Skip to content
Skip to content
BRAND STRATEGY

The model searches for the brands it already knows

Brand building has spent twenty years losing budget battles to performance for want of a mechanism. A study of 3,960 model responses just handed it one, and it sits upstream of every AI brand visibility tactic currently being sold.

Summarise with
Set as preferred source on Google
15 min read
Quick answerAI brand visibility is whether a model searches for you at all. The geoSurge study found models searched familiar brands 55.7 per cent of the time against 17.4 per cent for unfamiliar ones, so retrieval precedes citation.
Section 01

What the study actually measured

ai brand visibility

A fox does not dig every den it passes. It returns to the burrows it remembers, because memory is cheaper than searching and a hungry hunter cannot forage blind. Large language models keep the same habit, and that habit is the load-bearing wall beneath AI brand visibility.

The evidence surfaced on 30 July 2026, when Search Engine Land reported a study from geoSurge titled Report: AI Searches What It Remembers. No publication date appears on the study's own page, so the trade coverage carries the dateline of record. The method sits in the open: 66 United States buyer questions across nine industries, each run 60 times between 29 May and 9 June 2026, yielding 3,960 responses, 13,281 fan-out queries and 1,416 brand-level observations.

One number carries the argument. Brands already sitting in a model's memory were searched for at 3.2 times the rate of brands that were not, 55.7 per cent against 17.4 per cent, per geoSurge. Being unknown is not neutral ground, it is exclusion at the fetch step, long before a quotable sentence is ever read.

The geoSurge study in four figures

Model responses

3960

66 US prompts, run 60 times each.

Fan-out searches

13281

Issued before an answer was written.

Familiar brands searched

0.0%

274 of 492 cases, brands inside the remembered top ten.

Unfamiliar brands searched

0.0%

161 of 924 cases, brands outside it. Same prompts.

Every figure here is reported by geoSurge from 3,960 responses to 66 United States prompts run between 29 May and 9 June 2026. None is a folkfox measurement.

Memory is a dial, not a switch

The split is not binary. geoSurge reports a gradient: top-five brands were searched for 67 per cent of the time, the rest of the top ten 39 per cent, brands outside memory 17 per cent. Familiarity is a ladder, and climbing from nowhere to tenth is a shorter stretch than tenth to third.

Search rate by memory tier
Bar chart of brand search rates by memory tier: top five 67 per cent, rest of top ten 39 per cent, not remembered 17 per centTop 5 recalled: 67Rest of top 10: 39Not recalled: 1767%50.2%33.5%16.8%0%Top 5 recalledRest of top 10Not recalled
Search rate climbs with recall strength. Reported by geoSurge from 3,960 responses, not measured by folkfox.
Section 02

Retrieval happens before citation

Answer engines do not run one search. Google's Search Central guidance on AI features describes a query fan-out technique, issuing multiple related searches across subtopics and data sources before an answer is assembled. Those fan-out searches are where brand names live or die.

OpenAI's web search documentation makes the same architecture explicit from the other side: the sources field lists every URL the model consulted, and the number of sources is often greater than the number of citations. Two gates, then. The first decides whether your pages are fetched, the second whether a fetched page is quoted.

Most generative engine optimisation decks sold in the last eighteen months address the second gate: schema, answer-first paragraphs, quotable sentences, passage anchors. All of it is sound. None of it operates if the first gate stays shut, and AI brand visibility is the name for that gate.

A page that is never fetched cannot be cited, however beautifully it is marked up.
folkfox, on why AI brand visibility precedes on-page GEO

This is not a reason to stop on-page work, it is a reason to sequence it. Get retrieved, then get cited. Google's helpful content guidance asks whether a reader would see the site as widely recognised on its topic, which is a retrieval question wearing an editorial coat.

The order of operations most GEO decks get wrong

The gates AI brand visibility has to clear, in order.
GateWhat decides itOwnerEvidence it moved
1. RetrievalThe model names you in a fan-out searchBrand and PRBranded grounding queries appear
2. CitationA fetched page answers the query cleanlyContent and SEOCited pages and share of authority rise
3. AttributionThe answer names you rather than paraphrasingEntity and engineeringOne Organization entity resolves

Teams treating gate two as the whole task get very good at writing sentences nobody fetches. Work that lands begins with brand strategy, then moves to SEO and GEO services.

Section 03

The gap is wider in some categories than others

Not every hedgerow hides the same quarry. Across nine industries (travel, automotive, finance, business software, education, food and restaurants, luxury, fitness and wellness, fashion), remembered brands were searched for between 41 and 82 per cent of the time, unremembered brands between 9 and 23 per cent, per geoSurge and Search Engine Land.

The reported range, end to end
Remembered, widest category
82%
Remembered, narrowest category
41%
Unremembered, widest category
23%
Unremembered, narrowest category
9%
Range endpoints reported by geoSurge across nine industries, not the score of any single industry and not a folkfox measurement.

Resist reading a league table into that. The report calls per-industry figures suggestive rather than settled, because each industry rests on only six to twelve prompts, and names the cohort-wide numbers as the firmer ones. What survives is the shape: the two ranges never overlap, so in every category tested the remembered brand was searched for more often than the unremembered one.

In regulated categories the stakes climb. A fintech, a clinic or an operator whose name never surfaces in a fan-out search is not merely losing traffic, it is losing the chance to correct what the model says about its licence.

That last bullet is not a hedge, it is the point. One study, honestly reported, beats a confident claim with no method attached.

Section 04

Building AI brand visibility without a television budget

Here is where most coverage of this study goes wrong. The conclusion is not be a big brand, since that is advice only a big brand can take. The useful reading is narrower and kinder: AI brand visibility comes from category association rather than category dominance, and association is buyable in ways dominance is not.

The Ehrenberg-Bass Institute defines mental availability as buyers easily thinking of a brand when buying and recognising it quickly, and holds that connecting a brand to more category entry points raises the chance it gets bought. Swap buyer for model and the mechanism rhymes.

What builds machine-side familiarityThe four off-site signals folkfox prioritises, illustrative rather than measured. Drawn from client work, not from the geoSurge dataset, which ranks no signals. None needs a media budget. All four need consistency.The entity the model remembersThird-party listsWikidata entityOrganization schemaCategory coverageThird-party listWikidata entityOrganization schCategory coverag
The four off-site signals folkfox prioritises, illustrative rather than measured. Drawn from client work, not from the geoSurge dataset, which ranks no signals. None needs a media budget. All four need consistency.

Entity consistency is the boring lever that moves

A model cannot remember a brand it cannot resolve to one entity. That is what schema.org's Organization type exists for, and in particular sameAs, defined there as the URL of a reference page that unambiguously indicates the item's identity. Google's own guidance is blunter about the block as a whole: adding it to your home page helps Google understand your administrative details and disambiguate you in search results. Ship it as JSON-LD, the format Google's structured data introduction recommends as easiest to maintain at scale, then keep name, URL and sameAs identical everywhere. Divergent entity data is the commonest own goal in AI brand visibility work.

Notability is a lever you can genuinely pull

Wikidata is testable and unglamorous, which makes it perfect. Wikidata's notability policy admits an item with a valid sitelink to a Wikimedia project, or one referring to a clearly identifiable conceptual or material entity describable with serious and publicly available references, or one filling a structural need. That second door stands open to many small companies.

Wikipedia is stricter. Its notability guideline for organisations demands significant coverage in multiple reliable secondary sources independent of the subject, and rules out press releases, paid or sponsored articles, and routine announcements of funding, hires and launches. Put plainly: the coverage you bought does not count.

The off-site plan, in order
Fix the entity

One legal name, one canonical URL, one Organization block with a stable sameAs set.

Pick three category entry points

Not keywords. Three situations a buyer is genuinely in when your category matters.

Get named where pages rank

Comparison pages, industry lists, directories, speaker pages. Fan-out searches return those.

Publish where the model forages

Trade press, research write-ups, credible community answers. Owned content is not enough.

Re-measure monthly

Branded against non-branded grounding queries, not rankings. This signal moves in months.

Run it in that sequence and AI brand visibility becomes a workstream that content marketing, paid social and PR can be pointed at, the same discipline we apply for FinTech marketing clients in crowded categories on finite budgets.

Section 05

Measuring AI brand visibility for free

Until three days ago the honest answer to how do I track this was expensive. On 3 August 2026 Search Engine Land reported that Microsoft Clarity had added branded and non-branded segmentation across its AI reports, with the Share of Authority card now breaking results out by query type.

Clarity's own write-up on AI citations defines the metrics that matter. Grounding queries surface the key phrases the system used when retrieving cited content. Cited pages show which of your URLs get selected and how often. Share of Authority shows how citations divide between your domain and other sources across overlapping grounding queries. Two caveats come from Clarity itself, and both belong in the boardroom: the feature is badged early access, and the figures are built for reading trends rather than precise accounting.

That split is the geoSurge finding made measurable on your own site, and Microsoft Clarity bills itself as free forever with no traffic limits, so the cost of looking is nil. Citations arriving almost entirely on branded grounding queries mean the model knows you but the category does not. Citations on non-branded queries mean category association, the harder and more valuable half of AI brand visibility.

What the fan-out searches asked for
What the fan-out searches asked forDonut chart splitting fan-out queries between those naming a specific brand and those naming only a topicNamed a specific brand: 31%Topic only: 69%31%
Named a specific brand 31%Topic only 69%
Of 13,281 fan-out searches, 31 per cent named a specific brand and the rest named only a topic, per Search Engine Land reporting the geoSurge study.

Pair it with the older instrument. Search Console's performance report still defines impressions as how many times your site appeared in Search results. Impressions holding while clicks fall is a summarisation problem. Impressions falling is a retrieval problem, and retrieval problems are AI brand visibility problems.

The one number to take to the board

Pick share of non-branded grounding queries where your domain is cited, and defend it for two quarters. It costs nothing to collect and it survives the next model release, because it measures being remembered rather than being ranked.

Quiet disciplines are how smaller brands outfox larger ones. Our work on AI citation reporting and our piece on why strong pages still miss AI citations start in the same place: measure the mechanism, never the mood. The brands that win the next two years are the ones a model calls to mind before it calls the web.

Questions

Frequently asked questions

What does AI brand visibility actually mean?

It is whether an AI system searches for your brand at all while researching an answer. Models issue background fan-out searches before responding, and geoSurge found they name familiar brands far more often. AI brand visibility is your presence in those searches, upstream of any citation.

Does this study prove brand building causes AI citations?

No, and the report says so itself. The study describes an association in exploratory data rather than proven cause, and names brand fame as the confound, since well-known brands are both likelier to be remembered and likelier to be searched. Twelve days, United States prompts, memory and searching measured on separate models. The direction is credible, the magnitude is not settled.

Is on-page GEO work pointless for an unknown brand?

No, it is second in the sequence. On-page work decides whether a fetched page gets quoted, and it cannot help a page that was never fetched. Do the entity and off-site work first, then the answer-first writing, and both start earning.

How do I measure AI brand visibility without buying a tool?

Microsoft Clarity now splits AI citation data into branded and non-branded queries, with a Share of Authority breakdown, in an early-access feature on a product billed as free forever. Pair it with Search Console impressions, then track the non-branded share monthly rather than chasing weekly movement.

Rand Fishkin
I'm worried that citation analysis research is misleading us... badly.
2026View on X
Will a Wikipedia page fix this?

Sometimes, and it is harder than it looks. Wikipedia requires significant independent coverage and excludes press releases, sponsored posts and routine announcements. Wikidata is more achievable, since an item qualifies if it is a clearly identifiable entity described with serious, publicly available references.

Keep reading

Read more on this topic

Want to be the brand the model already knows?

folkfox builds AI brand visibility for smaller brands in crowded categories: entity work, off-site presence, and reporting that measures retrieval rather than mood.

The den

Where to?

Pricing

Choose a section. Enter opens it, Escape continues reading.

Cookie preferences

folkfox uses data the way we use strategy: only when it earns its place.