AI models search for the brands they already remember

AI assistants do not weigh every brand equally when they answer a buyer’s question. A new study from geoSurge, an AI-visibility research firm, found that models search the open web far more often for brands they already recognise from training than for brands they do not, by a factor of more than three to one. For multi-location brands hoping to appear when a customer asks an AI assistant where to go or what to buy, the finding reframes AI visibility as a question of being a known entity, not just a well-optimised page.

What happened

geoSurge measured two things separately, on separate models: what a model already remembers about a set of brands, using the firm’s own memory methodology, and which brands a model then searches for when it answers a real question, measured on Google’s Gemini 3.5 Flash. Because the two sit on different models, the finding is an association between memory and search behaviour, not proof that a single model’s memory drives its own searches. Across nearly 4,000 model responses to 66 United States buyer prompts, firing more than 13,000 fan-out searches and 1,416 brand-level observations, the pattern was consistent. A brand in a model’s top-10 memory was searched at 3.2 times the rate of a brand it did not remember, 55.7 percent of the time against 17.4 percent.

The gap held across all nine industries geoSurge tested, from travel and automotive to finance and retail. Search rates for remembered brands ran between 41 and 82 percent; for brands the model did not remember, between 9 and 23 percent. When a model searched for a specific company by name rather than a generic category, 63 percent of those searches named one of its five most familiar brands.

“A brand in the model’s top-10 memory was searched at 3.2 times the rate of one it did not remember: 55.7% versus 17.4%.”

geoSurge, AI Searches What It Remembers

geoSurge is careful about what the numbers prove. The company describes the result as an association found in exploratory data, not a proven cause, and notes that live search can still surface a lesser-known brand when a category depends on fresh information. The direction is still clear: what a model already knows shapes what it looks for.

Why it matters

Generative engines are becoming a front door to local discovery. When someone asks an assistant for the nearest option or the best provider in a category, the answer depends first on which brands the model even considers. geoSurge’s data suggests that consideration set is skewed toward brands the model has already absorbed, which favours large, established names with a broad, consistent footprint across the web.

That is a structural change from classic search, where a single strong page could rank on its merits. If a model rarely searches for a brand it does not recognise, the older approach of optimising one page for one query does not reach the part of the pipeline that decides who is in the running. Visibility starts earlier, with whether the brand is a coherent, well-represented entity in the data these models are built on.

geoSurge’s broader argument goes a step further, beyond this study. The firm has said that because models can rewrite what they remember when they are updated, a brand well represented in one version can lose ground in the next, with no change to its own site or SEO. That claim sits outside the search-behaviour data reported here, but it follows the same logic: presence in AI answers is not a setting a brand switches on once.

What this means for multi-location brands

For a national retail chain, a dealer group, a pharmacy network, or a fuel or charging operator, the takeaway is not to chase a single AI ranking. It is to make the brand legible at scale, so that across hundreds or thousands of locations the model reads one consistent, authoritative entity rather than a scatter of mismatched listings.

That is a data-governance problem before it is a content problem. Consistent names, addresses, categories, and opening hours across every location, maintained through bulk and API management rather than location by location, are what let a brand read as a single coherent presence in the sources these models draw on. Fragmented or contradictory location data pulls the other way, splitting the brand’s signal across dozens of half-matching records.

The same logic extends to reviews and to the structured content that describes each location. A deep, current review corpus and accurate place data give a model more consistent material to remember. PinMeTo’s Places AI layer is built for this, keeping the entity a brand presents to AI assistants aligned across markets, and our guidance on ranking in AI search results and generative engine optimization sets out how multi-location teams can measure and defend that presence over time.

The practical first step is measurement: track whether your brand appears in AI answers for the queries that matter across your markets, treat a drop after a model update as an event to investigate, and hold location data and reviews to a standard that keeps the brand recognisable at estate scale.

The bottom line

AI assistants increasingly search for what they already know. For multi-location brands, that turns AI visibility into a long game of being a consistent, well-governed entity everywhere at once, rather than a page-by-page optimisation exercise. The brands that stay recognisable to each new model version will be the ones that keep showing up when customers ask.