geoSurge, a vendor that sells AI-visibility tools, published a study arguing that what one model already recalls about a brand tracks with whether that brand turns up in the fan-out search queries of another. Search Engine Land’s Danny Goodwin covered it on July 30. The headline number: brands inside the recall set appeared in fan-out search queries at 55.7%, versus 17.4% for brands outside it, a 3.2x gap.
The gap: 3.2x more fan-out searches for remembered brands
geoSurge built the comparison from 66 US purchase-related questions, run across roughly 3,960 model responses over 12 days, May 29 to June 9, at 60 iterations per prompt (five per day). Those responses generated 13,281 fan-out queries, the follow-up searches a model issues while working out an answer. From that pool, geoSurge isolated 1,416 brand-level observations: 492 memorized, 924 not. The 55.7% versus 17.4% split comes from that comparison.
What are fan-out searches?
A fan-out search is one of the follow-up queries a model issues on its own, after a user asks a question, to gather current information before it finalizes an answer. In this dataset, 13,281 such queries came out of roughly 3,960 responses. geoSurge’s study counts a brand as “searched” if its name shows up inside one of these queries, not whether a page for that brand later gets cited, ranked, or clicked.
Search rates rise with recall rank
geoSurge also broke the results down by where a brand sat in the top-10 recall measured for that prompt topic. Brands in the top five appeared in fan-out queries at a 67% rate. Brands ranked sixth through tenth still appeared at 39%. Brands outside the top ten, the unmemorized group, appeared at 17%, the same group that showed 17.4% in the aggregate comparison.
| Memory tier | Fan-out search rate |
|---|---|
| Top 5 recalled brands | 67% |
| Rest of top 10 recalled | 39% |
| Unmemorized brands | 17% |
Brand-led queries lean toward the same names
Not every fan-out query names a brand. In this dataset, 31% of all 13,281 fan-out queries were brand-led, meaning they named a specific brand. Of those brand-led queries, 63% named a brand from the memorized top five.
Two different models measured two different things
Memory and search came from separate systems. geoSurge measured memory on a separate model, asking it to recall a top-10 list of brands per prompt topic, via the company’s own recall methodology. Search behavior came from Gemini 3.5 Flash, tested on production fan-out queries: a brand counted as “searched” only if its name appeared inside a fan-out query, not if it was cited, ranked, or clicked.
The cohort is geoSurge’s own open demo dataset: nine organizations, one per industry, covering Travel, Automotive, Finance, Business Software, Education, Food & Restaurants, Luxury, Fitness & Wellness, and Fashion. geoSurge sells AI-visibility software, and this is its own demo set, not an independently sampled panel.
What the authors say the data doesn’t show
geoSurge is direct about its own limits: “This study measures an association in exploratory data, not a proven cause.” Brand familiarity is a plausible confound: a well-known brand is more likely to be both recalled by a model and searched for, independent of any relationship between the two. The size of the 3.2x gap depends on the scope chosen; the direction is more reliable than the exact multiplier. Industry-level breakdowns rest on 6 to 12 prompts per vertical, which the authors call suggestive, not settled. And the dataset is bounded: US-only prompts, a specific 12-day window, two specific models.
Where this sits relative to citation-stage GEO advice
GEO guidance commonly targets a later step: getting a page selected once a model already runs a query for that brand. This dataset describes an earlier one, whether the brand’s name enters the query the model sends out at all. Those are two separate measurements, and geoSurge’s numbers speak only to the earlier one; nothing here says what happens once a brand’s name is inside a fan-out query.
That earlier-step framing runs parallel to a distinction in our look at the self-authorship loop that can homogenize AI search answers, another case where a mechanism inside the model’s own process, not the page itself, does the shaping. Neither study measures clicks, sessions, or revenue. For marketers who want to know whether any AI-driven interest reaches the site, GA4’s AI Assistant default channel is where that traffic would register, a separate, later measurement from a fan-out query count.
Sources: geoSurge, “AI Searches What It Remembers”; Search Engine Land, “AI models favor familiar brands in search: Study”.
