The More You Optimise for AI Search, the Less AI Recommends You

The AEO playbook, as currently practiced, has a few recognisable moves. Add schema markup. Write title tags that match fan-out queries. Structure pages with FAQ sections that answer common questions. Use natural-language URLs. Publish opinion content. Get cited in third-party publications. Earn UGC mentions.

Each of these moves, taken individually, has evidence behind it. The problem is what happens when an entire category of brands runs them simultaneously.

Consider what AI sees. When a buyer asks “who are the leading providers in X,” AI retrieves candidate pages, evaluates them, and constructs a recommendation. The evaluation is essentially a similarity calculation: which pages are most relevant to the fan-out queries the model generates from the buyer’s prompt, and which ones look most credible against the patterns the model has learned?

If five competing brands have run identical AEO practices, AI now sees five pages that:

• Have nearly identical schema markup.
• Have nearly identical title structures aligned to the same fan-out queries.
• Have nearly identical FAQ sections answering the same questions in the same order.
• Use the same category-level phrasings the AEO guides recommended.
• Were cited in similar third-party publications the AEO playbook pointed them at.
• Have similar UGC mention patterns from the same kind of seeding effort.

These pages are not distinguishable from each other on the dimensions AI evaluates. AI does what any well-engineered system does when given indistinguishable inputs. It picks somewhat arbitrarily, sometimes alphabetically, sometimes based on small differences in metadata that nobody intended to optimise for. Sometimes it picks whichever brand has slightly more earned-citation volume, which is the part of the system that is not optimisable by AEO practice anyway.

The brand that “won” AEO is not the brand recommended. The brand recommended is the one AI could distinguish from the others. Sometimes that is the brand with the strongest opinion. Sometimes that is the brand with the most authentic UGC mentions. Sometimes that is the brand that did not over-optimise.

This is the AEO paradox: each brand running the playbook is making themselves harder to distinguish from every other brand running the playbook. The work is locally rational. The outcome is collectively self-defeating.

AI cannot recommend a brand it cannot distinguish.

There is an analogy from machine-learning research called model collapse. When generative AI models are trained on output from other AI models, the diversity of the training distribution shrinks with each generation, and eventually the model collapses to producing a small set of homogeneous outputs. The dynamic, transposed to marketing, is what AEO is doing at scale. Every brand learning from every AEO guide is contributing to a homogenisation of the input AI sees in their category. Eventually, all the brands look like each other to the engine. The recommendation defaults to whoever escaped the homogeneity.

Five separate studies, each with different methodology and different commercial incentives, converge on the same finding. AI engines do not reward optimised content. They reward content that is structurally distinct from the optimised baseline.

Finding one: 95.1% of AI citations are earned, not owned. Per Profound’s analysis of 250 million AI responses, 95.1% of citations come from sources the brand does not directly control. Owned content (the brand’s own website) accounts for 3.3%. Operated (sites the brand influences but does not own outright) accounts for 1.6%. If you spend 100% of your AEO effort on owned-content optimisation, you are addressing 3.3% of the citation surface. The other 96.7% is shaped by what gets said about you in places you cannot directly optimise: Reddit threads, YouTube transcripts, specialist publications, analyst notes, peer commentary.

Finding two: opinion content beats comparison content. Same Profound study, 8,500 top citations analysed by category:

• Blogs and opinion content: 34.2%
• Comparative pieces and listicles: 27.3%
• Documentation and wikis: 15.8%
• Commercial and store pages: 14.8%
• Community and forum content: 5.0%
• Homepages: 2.2%

AI cites brands that take a position more than brands that hedge. The “10 best providers for X” comparison post that AEO guides recommend writing produces fewer citations than a single opinionated piece that argues for one position. Taking a stance is a citation signal. Hedging is an erasure signal.

Finding three: Reddit dominates the citation set. Reddit is the single most-cited source in AI Search, by share, on every major engine. Per Profound’s 90-day window: Perplexity 5.90% (#1 source), ChatGPT 2.87% (#2 overall), Google AI Overviews 2.52%, Google AI Mode 2.08%. UGC as an information class: ChatGPT 17.4% of citations, Perplexity 15.8%, Google AIO 12.3%.

Why does AI weight Reddit so heavily? Because Reddit content is, by construction, unoptimised. Reddit posters are not writing for AI engines. They are writing for other humans. The voice is authentic, the opinions are real, the language is human. AI editorial teams have deliberately weighted this kind of content because it carries trust signals that proprietary algorithms cannot generate. As Profound’s Josh Blyskal put it in the source material: Reddit is not an accident. It is an editorial choice by AI builders.

The implication for brands: the closer your content sounds to “a real person with a real opinion,” and the further it sounds from “a brand running a content playbook,” the better AI’s chance of citing you.

Finding four: schema markup produces no causal effect on AI citations. Per Ahrefs’s 1,885-page causal test (matched difference-in-differences design, four robustness tests, May 2026): adding JSON-LD schema to a page produces no meaningful citation uplift on Google AIO, Google AI Mode, or ChatGPT. The widely-pitched “add schema for AI visibility” tactic does not survive causal testing on pages already in the consideration set.

A companion experiment from searchVIU, cited in the Ahrefs study, found that none of ChatGPT, Claude, Perplexity, Gemini, or Google AI Mode used schema markup during real-time page fetch. All five systems extracted only visible HTML content. JSON-LD, hidden Microdata, and hidden RDFa were all ignored.

This is the cleanest evidence we have that the AEO playbook’s most common technical move does nothing. Ahrefs sells SEO tools, including features related to schema markup. They published a result that undermines that pitch. The vendor publishing a finding against their own commercial story is the strongest credibility signal a research function can produce.

Finding five: relevance to the buyer’s actual question beats relevance to category keywords. Per Ahrefs’s 1.4-million-prompt analysis of ChatGPT 5.2: once a page is in the consideration set, the strongest measurable predictor of citation is the cosine similarity between the page title and the fan-out queries ChatGPT internally generates from the buyer’s prompt. Cited URLs score 0.656 against fan-out queries; non-cited URLs score significantly lower.

This sounds like an argument for optimisation. Read it again. The optimisation target is not “match the keywords.” The optimisation target is “be specifically relevant to the question the buyer is actually asking.” That is a distinctiveness move, not a conformity move. The brand that wins is the one whose title is unambiguously about the buyer’s actual question, not the one that hedged with category-level language to capture the broadest possible keyword set.

The convergent finding. Five studies, different methodologies, different commercial incentives, all pointing at the same conclusion: AI rewards content that is distinct. Distinct in source (earned over owned). Distinct in voice (opinion over comparison). Distinct in tone (authentic over branded). Distinct in tactical signal (the absence of optimised conformity beats the presence of optimised conformity). Distinct in relevance (specifically answering the buyer’s actual question beats generically aligning with category keywords).

The AEO playbook produces the opposite of every one of these. It pushes brands toward owned content over earned. Toward branded voice over opinionated voice. Toward optimised tone over authentic tone. Toward technical conformity over content distinctiveness. Toward category keywords over buyer-specific questions.

The playbook is not just unhelpful at scale. It is structurally backwards.

What replaces the AEO playbook is not “a better playbook.” The category does not need a sharper set of tactics. It needs a different orientation.

The right orientation is what we call a distinctiveness moat: the deliberate set of choices that make your brand recognisably different from every other brand in your category, so that AI can distinguish you when it constructs a recommendation.

A distinctiveness moat is built from five kinds of choice, each tied to one of the findings above:

One. Earned source distinctiveness. The recommendations AI gives are built primarily from sources the brand does not control. Investing in being the brand most-cited by the third-party sources AI already trusts in your category is the highest leverage move available. This is not generic PR. It is specific placement in the specific publications, threads, and analyst surfaces that AI already weights. The distinctiveness move: identify the five sources AI cites most in your category, then be the brand that has the strongest editorial relationship with each of them.

Two. Position over coverage. Publish content that takes a position that almost no other brand in your category will take. The opinion content category is 7 percentage points above comparison content in the citation share. The position you take is what makes you distinguishable in the AI’s prose. If your strongest position could be made by any of your competitors with a slightly different anonymisation pass, your position is not strong enough.

Three. Authentic voice in the surfaces AI weights. Reddit’s citation dominance is not a Reddit story. It is a voice story. AI weights surfaces where authentic human voice produces authentic human language. Building a presence in those surfaces (not optimising your owned content to sound like them, which is a different and weaker move) is a distinctiveness signal. The right work is specific participation in specific communities where your category is discussed, written in the voice of the community and not the voice of your brand guidelines.

Four. Distinct architectural choices, not optimised ones. The schema finding is one example of a broader pattern. The AEO playbook’s technical moves (schema, title-tag optimisation, FAQ-heading structure) produce no causal effect when adopted at scale. What does produce effect is having content the engine can extract cleanly because the visible HTML is structured the way humans read it, not the way crawlers parse it. The distinctiveness move: write for the reader, not for the parser. Counter-intuitively, this is more, not less, optimised for AI.

Five. Specific relevance to the buyer’s real questions, not the category’s search volume. The fan-out finding from Ahrefs is the most actionable single insight in the citation literature. The brand cited is the brand whose titles and headings specifically address the sub-questions the buyer is actually asking, not the ones the keyword tools say have the highest volume. The distinctiveness move: do the research to know what your buyer actually asks AI, then write titles that unambiguously answer those questions in language no competitor would have chosen by accident.

Each of these moves is, in the language of strategy, costly to imitate. A competitor cannot copy your earned-source relationships in a week. They cannot copy your opinion-taking content without sounding like they copied it. They cannot copy your community voice without doing the participation work themselves. They cannot copy your buyer-specific question architecture without first doing the buyer research. These are moats. The AEO playbook’s tactics are not. Adding schema, optimising titles, and earning UGC mentions in the same way everyone else does are tactics any competitor can replicate in a sprint.

The discipline that operationalises the distinctiveness moat is Recommendation Design. It is to AI what SEO was to search: the deliberate practice of being chosen, not just seen. The loop runs continuously across the five stages of the AI-mediated buyer decision (Consideration, Evaluation, Decision, Retention, Advocacy). Each stage is its own observation, decoding, and seeding cycle. The full mechanics are in the flagship piece on Recommendation Design; this section reframes the discipline through the distinctiveness lens.

What changes when the loop is run with the distinctiveness moat in mind is not the structure of the work. It is the question the work asks. The standard reading of Recommendation Design asks: what is AI saying about us, and what should we publish to change it? The distinctiveness reading asks: what is AI saying about every brand in our category, and what could only we say that would change which one of us AI recommends?

Observe, with the homogeneity test. Run the prompts your buyers actually use, across the AI tools they actually use, with one specific additional move: also run the same prompts against your three closest competitors’ brand names. Capture how AI describes you, how it describes them, and how similar the descriptions are. The observation that matters is not whether your brand appears. It is whether your description is distinguishable from your competitors’ descriptions when you read them next to each other.

Decode, with the distinctiveness diagnostic. For every prompt where you are mentioned, identify the specific words and phrases AI uses about you. Then ask: how many of these words and phrases would also be true of our top three competitors? If the answer is “most of them,” AI does not yet have a distinctive read on your brand. If the answer is “some of them,” AI has the beginnings of a distinctive read but the signal is weak. If the answer is “almost none of them,” AI has a distinctive read and the seeding work has been effective.

Seed, with the no-one-else-can-credibly-say-this rule. When you produce content against a decoded gap, the test is not “does this fill the gap.” The test is “can any of our competitors credibly publish the same thing?” If yes, the content is not distinctive. Publish anyway, but understand it will produce a partial result. The content that moves the recommendation is content that only your brand can credibly produce. That credibility comes from named experience, named customer outcomes, named opinions, or named methodologies that your brand alone can point to.

The discipline is the same. The question is sharper. Recommendation Design without the distinctiveness lens produces brands that are visible. Recommendation Design with the distinctiveness lens produces brands that are chosen.

The first move is a diagnostic any marketing team can run in an afternoon. It produces a specific number that tells you how far you are from a distinctiveness moat, and it tells you which competitors AI currently sees as interchangeable with you.

Step one. Pick five prompts your buyers actually ask AI in your category. If you do not yet have a real list, ask three recent customers, “If you were looking for a brand in our category today, which questions would you ask AI?” Record the questions verbatim.

Step two. Run each prompt across ChatGPT, Claude, Gemini, and Perplexity. Capture every brand named in each recommendation. Capture the specific language AI uses about each brand.

Step three. Build a 2x2 matrix. Rows: the brands named. Columns: the descriptive phrases AI used. Fill in which brands got which descriptions.

Step four. For each row (each brand, including yours), count how many descriptive phrases your brand shares with every other brand. The brand sharing the most phrases is the brand AI sees as most interchangeable with the others. The brand sharing the fewest phrases is the brand AI sees as most distinct.

Step five. Calculate your distinctiveness gap. It is the number of descriptive phrases that AI applies to you that AI also applies to at least one other brand. The lower the number, the better your distinctiveness moat. Anything over 60% shared phrases means your brand is in the homogeneity zone where AEO practice has erased your distinguishability. Anything under 30% means AI has a distinctive read on your brand that no current competitor is matching.

Step six. Identify one descriptive phrase that AI applies to a competitor that you wish AI applied to you. Then identify one descriptive phrase that AI applies to you that you wish it applied uniquely. Both of those phrases are seeding targets. The first is a credibility gap to close. The second is a credibility gap to widen, by publishing content that makes the phrase impossible to apply to anyone else.

The diagnostic produces both a measurement (your distinctiveness gap) and an action plan (the two phrases to seed against). Run it once. Seed against the two phrases for a quarter. Then run it again. The change in the gap is your evidence that the moat is being built.