When the Buyer Is a Model

The armoring industry sold fear. That was the category premise when I arrived at Texas Armoring Corporation in 2008, and it was not a stylistic choice, it was the entire logic of the business. Every competitor’s marketing was built on the same implicit sentence: something terrible is going to happen to you. It worked, in the narrow sense that it closed deals with people who were already afraid. It also permanently capped the category, because fear is a terrible foundation for a purchase somebody has to feel good about for the next decade, and because a person who is frightened buys once and never tells anyone.

We rejected the premise and repositioned the category around trust, responsibility, and status. That was not a campaign. A campaign is a thing you run and then stop. This was a decision about what the product meant, and it is the reason a decade of compounding growth followed instead of one strong year. Revenue went from two million to twenty-five million dollars. It was built almost entirely on demand the brand earned rather than demand the company purchased, including a video series that generated billions of cumulative organic views across YouTube, Reddit, Facebook, syndicated news, and global media, with no performance media spend behind it at all.

I bring that up not as a credential but as a data point about mechanism, because the mechanism is the part that transfers, and the mechanism is now being tested against a genuinely new condition: for the first time in the history of marketing, a meaningful share of your buyers will encounter your category through something that is not a human being.

Section 01Brand was always a memory system

The most useful idea in modern marketing science is also the least romantic. John Dawes of the Ehrenberg-Bass Institute wrote it down in 2021 for LinkedIn’s B2B Institute, and it has since become known as the 95:5 rule: at any given moment, roughly five percent of your potential business buyers are in the market, and ninety-five percent are not.¹ The arithmetic underneath it is mundane. If a company changes a given service provider about once every five years, then twenty percent of the market is in play in a year and roughly five percent in a ninety-day window.

Dawes has been careful to call it a heuristic rather than a law, and the honest version is that the ratio varies by category and you should calculate your own from your actual interpurchase interval.² But whatever the ratio turns out to be, the implication does not change, and the implication is brutal for the way most companies spend money.

This reframes advertising entirely. If most of the people you reach cannot buy, then advertising is not persuasion and never was. It is the construction and refreshing of memory links that will activate later, in a buying situation that has not happened yet. Jenni Romaniuk’s phrase for it is that you cannot push buyers down a funnel, but you can catch them as they fall.

Bain put a number on the consequence that I think every marketing leader should have taped to their monitor: roughly eighty-five percent of B2B buyers purchase from a vendor who was already on their day-one list, formed before they did any research at all.³ If that is even directionally right, then the funnel most companies fund with most of their budget begins after the decisive event has already occurred. You did not lose that deal in the evaluation. You lost it eighteen months earlier, in a moment you never observed, when somebody built a mental shortlist and you were not on it.

That is why I have never treated earned demand as a euphemism for cheap. Paid demand rents attention on a metered basis and stops the moment the meter does. Earned demand builds an asset that keeps paying, because the memory structure persists in the buyer’s head whether or not you spent anything this quarter. The AK-47 videos were still generating inbound inquiries years after they were made, from people who had never seen an ad.

Section 02The second reader

For twenty years, the memory that mattered was human. There is now a second one, and it does not work like the first.

The measurable shift is not subtle. Similarweb tracked zero-click searches rising from fifty-six percent to sixty-nine percent between May 2024 and May 2025, a thirteen-point move that maps almost exactly onto the AI Overviews rollout. Pew’s behavioral study, which tracked nearly sixty-nine thousand real searches rather than asking people what they thought they did, found click-through at roughly eight percent when an AI summary was present against fifteen percent when it was not. Seer Interactive’s longitudinal analysis of tens of millions of impressions put the organic CTR decline on AI Overview queries at around sixty-one percent. In Google’s AI Mode, which replaces the results page rather than sitting above it, the zero-click rate approaches ninety-three percent.⁴

Before anyone builds a budget on those figures, the caveat matters more than the figures do. This research is young, much of it is vendor-produced, and the methodologies are not comparable. Across the studies that have measured AI Overview impact on click-through, the estimated decline ranges from about fifteen percent at one end to nearly ninety percent at the other, depending entirely on the keyword set and the method.⁵ Every serious study agrees on the direction. None of them agree on the magnitude. Plan against the direction. Do not build a forecast against any single one of those numbers, and be suspicious of anyone who does, particularly if they are selling the remedy.

Section 03Why this is the old physics on a new substrate

Here is the part I find genuinely clarifying rather than alarming. When a search engine answered a query, it ranked ten links and let the human choose. When a model answers a query, it does not rank ten. It names two or three, in prose, with reasons.

That is not a search result. That is a consideration set, and consideration sets are the exact object the Ehrenberg-Bass tradition has been measuring for forty years. Mental availability, the probability that a brand comes to mind in a buying situation, has just acquired a second substrate. The buyer’s memory still matters. There is now also a retrieval layer that has read what the world has written about you, and it is forming its own shortlist on a similar principle: what is salient, what is corroborated, what is distinctive enough to be worth naming.

So generative engine optimization is not a new discipline requiring a new team. It is the discipline you already had, with a second reader who has different reading habits. Five things change:

1. You are optimizing to be described, not clicked. The unit of success is an accurate restatement of your position inside somebody else’s answer. That rewards writing that makes a clear, specific, falsifiable claim and punishes the vague consensus prose most B2B companies publish. If your positioning statement could be pasted onto four competitors’ websites without anyone noticing, a model has no reason to name you and no material to name you with.
2. Entity beats keyword. A model does not retrieve strings; it retrieves things with relationships. If your company exists on the web only as a name that appears near some adjectives, you are not retrievable. If it exists as an entity with resolvable relationships to people, products, categories, and credentials, you are. This is what structured data is actually for, and it is worth doing properly.
3. Third-party corroboration is the ranking signal. Models weight what others say about you far above what you say about yourself, which is precisely the logic of earned media. The discipline that produced tier-one press has not been made obsolete; it has been handed a second consumer who reads everything, never forgets, and cannot be flattered.
4. Blocking crawlers is a visibility decision, not a security decision. Similarweb’s brand visibility work found that publishers who block model access rank far below what their actual readership would predict, while narrow specialist sites with structured content and open access punch dramatically above their search demand.⁶ Whatever you decide, decide it knowingly, at the executive level, rather than discovering it was set by a default in a configuration file.
5. Measure share of answer, not share of traffic. If a model names three vendors and you are one, you own roughly a third of that answer. That is a legitimate metric and it correlates with the thing you care about. It also comes with a compensation: the visitors who do arrive by way of an AI citation appear to convert at a substantially higher rate than ordinary organic traffic, because the model has already done the qualification.⁷ Fewer, better.

Section 04A failure of my own, offered as evidence

I rebuilt the entity graph for my own site for exactly the reasons above, defining the person, the organizations, the work, and the relationships between them in structured data so that a machine could resolve “Jason Forston” from a string into something with edges.

It did not parse. For weeks. There were two syntax errors buried in the block, and the consequence was that every structured-data consumer on earth received precisely nothing, while the page itself looked flawless. Nothing on screen changed. No error appeared anywhere a human would look. There was a second, subtler bug underneath it: duplicate keys, which JSON permits and silently resolves by keeping the last one, so several of the relationships I thought I had declared had been quietly discarded.

Section 05What does not change

Every technological shift in marketing produces a wave of people insisting the fundamentals are dead, and they are always describing a change in the delivery mechanism as though it were a change in human behavior. I have watched this happen with search, with social, with mobile, with video, and now with retrieval. The medium has changed four times in my career. The mechanism has not changed once.

A model, exactly like a person, can only retrieve what was distinctive enough to be encoded in the first place. A category in which every company says the same eight things in the same register gives a retrieval system nothing to differentiate on, and it will fall back on whatever proxies for authority it can find, which will mostly favor incumbents. The way out is the same way out it has always been: say something true that nobody else in your category is willing to say, and then be structurally capable of backing it up.

The AK-47 series worked because no other armoring company on earth would have made it. It was not a content strategy. It was a category argument delivered as an artifact, and it was only credible because the product actually performed the way the argument claimed. That is the whole formula, and it survives the arrival of the machines intact, possibly strengthened, because a system trained on the aggregate of everything written about your industry is exceptionally good at detecting which companies are saying the same thing as everyone else.

I have spent twenty years building demand that a brand earned rather than demand a company purchased, and the single most consistent finding across all of it is that the compounding comes from the decisions that were hard to make, not the ones that were expensive to execute. Rejecting the fear premise cost nothing and returned a decade. Buying the traffic instead would have cost a fortune and returned a quarter.

Ninety-five percent of your buyers cannot buy from you today. What you do about that is the entire question, and it always was.