September 13, 2026 · 6 min
What AI assistants actually search before they answer
When someone asks an assistant which CRM to buy, it usually runs a few web searches before writing anything. Those searches pick the pages it reads, and those pages decide which brands end up in the answer. If you care whether your product appears, the searches are the part that matters — and nobody publishes them.
So we recorded them. We put buying questions across a wide range of commercial categories to ChatGPT, Claude and Gemini and captured every search each one ran first. Four things came out of it, and the first one changes what a tool like this can honestly promise.
0%
query overlap between two identical runs
Ask ChatGPT the same question twice and the two sets of searches share none of their query strings. Gemini repeats itself about 10% of the time, Claude about 24%.
There is no fixed list of searches waiting to be discovered. There is a distribution, and any single observation is one draw from it. After twelve runs of one prompt Gemini had produced 24 distinct queries and was still adding more — so watching an engine once tells you roughly an eighth of what that question actually triggers.
Anyone who offers you the exact queries ChatGPT will run is selling you one sample and calling it the truth.
88%
of advice questions Claude answered without searching
On “how do I choose X on a budget”, Claude searched on roughly one in eight. It answered the rest from memory. On comparisons and recommendations it searched almost every time, so this is not a quiet model in general — it has decided advice does not need a citation.
ChatGPT searched on 100% of the same questions. If your category lives in advice queries, one of the major assistants is not reading anybody’s pages before it answers, and no amount of optimising changes that.
99% → 18%
search rate across two versions of one engine
Two Gemini versions on identical questions: the older searched on 99% of them, the newer on 18%, answering the rest from memory.
That is a large change in how much of the open web reaches an answer, and it arrived with a model version rather than an announcement. Any content strategy resting on “the assistant will look it up” is resting on a number that moves without warning.
1st and 2nd
most-cited domains: YouTube and Reddit
Pooled across every engine, YouTube and Reddit were cited more than the review sites you would expect to own these categories.
The engines also have habits. ChatGPT writes long queries stuffed with brand lists and site: operators, and half of everything it searched was looking for a price. Claude writes one short query, usually the category and the year. Gemini spreads three or four mid-length queries across the question.
What we built from it
The first finding rules out the obvious product. You cannot sell the exact searches, because they do not exist twice. What does hold steady underneath the churn is the ground a question covers — the price comparisons, the alternatives, the category-plus-year searches that recur in different words every time.
That is a prediction problem, and a prediction problem with real observations behind it is a model. Query Fanout 1 takes a question a buyer would ask and returns the searches an assistant is likely to run first, so you can see which pages would have to rank for you to be in that answer.
On categories it never saw in training it reaches 0.563 coverage of what a real engine searched. Two runs of the same real engine agree with each other at 0.75 to 0.91 — so it finds much of what an engine looks for, and misses some, and that ceiling is the honest upper bound for anybody. It will also occasionally name a competitor that does not exist, so check the names.
You can try it on your own category without an account, or call /v1/fanout in four lines.
Buying questions spanning commercial categories and five buying intents, put to each engine and recorded from its responses, September 2026. Engine behaviour moves; these numbers describe what those versions did then.