AI models issue fan-out searches, the follow-up queries a model runs while assembling an answer, for brands they already recognize 3.2 times more often than for brands outside their memory. That is the central finding of a study by geoSurge, which sells visibility tooling for GEO, generative engine optimization, the practice of optimizing content for AI-driven answers. Familiar brands appeared in fan-out searches 55.7% of the time, against 17.4% for brands outside a model’s top 10, across nearly 4,000 AI responses to 66 U.S. buyer questions. The gap points to a cold-start problem. If a model is already more likely to go looking for a brand it recognizes, content optimization alone cannot fully close the distance for a brand it does not.
geoSurge ran 66 U.S. buyer prompts, each repeated 60 times, over May 29 to June 9, then analyzed the resulting 3,960 responses. Researchers treated model memory and live search behavior as separate variables: what a model already associates with a category, and what it actually looks up while answering. geoSurge counted 13,281 fan-out searches across those responses and isolated 1,416 brand-level observations within them.
Most fan-out searches skip brand names entirely. Only 31% named a specific company. Among the ones that did, 63% named one of the model’s five most familiar brands in that category. Familiarity does not just tilt which brands get searched. It concentrates the searching around a small, already-known set.
The pattern held across nine industries geoSurge tracked: travel, finance, automotive, business software, food and restaurants, education, fashion, luxury, and fitness and wellness. Familiar-brand search rates ranged from 41% to 82% by category, against 9% to 23% for unfamiliar brands. Some of those industry figures rest on as few as six prompts, according to the study, which makes the per-industry breakdown directional rather than definitive.
geoSurge sells the kind of AI-visibility monitoring its own findings argue brands now need, and that incentive is worth naming plainly. Search Engine Land’s coverage, reported by Editorial Director Danny Goodwin, is staff-written editorial coverage of a vendor study, not a placement commissioned by geoSurge. A vendor publishing its own data and a newsroom independently deciding to cover that data are different things, even when the newsroom’s account faithfully reports the vendor’s numbers.
Memory did not decide every outcome. In one case the study cites, Gemini searched for Lemon Squeezy, an online payment platform, while responding to a question about payment providers. Lemon Squeezy sat outside Gemini’s measured memory for that category at the time. Researchers said live search can still surface brands a model has no strong prior association with, particularly in categories where models lean less on stored memory. The result keeps the study’s core finding from reading as deterministic: memory shifts the odds, not the outcome.
The authors stopped short of claiming causation. They describe only an observed link between what a model already remembers and what it later searches for, not evidence that the memory causes the search pattern. That is the honest limit of a study built on 66 prompts, however many times each was rerun.
For a search team, the cold-start math points to a specific allocation choice. A newer brand competing on broad, established category terms is competing on ground where model memory, not live retrieval, decides who even gets searched. That fight favors incumbents by default. The more productive target is the newer, narrower, faster-moving query, the kind where no brand has built up model memory yet and live retrieval has to do the deciding. Those are close to the conditions under which Gemini found Lemon Squeezy. Brand-awareness work such as PR, category mentions, and presence beyond a brand’s own site now carries a measurable GEO payoff for brands that content optimization alone cannot help.
Search Engine Land reported this geoSurge study on July 30, 2026, in an article by Editorial Director Danny Goodwin.
