, somewhere between walking Zoomer and staring at a Supabase dashboard at 11pm, is that this is what I should have been writing about the whole time.

Signals is dead. Long live Systems-Led Growth.

Here’s what that LinkedIn newsletter is now: weekly breakdowns of the actual, messy, honest work of building go-to-market systems with AI. What I’m building, what’s working, what’s not, and everything in between.

Each issue will be one of three things:

1. A system I built: how it works, what it cost, and how you can build it too

2. Something that broke: why it broke, and what I changed

3. A framework or workflow you can steal and implement this week

Nothing fancy. Just the work from someone who does it every day, usually while also managing SEO across multiple properties, consulting for friends, raising two kids, and trying to keep his beagle from eating things that aren’t food (I’m really great at two of those things, trying my best at the third, and ironically I’m total garbage at the fourth).

If you stuck around through five months of silence, thank you. Genuinely. I don’t take that for granted, and I intend to make it worth your time.

If you’re new here, welcome. I’m Nathan. I build AI-augmented go-to-market systems for skeleton crews, and I document what I build. That’s the whole pitch.

Let’s lay some pipes.

---

Next week’s issue will be something that’s been on my mind recently: Is there still a world for workflows, or will it all be agentic?

Spoiler: I have strong opinions on this, and the “agents will replace everything by Tuesday” crowd is not going to love them.

Everyone says they want their teams to use AI. But here’s the contradiction: leadership often demands “better results with AI” without giving people the time, tools, or safety to figure it out.

This equates to teams feeling pressure without clarity, curiosity without direction, and fear without support.

Let’s fix that.

The Contradiction at the Heart of AI Adoption

Most leadership teams say:

• “AI is the future of work.”

• “Be innovative. Be creative.”

• “Transform how we operate.”

But what they really mean is:

• Do it faster.

• Don’t take too much time experimenting.

• Don’t mess anything up.

This is the type of vague pressure that leaves marketing teams spinning in circles. Employees today are wrestling with a set of impossible expectations:

• Experiment without time: No space in the calendar for testing.

• Adopt without freedom: “Use AI, but not that tool” (privacy restrictions).

• Improve without budget: No resources for subscriptions or training.

• Build without security: The fear that what they make will replace them.

On top of that, the tech is moving faster than strategy. Tools ship weekly. Models change monthly.

Leadership hasn’t caught up, and it leaves teams guessing.

The Four Pillars of AI Adoption

Here’s how leaders can flip the script and actually set their teams up for success:

1. Incentivize Experiments

Reward the act of testing, even if it fails. Experiments are supposed to fail sometimes. That’s kinda’ the whole point (otherwise, it’d be a set strategy)

2. Block Protected Time

Don’t just tell people to “make time.” Schedule it. Two to three hours a week, blocked off for AI play and follow-ups on what worked (and what didn’t).

3. Provide Real Resources

Give your team safe environments, clear training paths, and communities to learn from. Decide if you want them leaning into chat, agents, or workflows (and provide support for each).

4. Normalize Failure

Failure is not a career risk, it’s a learning moment. Share stories of failed experiments and how they led to breakthroughs. Make this cultural.

Pitfalls to Avoid

• Set it and forget it: One-time training doesn’t cut it. AI is moving too fast.

• The “silver bullet” trap: No single tool solves everything. Let teams explore.

• Results-only obsession: Focus on improving job quality and creativity before demanding pure productivity.

• Fear-based messaging: AI is an opportunity, not a threat. Communicate it that way.

What Happens When You Get This Right

When teams have time, resources, and psychological safety, curiosity comes alive. Curiosity sparks creativity. Creativity leads to better solutions.

And that culture of experimentation? It becomes your competitive edge.

The companies that win with AI won’t be the ones who adopted the most tools. They’ll be the ones who built the strongest cultures around curiosity, creativity, and support.

The Takeaway

If you want AI adoption, stop obsessing over the tools. Invest in your people instead:

• Time

• Budget

• Mental safety

Get those right, and your team won’t just use AI—they’ll run ahead of your competition by the end of the year.

What’s dead is the idea of random acts of marketing, just publishing to stay visible without any real purpose.

AI has lowered the barrier to entry. With ChatGPT, Canva, and endless templates, anyone can publish. Which means the hard part isn’t creating content anymore. The hard part is creating content that matters.

Too many teams are still churning out blogs, webinars, and social posts just to stay active.

But activity ≠ outcomes.

If you don’t know what success looks like for each piece, you’re just making noise.

The Two Magic Questions

Five years ago, I started running all of my content through two simple filters. Today, they’re more important than ever:

1. What do they get? What does your audience walk away with?

   1. Information?

   2. A process?

   3. Inspiration?

   4. Confidence to solve their problem?

2. What do we want? What do we want to happen after they consume it?

   1. Build trust?

   2. Earn a demo?

   3. Reduce churn?

   4. Strengthen customer relationships?

If your content can’t clearly answer both, it’s not purposeful marketing. It’s just hoping that something sticks to the wall.

The Trap of One-Sided Content

• All for the audience, none for you: The “Ultimate Guide” that gets shared widely but never ties back to your product.

• All for you, none for the audience: The “We, we, we” press release that nobody cares about.

The sweet spot is both.

Like a case study: your audience learns how to solve a problem, and you earn credibility by showing how your product made it possible.

The Challenge

Look at your last five pieces of content. For each one, can you answer:

• What did the audience get?

• What did we want as a business?

If you can’t answer both, you’re not marketing. You’re just dreaming.

The era of random acts of marketing is over. The future belongs to purpose-driven publishing, content that creates value for your audience and moves your business forward.

So ask the two questions. Every time.

But here’s the problem: voice without rules is chaos.

That’s where editorial guidelines come in. They’re not the same thing as voice, and if you confuse the two (or skip documenting them), you’ll end up with inconsistent content, frustrated editors, and AI systems that don’t know what to do with your drafts.

Today’s crash course explains exactly how to fix that.

Brand Voice vs. Editorial Guidelines

Think of brand voice as your personality, the feeling you want your audience to have when they read, listen, or watch your content.

• Liquid Death vs. Bubble (Bubbly?): both sell sparkling water. Their voices couldn’t be more different.

Now, think of editorial guidelines as the rulebook that voice must follow.

• Do we use the Oxford comma?

• Do we write headlines in Title Case or sentence case?

• Do we say “customers” or “clients”?

• Are contractions allowed?

Voice = the emotion.

Guidelines = the rules.

Why Editorial Guidelines Matter

• Consistency builds trust. Every typo or style shift chips away at credibility.

• Scaling requires rules. Writers, editors, and AI systems can only follow what you’ve clearly documented.

• AI needs frameworks. Voice alone doesn’t teach AI how to format, punctuate, or spell. Guidelines do.

Without guidelines, brand voice becomes a vague aspiration instead of a repeatable system.

Channel-Specific Guidelines

At a minimum, you need three sets of editorial guidelines. One is for your website, one is for social, and one is for email. Here are some quick examples:

• Website:

  • Professional, clear, conversion-driven.

  • Title Case for headings.

  • Avoid contractions in main copy.

• Social Media:

  • Conversational and concise.

  • Sentence case for posts. Contractions allowed.

  • Emojis permitted (max two).

• Email Marketing:

  • Personal but professional.

  • Sentence case subject lines, no emojis.

  • One clear CTA per email.

Each channel is a different room you’re walking into. You don’t change who you are, you just adapt how you show up.

And remember: you can take this even further. Many brands benefit from splitting guidelines more granularly:

• Within social media, you may want separate rules for LinkedIn, Instagram, and TikTok, since tone, formatting, and audience expectations vary.

• You can also create cross-channel guidelines per ICP segment. For example, your messaging to a CMO might lean more formal across all channels, while messaging to an end user might allow for more casual, playful language.

The deeper you define these distinctions, the easier it becomes to keep every message aligned, no matter the platform, audience, or creator.

Putting It Into Practice

1. Document your rules. Simple checklists with examples work best.

2. Make them accessible. Keep them in one place where your team and tools can use them.

3. Train your AI editor. Feed your guidelines into Copy.ai, ChatGPT, or another tool so it enforces your rules automatically.

This saves time, reduces errors, and keeps every message on brand.

Final Takeaway

Brand voice makes you memorable. Editorial guidelines make you consistent.

Together, they’re the playbook your team (and your AI) needs to scale content without losing the thread.

A voice.

The consistent personality that shows up across everything you publish, no matter the format.

Today, let’s break down why brand voice is different, how to codify it, and how to make it repeatable across channels.

What is a Brand Voice?

Think of your brand voice like a personality.

• It’s not what you say, but how you say it.

• It lives in vocabulary choices, sentence rhythm, and tone.

• It should feel recognizable even when the medium changes.

Example: the way you talk to your closest friends is different from the way you talk to your partner’s parents. Both are “you,” but adapted to the context.

Your brand should have that same range. One that’s adaptable, but never inconsistent.

Why Voice Isn’t the Same as Editorial Guidelines

Editorial guidelines tell you the rules: grammar choices, formatting, punctuation, capitalization.

Voice tells you the vibe.

• Editorial: “We use Oxford commas.”

• Voice: “We sound approachable, confident, and a little playful.”

You need both. But without codifying voice, your content will come out sounding robotic (or worse, like ten different people wrote it).

How to Codify Brand Voice Across Channels

Here’s a simple, repeatable system:

1. Define the Core Personality

   • Is your brand witty or straightforward? Warm or authoritative? Optimistic or pragmatic?

   • Pick 3–4 traits that describe your voice.

2. Map Voice to Channels

   • Social: shorter, punchier, more casual.

     • These are obviously examples and will vary from brand to brand

   • Blog: clear, explanatory, with room for storytelling.

   • Email: personal, inviting, conversational.

3. Build Examples You’d Actually Use

   • Don’t leave it vague. Show exactly how a tweet, a blog intro, and a sales email should look when written in your voice.

   • No placeholders, no “almost right.” Real, repeatable samples.

4. Create a Living Doc

   • House these examples in your brand brain.

   • Update them as you publish and refine.

The Bottom Line

Most teams obsess over what to say: the message, the campaign, the CTA.

But your brand voice—the how—is the multiplier here. It’s what makes a blog post sound like you. What makes an email instantly recognizable. What makes your audience trust that, no matter the channel, they’re hearing from the same brand.

Codify it once. Apply it everywhere. That’s how you make your brand brain smarter and your marketing stronger.

On Friday, we’ll cover the docs you’ll need to make channel-specific editorial guidelines.

Welcome to the very first episode of Nathan.ai, the podcast that runs alongside the Forever Free Substack. This show is about building content systems, testing copy, and using AI to scale what already works (without losing your brand’s soul in the process).

In today’s episode, I talk about why AI won’t fix a broken message… it just amplifies it. If your message is unclear, harmful, or misaligned, AI doesn’t save you. It makes the problem louder, faster, and more damaging.

We cover three big points every marketer should hear:

• AI = Sociopath → AI doesn’t care about truth or empathy. It just scales whatever you feed it.

• Broken Messaging = Broken Outcomes → If your foundations aren’t clear, AI multiplies the chaos.

• The Goal Hasn’t Changed → Business is still about solving human problems. AI only accelerates the speed at which you can connect solutions to those problems.

You’ll also hear:

• Why speed without direction is just chaos.

• How “spam the TAM” is the wrong play in 2025.

• A simple order of operations for using AI the right way: clarify → test → then scale.

At the end, I leave you with a reminder: AI is a big ol’ megaphone. If you’re yelling the wrong thing, it doesn’t matter how loud you get.

Hey folks,

I’ve posted a bit sporadically over the past few months, but mostly because I promised to never let a personal brand interfere with my professional work. And I’ve been heads down on some pretty fun stuff!

But today I wanted to talk about content systems. I made the video above, and what you’ll read below is an AI generated, human edited text-based follow along.

Checkbox Marketing Are Motion Without Progress

Checkbox marketing is what happens when activity substitutes for intention.

• Posting on LinkedIn because “we should be active.”

• Publishing random blogs because “SEO is important.”

• Throwing together a webinar because “everyone else is doing it.”

It feels productive. You can point to a checklist and say, “Look, we’re doing stuff.”
But without a clear outcome, it’s just busy work.

I call this “work-around work.” Planning, brainstorming, ideating… but rarely orchestrating toward a common goal with a unified plan.

Checkbox marketing is the treadmill: lots of effort, no forward motion.

Content Systems Are Built With Intent

A content system looks almost identical on the surface (after all, you’re still checking boxes in a system to make sure the system is running smoothly). The difference is those boxes are designed to move you toward a specific outcome.

That’s why I love the Henry Ford vs. Willy Wonka analogy:

• Henry Ford’s factory was a system designed to mass-produce affordable cars. Precision, efficiency, repeatability. His goal was clear: make automobiles cheap enough that the average worker could own one.

• Willy Wonka’s factory (yes, fictional) was also a system. It just had a different goal: make candy a magical, visceral experience. His pipes, machines, and processes were all engineered for delight.

Two factories. Both efficient. Both intentional. But radically different outputs because they were built around radically different goals.

Different Goals, Different Systems

That’s where marketing teams often miss the mark.

They build content without first asking: What do we want this to achieve?

• More organic traffic?

  • → You need a repeatable system for top-of-funnel content.

    • Keyword research → AI-assisted briefs → human editing for search and story → consistent publishing.

• More demo sign-ups?

  • → You need a system for thought leadership and conversion.

    • Deep conversations → webinars → repurposed clips → nurture flows leading to demos.

• More social reach?

  • → You need a system for audience-native content.

    • Break one strong insight into multiple engaging formats designed for the channels your audience actually hangs out in.

The system changes depending on the outcome.

Checkbox marketing just hopes everything magically comes together. Systems are built to make progress inevitable.

Two Questions to Keep You Honest

Every piece of content should answer two things:

1. What do they get? What’s the value for the reader, viewer, or listener? An insight? A tool? A perspective?

2. What do we want? What’s the business outcome? Traffic, sign-ups, demos, pipeline?

If you can’t answer both, you’re checking a box, not building a system.

This Week’s Challenge

Audit your own marketing.
Be brutally honest:

• Are you running a system?

• Or are you just checking boxes?

Pick one goal for the next quarter and sketch the system you’ll need to reach it.

✍️ Thanks for reading. And if you want more on building content systems that actually drive pipeline, hit subscribe. I’ll keep this Substack forever-free :)