The AI model didn't read your press release. It read your customers.

Roughly six in seven citations behind an AI answer about your brand point somewhere you don't own. Two good diagnoses of the AI visibility problem are both arguments about the other seventh.

Dmitrij Żatuchin, of Rankfor.AI and the Estonian Entrepreneurship University, recently published an analysis of where AI answers about brands actually come from. Not what the answers say, which is where most of this research sits, but which sources they are grounded in. Across 167,551 URL-grounded citations covering 128 brands, 12 markets and 13 languages, 85.7% pointed to sites the brand does not own. Owned properties accounted for 14.3%.

Two caveats before that number does any work. The dataset is European, drawn from Nordic-Baltic, Polish and Central European markets, so the exact ratio should not be assumed to hold in the Gulf or Southeast Asia. And Żatuchin runs a company selling AI brand intelligence, a competing interest he declares. The underlying data is published openly, which is more than most vendor research offers.

Even discounted heavily, the shape of the finding is hard to argue with. The thing a model reads when someone asks about you is overwhelmingly not written by you.

That should reframe the two most serious diagnoses of the AI visibility problem published in the last few weeks. Both are right about something important. Both, we would argue, are arguments about the 14.3%.

Diagnosis one: you are in the wrong sources

Writing for Onclusive, Aya Elbaoudi made the case that earned media volume and AI visibility have decoupled almost completely. A company can hold an enormous clippings file and be functionally absent from AI answers, while a challenger with a fraction of the coverage owns the category. The variable is not how much you are covered but whether you are covered in the specific sources a model has learned to trust.

The same citation study supports her directly. Roughly 80% of citations came from about 18% of domains, a distribution steep enough to fit a Zipf law. Concentration is real, it is measurable, and being absent from the short head is expensive. This diagnosis is correct.

It is also, necessarily, a diagnosis about placement. It tells you which doors to knock on. It does not tell you what the people behind those doors will say.

Diagnosis two: you are incoherent and it now shows

Edelman's Andrew Mildren makes a sharper argument. AI did not fragment your brand. It made a fragmentation that was already there impossible to ignore.

Organisations have always tolerated different units, regions and functions telling slightly different stories about what the company does. That was survivable because nobody encountered all of it at once. The prospect saw the deck, the analyst read the annual report, the customer used the product, and each stayed in its lane. A model collapses the lanes, synthesising everything simultaneously, and whatever incoherence exists surfaces in the answer.

The supporting evidence is genuinely striking. Żatuchin's earlier category-ownership study, covering 3,750 responses across 50 brands and five industries, found three leading models agreeing on the top-recommended brand in only 41.6% of categories. INSEAD Knowledge found the same effect at brand level: asked about Airbnb, Llama weighted uniqueness, ChatGPT local options, Perplexity flexibility. The same research found Ariel taking almost a quarter of category mentions on Llama and under 1% on Gemini, and Chanteclair holding 19% share on Perplexity while vanishing from Llama altogether. As INSEAD put it, there is no page two on LLMs.

Mildren's conclusion is that the job description has changed. Marketing is moving from managing messages to managing interpretations.

That is the most useful sentence written about this subject so far. It is also where we want to pick up the argument, because it raises a question it does not quite answer: interpretations formed by whom, on the basis of what?

A wrinkle in the coherence diagnosis

Before that, a note on the evidence. Cross-model divergence is being read as proof of self-inflicted incoherence, and only part of it is.

If models disagreed because your account of yourself was internally contradictory, you would expect them to be similarly confused, all reading the same muddled record and arriving at comparably muddled answers. That is not the pattern. They are differently confident. Each names a clear winner and the winners differ, which is what you get when three systems retrieve different slices of a large third-party corpus, not when three systems read the same incoherent brochure.

Governing your own language will not fix that, because your own language is 14.3% of what they read.

What both diagnoses assume

Both accounts share a premise: that the record a model learns from is one your organisation authors, directly or by influence. Place better, brief tighter, govern the language across regions, and the model updates.

That premise held when discovery was assembled from documents you could commission. It holds much less well against a corpus where most of the citations point somewhere else.

Look at where description actually lives. G2 and Capterra reviews. Reddit threads where someone asks the category question and forty strangers answer it. Comparison posts by affiliates who have never spoken to you. Support forums. Community answers. YouTube comments under a review video. LinkedIn posts from customers explaining why they switched. Procurement discussions in trade forums. Trustpilot. Glassdoor, which shapes how a model characterises you as an organisation even when the question was commercial.

Almost none of that is a channel. All of it is a consequence.

Corroboration beats authorship

Mildren identifies the mechanism correctly. A human buyer is asking whether they believe you, and the answer runs through relationships, direct experience and emotion. A model is asking whether there is enough consistent, credible evidence to describe or recommend you with confidence, and it infers that through corroboration and agreement across independent sources.

Follow that through and the implication is uncomfortable. A model does not weight a claim by who made it. It weights a claim by how many independent sources make it and how consistently they make it. Your website says you are the fastest implementation in the category. That is one source, and a self-interested one. Four hundred customers independently writing that they were live in three weeks is a categorically different kind of evidence, and the model treats it that way.

So the thing you are optimising is not your messaging. It is the agreement between your messaging and the unprompted testimony of people who have actually used the product. Where those align, the model forms the confident, stable picture Mildren describes and reaches for you readily. Where they diverge, it either resolves in favour of the corroborated version, which is to say the customer version, or it hedges. In practice, hedging means naming someone else.

Elbaoudi's playbook identifies four gaps behind most underperformance: presence, position, narrative and source. We would add a fifth, and argue it sits underneath the others. Call it the corroboration gap: the distance between what you say about yourself and what your customers independently say about you, measured across the surfaces a model can actually reach.

Perfect internal governance does not close that gap. You can align every region, retire every legacy deck, publish one canonical description of the business, and a model will still read ten thousand reviews written by people whose account of you was formed by what actually happened to them. Coherence you author is necessary. At 14.3% of the citation base, it is nowhere near sufficient.

This is what we mean when we say experience quality has become a distribution channel. Not as a motivational slogan. As a mechanism. What you deliver determines the volume, consistency and specificity of what customers write about you unprompted, and that corpus is now a primary input into whether a buyer ever hears your name.

Four ways this goes wrong

The failure modes worth checking for:

Good product, generic language.Customers are genuinely happy but describe you in category-default terms. "Great tool, does what it says." That is a five star rating and a zero contribution to differentiation. The model learns you are an unremarkable member of a category, which is precisely what gets you left out of a shortlist of three.

A corpus skewed towards the unhappy. Satisfied customers do not write. Frustrated ones do, in detail, with specifics, at length. If you have never systematically invited advocacy at the moment value is realised, your public record over-represents your failure modes by an enormous margin, and the model reads that record as representative.

Praise for the wrong attribute. Customers consistently praise something you do not market, usually your support team or a workflow you consider incidental. Two things follow. The model learns an identity your positioning does not claim, and you keep spending against a message the evidence does not support. Sometimes the correct response is to change the positioning rather than the evidence.

Silence at the situation level.Customers describe what you are rather than when they needed you. A model asked "who is best for X" can only surface you if the corpus connects you to X.

Why the last one matters most

That final failure connects to something marketing science established long before any of this. Ehrenberg-Bass work on category entry points holds that brands are retrieved from memory through the situations, needs and moments that trigger a category purchase, not through attribute lists. Mental availability is a function of how many entry points you are linked to and how strongly.

INSEAD's researchers named the AI equivalent Share of Model, and the parallel is not decorative. A model retrieves you against the situation described in the prompt, exactly as a buyer retrieves you against the situation they are in. Share of Model is mental availability with the memory relocated to a machine.

Customer testimony is unusually rich in situational language, because people naturally explain what was going on when they went looking. "We were three weeks from an audit and our data was a mess." "Our previous vendor could not handle multi-currency." That is category entry point language, generated at scale, for free, by the only people who can write it convincingly.

If your customer corpus encodes the situations you want to be retrieved for, you have machine-readable mental availability. If it does not, you are retrievable only for your category noun, competing against everyone else holding the same noun.

What actually changes

None of this argues against brand building or against earned media. Large brands still dominate AI answers, for an unglamorous reason: more customers, saying more things, in more places, for longer. Scale helps enormously. The point is that scale of coverage and scale of corroboration are different assets, and only one of them appears on most media plans.

Mildren proposes a Trust Signal Audit to establish whether the evidence surrounding an organisation reinforces the brand it is trying to build or quietly contradicts it. That is the right instinct, and Gartner's projection that more than 80% of companies will make significant changes to identity, brand and culture by 2028 in response to AI suggests the contradiction will be widespread. The harder question is what you do once the audit tells you the evidence contradicts you.

Five things worth putting on the plan:

Audit the gap before commissioning anything. Find out what models currently say about you, in your market, against your named competitors. Most teams are working from an untested assumption about their AI representation.

Give customers language. The vocabulary in reviews comes from somewhere: your onboarding, your sales conversations, your support replies. If you want situational language in the corpus, use situational language where customers learn to describe you.

Ask at the moment of realised value, not at renewal. The timing of the request determines the specificity of the testimony, and specificity is what makes a review legible to a model.

Treat the unglamorous surfaces as owned. Support documentation, community answers and the way you respond to a public complaint are all corpus. They are read with the same seriousness as your homepage, and often with more trust.

Fix the thing customers are actually complaining about. The least sophisticated recommendation here and the highest leverage one. No content strategy outruns a product problem, because the people who experienced it will keep writing it down.