The model cites you everywhere and recommends you nowhere

The week ai visibility got measured properly#

The fox trusts the trail it can smell over the map it is sold. For two years the GEO industry has sold maps. This week, two research teams published trails: real observation counts, stated methods, and in one case the raw data itself.

The first is Kevin Indig's analysis of Semrush AI Visibility Toolkit data, published on Search Engine Land on 5 August. The method is stated in full: 1,094 categories, five prompt variants per category, US ChatGPT data from January to June 2026, yielding 283,215 citation observations and 76,493 named-brand mention observations across 1,458 brand entities.

The second is Moz's fan-out study, published on 6 August: 50,000 fan-out prompts generated through the Gemini API, covering 1,000 subtopics across 20 verticals at 50 prompts per subtopic, with the raw dataset released free. A transparency note our checker insists on: Moz's site blocks automated fetchers, so we verified the figures through a text render of the live page and the syndicated copy, which agree with each other.

Why these two, and not the usual vendor decks#

Both pass the test most ai visibility research fails: a stated method someone else could rerun. Semrush's own write-up of the underlying study describes the five prompt shapes tracked monthly: definition, comparison, alternatives, use case and buying question. Moz went further and published the raw rows, which turns every claim below into something a sceptical analyst can check on a quiet afternoon rather than take on faith.

Between them they answer the two questions that decide whether llm seo budgets are buying anything: what gets a brand quoted, and what gets a brand recommended. Different questions. Different answers. Different invoices.

Cited is not recommended, and the gap is the story#

Here is the finding that should reprice half the GEO market. In categories the model considers close to a brand's actual territory, 74% of appearances are citations, 44% are named mentions, and 34% get both. In distant categories, citations fall to 50% and named mentions to 25%.

Now look at the citation-only share, the brands that get a source link but no name in the answer. It barely moves: 41% in distant categories, 40% in close ones. Being quoted as a source is close to indiscriminate. The model links widely, names narrowly, and the naming is where commercial value lives, because nobody enquires with a footnote.

The depth finding needs stating carefully, because it is subtler than the takes will make it. Brands appearing in one of five prompt variants show a mention-share association of minus 0.051, and only at full five-of-five coverage does the association turn slightly positive. Depth does not catapult a brand into recommendations; it is the point at which depth stops working against you. That is a humbler claim, and it is the one the data supports.

This rhymes with an out-of-window study worth keeping beside it: Semrush's June ghost citations research found 62% of AI citations are ghost citations, a source link with no name in the answer, across 3,981 domain appearances in 14 countries. Two teams, two methods, one conclusion: the citation column in your dashboard is mostly ghosts.

Two query fan out types mint 97% of mentions#

The Moz dataset answers the other half of the question: where, in the machinery of an AI answer, do brand mentions actually get created? When a model receives a prompt, it fans out into multiple sub-queries behind the scenes, and each query fan out class behaves differently.

The concentration is severe. Across 50,000 prompts, brand mentions appeared in just 12.8% of them, and two fan-out types, Entity and Comparison, accounted for 97% of all brand mentions. The mean topical alignment score across the dataset was 0.67. Everything else, the definitional, exploratory and procedural expansions, mints almost nothing.

Sit those two studies side by side and a strategy falls out of the arithmetic. The answer-absolutely-everything content plan, five hundred pages against every conceivable sub-question, is spending most of its budget on fan-out classes that structurally do not mention brands. The work that pays is narrower and duller: being the entity the model resolves, and surviving the comparison it runs.

That means entity clarity, of the kind we set out in SEO vs GEO: why your best pages miss AI citations entirely, and honest comparison surfaces: pages that name rivals, state trade-offs and survive being quoted against you. Uncomfortable to publish, which is exactly why they are scarce, and scarcity is what the comparison fan-out rewards.

It also compounds the familiarity problem we measured in AI brand visibility: 3 brutal truths from 3,960 tests: models reach for brands they already know. Entity fan-outs resolve to established entities. The window for becoming established is the window in which your competitors are still reporting ghost citations to their boards with a straight face.

What an honest ai visibility score tracks#

Measurement is maturing fast, and unusually, some of the good instruments are free. Any ai visibility checker worth its dashboard now has to separate three things the trade conflates: cited, named, and recommended.

Microsoft has quietly become the volume player here. Its Clarity analytics product shipped Topic Insights on 9 July, free to every user, grouping AI citations by topic with competitive positioning and content-gap data, per Microsoft Clarity's announcement. PPC Land counts three AI-visibility releases in 25 days, with branded query segmentation landing on 3 August, and Search Engine Journal covered the August roundup that folded Topic Insights into Microsoft's advertising story.

The paid benchmark sits at the other end: Semrush's AI Visibility Toolkit tracks mentions across ChatGPT, Google AI, Gemini and Perplexity from $99 a month per domain. And the ground truth underneath both: Google's own AI features documentation maintains there are no additional requirements to appear in AI Overviews or AI Mode beyond ordinary indexing and quality. The plumbing is ordinary; the physics of who gets named is what the new datasets describe.

Five moves the data actually supports#

A fox hunts where the quarry is, not where the field looks prettiest. On this week's evidence, the quarry lives in two fan-out classes and five prompt variants, so point the effort there.

First, map your genuinely close categories and cover the five prompt shapes in each: definition, comparison, alternatives, use case and buying question. Full-depth coverage is where the mention association stops being negative, so treat five-of-five as the unit of done, not the page.

Second, build the comparison pages your brand team resists: named rivals, real trade-offs, quotable verdicts. The Comparison fan-out is one of the two places mentions are minted, and it can only mint from pages willing to compare.

Third, tighten entity signals: consistent naming, organisation schema, an about page a machine can parse, third-party corroboration of what you are. The Entity fan-out resolves entities, and an ambiguous one resolves to your better-documented competitor.

Fourth, restructure the monthly report so cited, named and recommended live in separate columns with separate trends. Whoever runs your ai visibility measurement should be made to defend each column's movement separately, because this week's data says they move for different reasons.

Fifth, redirect the long tail budget. The breadth-first plan is not wrong because the pages are bad, it is wrong because most fan-out classes structurally do not name brands, so the marginal page in a procedural class earns nothing a mention can be built on. The measurement shift this demands is the one we built in search traffic fell 34 per cent, the measurement finally caught up.

The quiet quarry in all of this: the raw Moz data is free, and almost nobody will read it. A two-day analysis of which fan-out class your category mints mentions in would outperform most retainers currently sold as llm seo. That analysis, plus the entity and comparison work it points at, is what folkfox SEO and GEO services does, with content marketing building the surfaces and the same discipline running through FinTech marketing, where being named wrongly is a compliance event and being unnamed is a commercial one.

Frequently asked questions#