TL;DR: When you ask an AI assistant a question, it doesn’t answer straight away. First it quietly runs its own set of web searches — usually a dozen or more — and reads what comes back. New research from geoSurge, reported by Search Engine Land on July 30, watched 13,281 of those hidden searches and found the model mostly looks up companies it already knew about before it went online: 55.7% of the brands it recognised got searched for, against 17.4% of the ones it didn’t. Separate work by Seer Interactive found 95% of those hidden searches are phrases nobody types into Google. So the pages that get pulled into answers are often pages nobody optimised for.
First, what actually happens when you ask an AI a question?
Say you ask an assistant which travel insurance covers skiing. It doesn’t just write an answer from memory, and it doesn’t run one search either.
It breaks your question into several smaller ones and searches each of them. “Ski cover travel insurance”, “does travel insurance cover off-piste”, “winter sports policy exclusions”, and so on. It reads the results, then writes one answer out of the lot. The industry name for this is query fan-out — one question fanning out into many searches.
You never see those searches. They don’t appear in your chat, and they don’t appear in your analytics, because the assistant is doing the browsing, not the reader. Researchers get at them by capturing what the model requests while it works.
That hidden step is the part worth caring about. If your page doesn’t come back in any of those searches, nothing you wrote can reach the answer.
What the study found
geoSurge took 66 buying questions from nine industries — travel, cars, finance, business software, education, restaurants, luxury, fitness, fashion — and asked each one sixty times over twelve days, between May 29 and June 9, 2026. That produced just under 4,000 answers and 13,281 fan-out searches. The searches were measured on Gemini 3.5 Flash.
Separately, they measured which brands each model already “remembered” — knew about from its training, before any web search. Then they checked whether memory predicted what got searched for.
It did, strongly.
- Brands the model remembered were searched for in 274 of 492 cases (55.7%). Brands it didn’t remember: 161 of 924 (17.4%) — roughly a threefold gap.
- Memory came in degrees. A model’s top five most familiar brands in a category showed up in searches 67% of the time. The rest of its top ten: 39%. Everyone else: 17%.
- About 31% of all fan-out searches named a specific company rather than searching the category generally. Of those, 63% named one of the model’s top-five remembered brands.
Read that last pair again, because it cuts both ways. Named-brand searches are dominated by a handful of famous names. But they’re only about a third of the searching. The other two thirds are generic — “best travel insurance for skiing”, not “does Acme cover skiing” — and those searches are wide open to whoever has a good page.
How much of this to believe
Two limits, and an agency should say both out loud before quoting the numbers to a client.
geoSurge sells AI-visibility software, so this is vendor research. The methodology is published and the sample is reasonable, but it isn’t independent.
And it’s one model, on one twelve-day window. ChatGPT, Claude and Perplexity all fan out differently, and Seer Interactive’s parallel work on ChatGPT’s fan-out patterns points the same direction without being the same measurement. Treat the direction as well supported and the exact percentages as one snapshot.
The searches nobody optimised for
Seer’s Gemini research adds the detail that changes what you’d actually do about this. Reviewing about 500 prompts and 11,029 captured fan-outs, they found 95% of the fan-out phrases have no search volume at all, and 26% contained a brand name.
No search volume means no keyword tool lists them. Nobody wrote a page targeting them. They’re phrases a machine invented on the spot to fill a gap in what it knew.
Which is a strange kind of opportunity. Two thirds of the hidden searching is generic and unbranded, and almost none of it is contested by anyone doing keyword-led SEO. What decides those is whether a machine, running an odd phrase you’ve never heard of, finds a page of yours that answers it plainly — and can actually read that page. AI crawlers take the HTML as served and don’t run your JavaScript, so a page that assembles itself in the browser comes back empty.
What to do about it
- Stop optimising only for questions you can see. Your keyword list is the visible layer. The fan-out layer sits underneath it, is several times larger, and is mostly phrases with no volume attached.
- Write for the sub-question, not the headline. If your topic is ski insurance, the model is separately searching exclusions, off-piste rules, equipment cover and medical limits. Answer each of those clearly in its own section rather than burying them in one long page.
- Check the page is readable as sent. Server-rendered HTML, no crawl blocks, no critical content injected by a script after load. This is unglamorous and it is the gate everything else sits behind.
- Measure what you’re actually in, not what you rank for. Being remembered by a model isn’t something you can look up in a rank tracker. You find out by asking the engines real questions and recording who gets cited.
- Set expectations with clients honestly. If they’re a small brand in a category dominated by five famous names, the named-brand third of the searching is not where they’ll win this quarter. The generic two thirds is.
The awkward part is the volume. Every client, every page, checked for whether it’s readable as served, then re-checked whenever something on the site changes — and none of it produces a screenshot anyone wants to look at. Which is exactly the work that slips.
That’s what Preferium does without being asked. Its 47 automated checks crawl every page and score the site out of 1000, then the system finds, fixes, deploys and verifies on its own, with a real browser re-opening the live page after each title, meta, H1 and schema deploy. Alongside that it measures citations across all 4 AI engines, so “are we in the consideration set” stops being a guess. More on how the system works.
Key takeaways
- Assistants search before they answer. One user question becomes a dozen or more hidden searches — query fan-out — and if your page isn’t returned by any of them, it can’t be cited.
- Familiarity predicts what gets searched. geoSurge measured 55.7% of remembered brands searched for versus 17.4% of unremembered ones, across 13,281 fan-out queries on Gemini 3.5 Flash (May 29 – June 9, 2026).
- The famous-brand advantage has a ceiling. Only ~31% of fan-out searches named a brand at all. The other ~69% were generic category searches, open to any page that answers well.
- Most of it is invisible to keyword tools. Seer Interactive found 95% of Gemini’s fan-out phrases have zero search volume.
- It’s vendor research on one model. Direction: well supported. Exact figures: one snapshot, quote them with the source attached.
- Technical readability decides whether you’re eligible. The model has to be able to read the page it lands on, in the HTML it was served.
