What is query fan-out?

Query fan-out is when an AI assistant splits one prompt into several search queries. The assistant runs each query, reads the pages it retrieves, and writes its answer from a mix of those results.

How query fan-out works

An assistant that decides a prompt needs outside information does not search for the prompt as typed. It rewrites the prompt into several narrower queries, each aimed at a different angle of the question.

The model reads the prompt and generates a set of search queries.
It runs those queries against a search backend.
It fetches the pages the searches return.
It reads the fetched pages.
It writes an answer grounded in what it read.
It cites a subset of the fetched pages next to the claims they support.

The queries, the fetched pages and the cited pages are three separate lists. Most chat interfaces show only the last one. The queries themselves usually stay hidden inside the product, which is why a person reading an answer has no direct view into what the assistant actually searched for.

How ChatGPT fans out

ChatGPT generates fan-out queries of four kinds: web, business, image and find. Web queries are general searches. Business queries look up local places. Image queries pull photos. Find queries target a specific site. A single turn often mixes several of these kinds at once, so a question about a product might trigger a web query for reviews and a business query for nearby stores in the same turn. Each query returns a group of fetched pages, and the final answer cites a subset of that combined set. Shopping prompts also return product cards with price, rating and merchant, separate from the cited sources.

How Claude fans out

Claude runs web searches and places searches per turn. Citations attach directly to the sentences in the answer text that rely on them. Claude never cites the places it plots on a map: a place can appear on the map without the answer referencing it.

How Grok fans out

Grok runs three kinds of search: web search, page browses and X search. It cites inline by index number, and each number points back into its own list of search results. The index lets you match a citation number to the exact source Grok pulled it from.

Fetched versus cited

A search step almost always returns more pages than the final answer uses. The assistant fetches a wide set to have enough material to draw from, then cites only the pages that back a specific claim. A page can be fetched, read and ignored, and that gap between fetched and cited is where most of a source's visibility gets lost.

The gap matters because being fetched proves an assistant found a page. Being cited proves the page changed what the assistant wrote.

Why query fan-out matters for GEO

GEO (generative engine optimisation) rewards content that ranks for the fan-out queries an assistant generates from a prompt. Those generated queries decide visibility more than the prompt itself. A page can be invisible for the prompt a person typed and still get fetched and cited, because it ranks for one of the four or five queries the assistant spun out from that prompt.

Being fetched is necessary but not sufficient. A page that never gets fetched cannot be cited, but a fetched page still competes with every other fetched page for a place in the final answer. Watching the actual fan-out queries shows which sub-intents an assistant cares about for a topic, which is more specific than guessing from the prompt alone. A brand tracking one broad keyword misses the narrower queries an assistant actually runs, such as a comparison query, a pricing query and a local-availability query all generated from a single user prompt.

How to see the fan-out queries yourself

Install Only Fan Out for Chrome, open its side panel next to a ChatGPT, Claude or Grok tab, and ask a question the way you normally would. The panel lists the queries the assistant ran, the pages it fetched, and the pages it cited, for that turn. It is free, and nothing leaves your browser.