In short

Query fan-out is the method an AI search uses to answer a question. It breaks the question into sub-questions, sends many searches for them at the same time, and assembles an answer from the results. Google describes it this way for its AI Mode: the question is broken down into subtopics, and searches are issued for each of them on the user’s behalf. According to Google, both AI Overviews and AI Mode use this method.

For your site this means: it is no longer found only through the term someone types. It is found through a sub-question the machine formulated itself.

What changes as a result

A classic search for “AI visibility agency” returns a list of results for exactly those words. A question to an AI search reads more like this: our management wants to know whether we show up in ChatGPT, find agencies that can analyse this. That turns into searches on several subtopics, such as what AI visibility is, how it is measured, which providers exist and what they cost.

A page that answers one of these sub-questions precisely gets into the selection, even if it never contains the wording of the original question. A page that only repeats the keyword helps no sub-question.

An example from our own records

Google reports impressions from AI Mode and AI Overviews in Search Console under ordinary web search. We analysed the queries for k-i-soft.ch over 90 days up to 1 October 2026.

Finding Value
Queries in total 149
of which whole sentences with twelve or more words 14
Impressions in total 5,225
of which through these 14 queries 284, so 5.4 percent

The long queries are complete prompts. Some end with an instruction to the AI, such as “suche online nach spezialisierten agenturen im dach-raum, die den status quo unserer ki-sichtbarkeit analysieren können, und nenne mir konkrete anbieter” (search online for specialised agencies in the DACH region that can analyse our current AI visibility, and name concrete providers). The prompts looking for an agency or a provider almost all led to the same page, our page on the AI visibility audit. One also hit our post on AI findability.

The most frequent was the question “warum erwähnt chatgpt unsere firma nicht, wenn man nach anbietern in unserer branche fragt, obwohl wir zu den etablierten unternehmen gehören?” (why does ChatGPT not mention our company when asked for providers in our industry, although we are among the established companies?). It appeared on 31 of 35 days between 27 August and 30 September 2026, with 117 impressions, every time in the same wording. Another question appeared in three versions that differed in only one clause: agencies in the DACH region, in Switzerland, or without a location.

Whether people or tools are behind this, Search Console does not show. Prompts repeated word for word over weeks, and the same question in variants, point to tools that ask the same questions regularly, as services monitoring AI visibility do. Whether Google reports the original question or the sub-searches generated from it, Google does not document. We see both: whole prompts, and the occasional keyword chain such as “evaluate the ai visibility products company agent analytics on ai citation analysis”.

The most common mistake

“We need to optimise for the term our customers type.”

With an AI search the customer types a sentence, and the search runs for something else. Whoever serves only that one keyword drops out of the sub-questions. Whoever answers the sub-questions cleanly, each with a section that makes sense on its own, gets selected for questions they never anticipated.

Where the boundaries run

Search term is what a person types in a classic search. With query fan-out, the machine formulates the searches itself, from the person’s question.

Prompt is the question or instruction to the AI system. The fan-out is the step after it, in which the system searches.

User-triggered fetch is the moment an AI system actually opens a page it found, for example under the identifier ChatGPT-User. The fan-out decides which pages come into question at all. Whether they were read afterwards is in the server logs, see Verified crawl.

AI crawlers fetch pages in advance, for training or an index. The fan-out happens at the moment of the question. More under AI crawler.

How to check it yourself

In Google Search Console under Performance, Queries, you can filter by a regular expression. With the filter “Query”, “Custom (regex)” and the expression ^(\S+\s+){11,}\S+ you see every query with at least twelve words. Set the period to three months and add the Page column.

Look at three things: how many whole sentences there are, which page they hit, and whether the same sentence appears word for word on many days. The page the prompts hit is the one answering a sub-question. That is where every additional section answering a further sub-question precisely pays off.

Sources

Provider pages retrieved on 1 October 2026.