Study Says GEO Doesn't Work. It's Asking the Wrong Question.

A <a href="https://arxiv.org/abs/2311.09735" target="_blank" rel="noopener noreferrer" class="text-amber-700 underline hover:text-amber-900">meta-analysis of 45 published studies on generative engine optimization</a> landed this week with a conclusion that will make some agency principals nervous: no evaluated technique <em>consistently</em> improves organic AI discoverability across platforms. The widely cited "40% visibility increase" claim — the one that has appeared in countless AEO pitch decks over the past 18 months — is challenged on methodological grounds. The paper argues it measured content prominence <em>after</em> retrieval, not the likelihood of being retrieved in the first place.

A <a href="https://tryorbt.com/publications/the-recommendation-economy" target="_blank" rel="noopener noreferrer" class="text-amber-700 underline hover:text-amber-900">new review of more than 30 published studies on generative engine optimization</a> landed this week with a conclusion that will make some agency principals nervous: the most commonly promoted GEO tactics have weak, null, or negative relationships with AI visibility. The widely cited "40% visibility increase" claim — the one that has appeared in countless AEO pitch decks over the past 18 months — is challenged on methodological grounds. The research argues it measured content prominence <em>after</em> retrieval, not the likelihood of being retrieved in the first place.

The response from much of the AEO industry has been defensive. It should not be. The study is largely correct. The problem is that it is evaluating the wrong thing entirely.

What the Study Actually Found

The 45 studies reviewed share a common design: take a piece of content, apply a GEO intervention — rewrite a paragraph, add schema markup, restructure a heading — then measure whether AI engines cite that content more frequently. The meta-analysis finds these interventions produce inconsistent results. Sometimes a modest improvement, sometimes no change, occasionally a decline.

This is an accurate finding. Anyone who has spent serious time working on AI visibility already knows that paragraph rewrites and keyword insertions do not reliably move citation rates. The problem is not the conclusion. It is the premise. The studies are measuring the wrong lever entirely.

It Is Not a Content Problem. It Is a Claims Problem.

AI citation is not primarily a content formatting problem. It is a <em>claim architecture</em> problem — and beneath that, a <em>domain ownership</em> problem.

When an AI engine decides which brand to cite in response to a purchase-relevant question, it is not scanning for well-structured paragraphs. It is looking for sentences that are <em>specific</em>, <em>defensible</em>, and <em>citable</em> — the kind of claim it can extract, repeat, and present as a recommendation with confidence. A specific percentage backed by a clinical study. A named certification no competitor holds. A concrete outcome tied to a defined time frame. These are the sentences AI engines lift and repeat. Broad quality assertions and mission statements are the sentences they skip.

The meta-analysis reviewed studies that tested whether restructuring existing content around better formatting improves citation rates. What it did not test — because it is nearly impossible to isolate in a controlled academic setting — is whether <em>publishing fundamentally different claims</em> changes citation behaviour. That is the actual intervention that matters.

The Domain Territory Question

Before a brand can publish citable claims, it needs to answer a more fundamental question: <strong>what territory do we want to own in AI search?</strong>

This is not a content strategy question. It is a brand strategy question. And most brands have not answered it.

Consider a well-known supplement brand competing in a crowded consumer health category. Its market position is strong — second in web footprint in its category, significant retail distribution, a parent company with over a century of scientific credibility. By every traditional metric, it should be earning AI citations at a rate commensurate with its market share.

Instead, it earns a fraction of what its smaller, digitally-native competitors earn. The AI gap — the difference between its market position and its AI share of voice — is the largest in its category.

The reason is not technical. Its website is fast, well-structured, and properly marked up. The reason is that its <em>claims</em> are the wrong type. Its content is full of broad, unanchored assertions about product quality and brand heritage — the kind of language that reads well to a human and means nothing to an AI engine looking for something citable.

Meanwhile, its most-cited competitor earns a disproportionate share of all category AI citations from a single sentence. That sentence combines a specific percentage, a named user group, a defined outcome, and a time horizon. It is not better written than anything the first brand publishes. It is <em>structurally different</em> — specific, defensible, and extractable in a way that general brand language never will be.

The first brand's biggest differentiator — its parent company's scientific heritage, its regulatory approval, its clinical backing — exists as a vague brand impression, not as a citable claim. No amount of paragraph rewriting fixes that. The fix requires stepping back from content tactics entirely and asking: what is the specific, defensible territory this brand should occupy in AI-generated answers? What are the two or three sentences that, if published and amplified, would make AI engines reliably associate this brand with that territory?

What We Have Found With Clients

The brands that improve AI share of voice most significantly are not the ones that optimise their existing content. They are the ones that do two things first.

<strong>They audit their claim architecture.</strong> Not their website structure — their <em>claims</em>. What specific, defensible statements does this brand have the right to make? What data exists — clinical, commercial, operational — that has never been expressed in a form an AI engine can extract? What is the brand entitled to say that no competitor can say?

<strong>They identify the domain they want to own.</strong> Not a broad category, but a specific, defensible territory. "The only platform in this category with a published third-party security audit and SOC 2 Type II certification." "The only supplement in this category formulated by a global pharmaceutical company with regulatory approval." That territory becomes the brief for every piece of content that follows.

One client — a B2B software company in a crowded enterprise category — had a strong product, a credible customer base, and a well-maintained website. Its AI share of voice was a fraction of its market share. The gap was not technical. Its content was full of feature descriptions and general value propositions. It had no citable claims.

We identified the domain it could defensibly own, engineered a small number of specific, high-weight sentences around that territory, and published them across the website and in earned media placements. Within 90 days, its AI share of voice had moved significantly. The content did not get better written. The claims got more specific, more defensible, and more aligned with what AI engines actually extract.

What the Study Gets Right — And What It Misses

The meta-analysis is correct that GEO as it is currently practised — paragraph rewrites, schema additions, keyword insertions — does not consistently improve AI visibility. That is a fair and important finding. There is a great deal of tactical noise in this space, and the study is right to challenge the easy claims.

What it misses is the distinction between <em>content optimisation</em> and <em>claim architecture</em>. The studies it reviews are almost entirely testing the former. The latter — identifying the domain a brand wants to own, auditing what citable claims exist and which are missing, then publishing specific, defensible sentences that AI engines can extract — is not something that can be tested in a controlled academic study. It requires brand strategy, competitive analysis, and editorial rigour working together.

Notably, the study's own recommendation is to "optimise content quality and authority." That is exactly what serious AEO looks like. The authors have inadvertently validated the approach while dismissing the tactics. Those are not the same thing.

The Harder Truth

The deepest implication of both the study and the client work is one that most brands are not ready to hear: AI visibility is largely a function of real-world brand authority, not content optimisation.

A brand that has published original research, earned third-party citations from authoritative sources, built a genuine entity presence in knowledge graphs, and accumulated a track record of specific, verifiable claims will earn AI citations. A brand that has done none of those things but has well-structured schema markup will not.

This is not a technical problem. It is a brand-building problem — one that happens to manifest in AI search results. The brands winning AI citations are not winning because their developers added FAQ schema. They are winning because they have spent years publishing specific, sourced, citable content that answers the questions their buyers ask.

The study asked whether GEO techniques work. The more important question is: what kind of brand do you need to be for AI engines to want to cite you? And what specific claims, published in what specific form, will get you there fastest?

Those are the questions worth answering.