September 13, 2026 · 6 min

What AI assistants actually search before they answer

When someone asks an AI assistant which CRM to buy, the assistant usually runs a few web searches before it writes anything. Those searches decide which pages it reads, and the pages it reads decide which brands end up in the answer. If you care whether your product appears, the searches are the part that matters.

Nobody publishes them. The engines show you their sources, sometimes, but not the queries that found those sources. So we watched. We put 200 buying questions across 40 commercial categories to ChatGPT, Claude and Gemini, and recorded every search each one ran before answering: 3,217 searches in total.

Four things surprised us.

They almost never search for the same thing twice

Ask the same engine the same question again and it writes different queries. Across our repeat runs, two ChatGPT answers to an identical question shared none of their query strings. Gemini overlapped about 10% of the time, Claude about 24%.

This is the finding that reshapes the whole problem. There is no fixed list of searches behind a question 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, which means a single look at what an engine searched tells you roughly an eighth of what that question actually triggers.

One of them mostly does not search at all

Claude answered 35 of 40 “how do I choose X on a budget” questions with no search whatsoever. It answered from what it already knew. On comparisons and recommendations it searched almost every time, so this is not a quiet model in general — it is a model that 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 will change that.

Newer does not mean more searching

Within Gemini, we tested two model versions on the same questions. The older one searched on 99% of them. The newer one searched on 18%, answering the other 82% from memory.

That is a large change in how much the open web reaches an answer, and it arrived with a model version rather than a product announcement. Anyone building a content strategy on “the assistant will look it up” should know that the assumption is version-dependent.

Each engine has a habit

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 they read differs too. Pooled across every engine, the most cited domains in our collection were YouTube and Reddit by a wide margin, ahead of the review sites you would expect.

What we did with it

We trained a model on it. Give Query Fanout 1 a question a buyer would ask, and it 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.

The honest limit is set by the first finding above. Because engines do not repeat themselves, no model can tell you the exact strings a given engine will use next time. What it can do is show you the ground the question covers. On held-out categories it reaches 0.563 coverage, where 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. It will also occasionally suggest 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 with four lines of code.

Collection method: 200 prompts built from 40 commercial categories and five buying intents, put to each engine and recorded from the responses. September 2026. Engine behaviour moves; the numbers above describe what these versions did then.