AI search sticks to brands it already knows, study finds

By Zara Kincaid · 2026-08-02

The number that rewrites the search playbook

A brand sitting inside an AI model's top-10 memory gets searched by that model 55.7% of the time, compared with 17.4% for brands the model does not remember, a 3.2x gap measured across 1,416 brand-level observations.^[1] geoSurge published that figure on 28 July 2026, and it is the clearest quantification to date of what the industry calls parametric visibility: where a brand sits in a model's trained associations matters more to AI-driven traffic than anything happening on a live search index.

The study covered 66 US buyer questions across nine industries, generating nearly 4,000 model responses and 13,281 fan-out web queries over 12 days.^[1] Memory was measured using geoSurge's own independent methodology; the search behaviour was observed on Gemini 3.5 Flash production fan-out queries.^[1] geoSurge characterised the results as an association in exploratory data, not a proven causal relationship.

How the machinery actually works

When an AI model assembles an answer, it does not run a single search. It first recalls what it knows about the query's category from its trained weights, producing an internal shortlist of remembered brands, then expands the question into multiple fan-out web queries, reads those results, and cites a subset in its reply. Memory sits upstream of every live fetch that follows.

That architecture has a concrete consequence: of all fan-out queries in the study, 31% named a specific brand and 69% were generic category searches; of the brand-led queries, 63% named one of the model's five most familiar brands.^[1] The model is not neutrally sampling the competitive landscape. It pulls disproportionately from the brands it already knows best, before a single live page is read.

Traditional crawlers work differently. A search engine ingests pages continuously and can rank new content within days. A large language model compresses associations during training, and those associations are fixed until the next update cycle. Optimising a web page at the moment a user types a query does nothing to shift what the model already believes about who the leading players in a category are.

The gap held across every industry tested

One finding that stands out in the geoSurge data is that the memory-to-search advantage produced no exceptions. Across all nine industries, not-remembered search rates ran between 9% and 23%, while remembered rates ran between 41% and 82%, with no industry inverting the pattern.^[1] The lower bound for remembered brands, at 41%, still sits well above the upper bound for unremembered brands at 23%, so the floors and ceilings do not even overlap.

Industry-level universality matters because one common counter-argument is that model memory only applies in high-profile consumer categories where brand awareness was always a factor. The data suggests the dynamic operates just as reliably in categories where buyers would normally conduct detailed comparative research, exactly the settings where a model's recall acting as a gatekeeper before any web lookup is most consequential.

Brands can disappear overnight

Francisco Vigo, CEO and co-founder of geoSurge, said: "These aren't edge cases, they're structural. LLMs don't pull from a live index. They generate answers from compressed memory that shifts with every update. That means a brand can go from 'high visibility' to 'completely gone' overnight, and most organisations won't even know it's happened."^[2]

geoSurge emerged from stealth in December 2025 to monitor where brands sit in model memory and flag when associations degrade between update cycles.^[2] The company's own commercial interest in this framing is worth noting, and the research carries no independent academic or regulator corroboration at this stage. Bushletter could not independently verify the underlying dataset.

What builds model memory, and what does not

Building the kind of associations that survive model compression is a slow accumulation, not a switch. The text corpus a model trains on is weighted toward sources cited repeatedly and across different contexts: press coverage that names a brand in connection with a category, listicle entries that position it among competitors, comparative content that anchors it to established players. None of those signals can be manufactured at query time.

Ethan Imboden, an investor at Tuesday Capital, said: "It's time to stop trying to SEO the LLMs because, well, they aren't search engines."^[3] The window that matters is the period before a model's next training update, not the moment a buyer types a question. Category authority built through genuine editorial coverage accumulates in training data, and a brand that waits until it needs AI-driven traffic to start building that presence is already behind.

geoSurge's own guidance acknowledges the exploratory nature of the data, and the study does not prescribe a specific content formula. What the numbers do establish is a directional signal: the brands capturing the majority of AI-initiated searches are the ones the model already knew, an asymmetry baked into how these systems work rather than a bug a future algorithm update is likely to correct.

Zara Kincaid

Zara Kincaid writes about artificial intelligence and search. Her focus is what happens to businesses when the front page of the internet stops being a list of links and starts being an answer.