AI models search brands they already know 3.2× more often

News: On 30 July 2026, the analytics firm geoSurge published a longitudinal study showing that a generative model issues 3.2 times more searches for a brand it already holds in memory than for one it does not: 55.7% of cases versus 17.4%. The study covers 3,960 responses and 13,281 queries generated by the model itself (geoSurge, 30 July 2026).

Direct answer: when an AI engine answers a question, it does not read the web and then choose. It first decides what to search for, and that decision draws on what it retained from training. The geoSurge study quantifies that bias for the first time by separating the two measurements: what the model holds in memory, and what it actually searches for. The result is a ratio of 3.2. In other words, most of the competitive filtering has already happened before a single one of your pages is read. A flawless page that no query surfaces has no chance of being cited, not because it was judged inadequate, but because it was never judged at all.

What the study actually measures

The protocol published by geoSurge runs from 29 May to 9 June 2026. Sixty-six US buyer questions, spread across nine sectors (travel, automotive, finance, business software, education, food and restaurants, luxury, fitness and wellness, fashion) were each asked sixty times, at five iterations per day. That yields roughly 3,960 responses, which themselves triggered 13,281 search queries, close to 3.4 searches per response. Search behaviour was observed on production fan-out queries from Gemini 3.5 Flash; memory was measured separately, using geoSurge's own methodology. In total, 1,416 brand observations, of which 492 were remembered and 924 were not.

The central result fits in one sentence from the authors: "A brand the model remembers from training is searched about 3× as often as one it doesn't." The gradient is regular, which makes the signal more credible than a simple binary gap would be.

One detail narrows the funnel further: among queries that do name a brand, 63% point to one of the five brands the model remembers best. Recognition does not merely open the door, it captures most of the traffic through it.

The point nobody picked up: 69% of queries name no one

The pessimistic reading of this study is obvious: the big get bigger, a small business has no chance. It is incomplete, and the number that corrects it sits inside the study itself. Across all measured fan-out queries, only 31% contain a brand name. The remaining 69% are generic queries: a problem to solve, a comparison criterion, a usage constraint.

Those two populations of queries do not follow the same rules. On the 31% that name a brand, the outcome is largely settled in advance and settled on recognition. On the 69% generic ones, the model is not looking for a brand: it is looking for an answer. The decision reverts to relevance, where an unknown company can appear in the result set and get cited. This is the mechanism we documented when explaining why pages ranked outside Google's top 10 still get cited by AI: the engine does not reward rank, it rewards fit to a specific query.

The operational consequence is sharp. If you are not an established brand, the content that pays is not the content about you: it is the content that answers precisely the generic question the model will ask on your behalf. The work is to anticipate those intermediate queries, not to optimise an "about us" page.

The other half of the problem: the effect is real, but invisible in your analytics

A second publication, independent of geoSurge, completes the picture on the outcome side. In "From Prompt to Purchase", filed on arXiv on 9 June 2026, Michael Iannelli and Alan Ai linked ChatGPT, Claude and Gemini conversations to the same users' actual clickstream data. When an assistant recommends a brand to a user with no prior engagement, the authors measure +4.3 percentage points in Google searches for that brand (95% confidence interval: 3.1 to 5.5), +2.4 points in visits to its website (1.4 to 3.5) and +1.0 point in retailer-page visits (0.3 to 1.7).

Two lessons. First, an AI recommendation is not a vanity signal: it shifts measurable behaviour. Second, and more awkwardly, it shifts it towards Google. The user reads the recommendation, then goes and searches for the brand by name. The visit therefore lands in your statistics labelled "brand search", and SEO takes the credit. The authors put it plainly: standard referrer-based and last-click measurement miss this upstream exposure. Your dashboard credits the wrong channel for an effect produced by another, exactly the same blind spot described in our analysis of share of voice in AI Mode.

One useful nuance from the arXiv study: incidental mentions, where the name appears without being recommended ("your Netflix download"), produce markedly weaker effects: +1.8, +1.1 and +0.3 points. Being named is not enough; it is the recommendation that moves behaviour.

What to do now

1. Map your generic queries, not your brand keywords. List the ten ways a buyer would phrase the problem without knowing any supplier, and check whether you have a page that answers each one head-on. That is the territory of the 69%.
2. Treat model memory as a PR objective, not an SEO one. geoSurge concludes that durable category authority, built through analyst and press coverage and partnership signals, feeds memory at training time and is "the more durable position". A mention on a recognised third-party site weighs more here than one more page on your own domain.
3. Stop judging AI citations by referral traffic. Track branded search volume instead: that is where the effect lands. A rise in brand queries with no paid campaign running is the most likely symptom of exposure inside AI engines.
4. Make your pages readable out of context. A cited excerpt travels without the rest of the page. The traceability criteria we detailed around the five trust signals in OKF v0.2, namely a named source, its date, then an identifiable author, apply directly.
5. Stop producing commodity content. On a generic query, a page that repeats what ten others already say gives the model no reason to retain it: the mechanism we described in our analysis of commodity content that never gets cited.

Our take

This study moves the GEO problem one step upstream, and that is uncomfortable. We spent two years optimising the page the model would read; geoSurge suggests the battle is partly decided at the previous step, when the model decides whether to look for you at all. But reading this as a verdict against small brands would be a misreading: the same dataset says seven queries out of ten name nobody. Recognition locks up branded queries; it does not lock up questions. For a small business the practical conclusion is not to give up, it is to stop fighting on the named-brand ground where the gap is 3.2 to 1.

What this article does not cover

The geoSurge authors are explicit about the study's limits, and it would be dishonest to omit them. The result establishes an association between memory and search behaviour, not causation: a brand's overall prominence may drive both measurements at once. Search behaviour was observed on a single model, Gemini 3.5 Flash, on its fan-out queries; nothing guarantees ChatGPT or Perplexity behave identically. The observation window does not exceed twelve days. The prompts are by construction exclusively American, which leaves open how this transfers to European markets, where a model's training memory of local players is probably thinner. Finally, the per-sector sample is small, six to twelve prompts, which makes the reported sector spread (41% to 82% for familiar brands against 9% to 23% for the rest) indicative rather than solid. geoSurge is also a vendor of AI visibility tools, and therefore an interested party in the subject it measures; we were not able to replicate the study independently. We will update this article if a multi-model replication is published.

Frequently asked questions

Alexis Dollé

CEO & Founder

Growth and SEO content strategist, I founded Cicéro to help businesses build lasting organic visibility, on Google and in AI-generated answers alike. Every piece of content we produce is designed to convert, not just to exist.

LinkedIn