Definition

What Is Query Fan-Out in AI Search?

By Rahil Jain, founder of Chalk LabsUpdated October 20, 2026How we research

The short answer

Query fan-out is when an AI search tool splits one question into several smaller web searches, runs them, and combines the results into one answer. Google uses the term for AI Overviews and AI Mode ¹. Because each small search is narrow, a focused page that answers just one of them can sometimes be cited even if it did not rank well for the original question ⁴⁵.

One prompt can become many narrow searches, run together or in a loop, and the answer is built from what they return.
Our test: Gemini chose to run 2 searches, neither repeating our wording. A Claude agent told to run 6 to 10 ran 5.
There is no standard count. One study found about 1.6 fan-out queries per ChatGPT answer that searched; another found 7.61 ²³.
Write one clear page per distinct need, not per phrase. Google says pages made for every variation can break its spam policy ⁴.

What is query fan-out in AI search?

Query fan-out is when an AI search tool turns your one question into several smaller web searches, runs them, and builds a single answer from the results. Google uses the name for AI Overviews and AI Mode ¹.

Call each of those smaller searches a sub-query. You type one prompt, but the tool may send out a handful of sub-queries, each worded differently from yours. It then reads what comes back and writes the answer with links to the pages it used.

Why should you care? Your page is not competing for your customer's exact words. It competes for the words the AI chose, and those can include terms the customer never typed.

Imagine asking a smart assistant, 'Which laptop should I buy for video editing?' A good assistant would not just look up that sentence. They would check reviews, compare prices and read a spec sheet, then give you one answer. Query fan-out is an AI search tool doing the same: your one question becomes several quick lookups behind the scenes. So your website needs clear pages that answer the small questions, such as price, who it is for and how it compares.

The model acts as a planner. It receives the prompt, decides whether a search tool call would help, and emits one or more search strings as tool calls. The runtime executes them in parallel or in a loop and returns the results to the model's context ⁶⁷.

Retrieval happens per sub-query, and some APIs bill each one separately. For Gemini 3 and newer models, Google charges $14 per 1,000 search requests, counted per query the model decides to run, while older Gemini models are billed per prompt ⁷. Anthropic's API lets developers cap searches per request with a max_uses setting ⁶.

What happened when we ran this exact question through two AI tools?

We gave the question 'what is query fan-out in AI search' to Gemini 2.5 Flash with Google Search switched on, and to a Claude agent with a web search tool, on 20 October 2026. Gemini chose to run 2 searches, and neither repeated our question word for word. The Claude agent had been told to run 6 to 10 distinct searches and ran 5, one of them our question unchanged, so read its rows as one plausible set of angles, not typical behaviour. The full record sits in our Query Fan-Out Explorer ⁹.

Both runs added words we never typed, such as 'information retrieval', Google AI Mode, ChatGPT, Perplexity, study and tools. We grouped the seven searches into six needs: a definition, an older information-retrieval angle, Google's own account, advice, evidence and tools. The two runs listed 15 websites between them, and only ahrefs.com appeared in both. Treat that as a rough signal: the Claude list holds the 8 domains its agent reported, and the Gemini list holds the sites Google Search grounding attached, so low overlap proves little.

Checking visibility in more than one assistant is still sensible, because each tool picks its own searches and sources.

Sites each run listed on 20 October 2026 (Claude: up to 8 domains reported by the agent; Gemini: sites attached by Google Search grounding)

Gemini: conductor.com, ahrefs.com, screpy.com, mdmarketingdigital.com, kopp-online-marketing.com, ekamoira.com, searchengineland.com, eseospace.com
Claude: ahrefs.com, peec.ai, searchenginejournal.com, elogic.co, wellows.com, texta.ai, llmpulse.ai, surferseo.com
In both runs: ahrefs.com

The 7 searches recorded for one prompt, with our reading of what each was after
Tool	Search it ran	What it seems to be after
Gemini	query fan-out AI search	The question cut down to keywords
Gemini	query fan-out in information retrieval	An older information-retrieval angle (the nearest older idea is query expansion)
Claude	what is query fan-out in AI search	Our prompt, unchanged
Claude	Google AI Mode query fan-out technique explained	Google's own account of the idea
Claude	how to optimize content for query fan-out	Practical advice
Claude	ChatGPT query fan-out sub-queries study how many searches	Evidence on how many searches run
Claude	tools to see query fan-out queries Perplexity ChatGPT	Tools that reveal the searches

How does query fan-out work?

The model decides whether to search, writes the searches, runs them, and answers from what it finds. OpenAI, Anthropic and Google document versions of this loop, and the steps below follow their guides.

Decide. The model chooses whether a search would improve its answer. OpenAI says the model can choose to search or not based on the prompt ¹⁰, and Anthropic says Claude decides when to search ⁶.
Split. If it searches, it writes one or more search strings. Google's Gemini guide says the model can generate one or several search queries on its own ⁷.
Run. The searches run at the same time or in a loop. Anthropic says the search step can repeat several times within one request ⁶. Google's version is covered in the next section.
Read and rank. The results are scored and the best passages kept. Perplexity says it splits documents into self-contained spans, retrieves them with both keyword and meaning-based search, then reranks them ¹².
Write. The model composes the answer and links the pages it used. OpenAI and Anthropic both describe answers with inline citations ¹⁰⁶.

What does Google say about query fan-out?

Google confirms that AI Overviews and AI Mode may use query fan-out, and says sites need no special optimisation for it. Four Google pages cover what it has published.

Search Central. Fan-out means issuing multiple related searches across subtopics and data sources. Google adds that there are no additional requirements or special optimisations for these features ¹.
The AI Mode launch post (March 2025). Related searches run at the same time across subtopics and several data sources, and the results are brought together ¹¹.
The optimisation guide. Fan-out is a set of concurrent, related queries generated by the model, and SEO best practices still apply ⁴.
Deep Search. Google's May 2025 post says it can issue hundreds of searches, but shows it as still in development, with details that may change ¹³.

How many searches does one AI answer run?

There is no single number. A tool may run no search, one, a handful or hundreds, depending on the tool, the mode and the question. The table lists the published figures and the setup behind each.

Mode matters as much as the model. OpenAI describes three: a single pass over top results, an agentic mode where a reasoning model keeps searching until it is satisfied, and deep research that can draw on hundreds of sources ¹⁰. Anthropic's research feature uses a lead agent that splits the question into subtasks for several subagents working in parallel ⁸.

Why do the numbers differ so much? Ahrefs says many other studies used commercial 'best X' prompts that trigger more brand lookups, and that free, logged-out sessions search less than paid or reasoning ones ². A small logged-in test it lists (57 conversations, mostly software prompts) reported 15 to 40 fan-out queries per prompt in Thinking mode, though Ahrefs only lists it in a table and says the studies are not directly comparable ². Read each figure as what happened in that setup, not as what ChatGPT always does.

1.6average fan-out queries per ChatGPT response that searched, in one Ahrefs study of free, logged-out use (August to September 2026)²

Published figures and the setup behind each
Source	Setup	What it reported
Ahrefs study ²	Free, logged-out ChatGPT, 66.6 million responses, August to September 2026	14.4% of responses searched, with about 1.6 fan-out queries each
Nectiv study ³	GPT 5.6 Sol through the API, about 4,000 prompts across industries (undated page, lightly documented method)	7.61 fan-out queries per prompt on average
Anthropic web search docs ⁶	Guidance on capping searches per request (not a measured average)	1 to 3 for simple facts; 10 or more for comparisons
Anthropic Research ⁸	Rules of thumb described in June 2025 for its Research feature (not measured averages)	Simple fact-finding: 1 agent, 3 to 10 tool calls. Comparisons: 2 to 4 subagents, 10 to 15 calls each
Google Deep Search ¹³	AI Mode capability announced in May 2025	Can issue hundreds of searches
Perplexity Deep Research ¹⁴	Deep Research mode, February 2025 (vendor claim)	Dozens of searches and hundreds of sources

Why does query fan-out matter now?

It matters because, in Google's AI Overviews, the cited page is often not one of the top 10 results for your customer's wording. An Ahrefs update published on 2 March 2026 found about 38% of cited URLs sat in the top 10, against about 76% in its July 2025 study ⁵. Ahrefs suggests fan-out may explain the drop. It also notes that AI Overviews moved to Gemini 3 in January 2026, that its parsing improved and that the sample was larger, so the two figures are not a clean before-and-after ⁵.

In February 2026 Google said queries in AI Mode are three times longer than traditional searches ¹⁵. Longer questions tend to hold more parts, so there is more to split (our reading, not a measured link).

The added words show what the model wants to find. In Ahrefs' ChatGPT data, '2026' appeared in 22.9% of fan-out queries in September, 'official' in 13.1% and 'price' in 8.9% ². Ahrefs reports that about 6.8% of fan-out queries used the site: operator, which limits a search to one website. Its summary and body describe the base differently, so treat the figure as approximate ². Among the roughly 160 most-searched sites, 75.9% of those searches targeted brand sites ². Treat these as hints, not rules.

Ahrefs' own advice is to state prices, specifications and support details in plain text on pages the model can find ².

Does query fan-out make long-tail pages matter more?

Possibly, but none of the studies we cite measures it. Each sub-query is narrow, and a page that answers that narrow question directly has a better chance of being retrieved for it, so treat this as a working rule, not a finding. A long-tail query is a specific, multi-word search, such as 'tools to see query fan-out queries', that few people type compared with a broad term like 'AI SEO'. Fan-out turns one big prompt into several of these.

Here is a made-up example, modelled on the word patterns in Ahrefs' data ². For 'best bookkeeping software for a small agency', sub-searches could look like 'Tool A price 2026', 'Tool A reviews' and 'Tool A vs Tool B'. We did not record these.

The engines also work below the page level. Microsoft says assistants like Copilot break content into smaller structured pieces, and those pieces are ranked and assembled into answers ¹⁶. Perplexity splits documents into self-contained spans and retrieves each one separately ¹². So a short, self-contained section may be quoted even when the rest of the page is not (our reading).

Google adds limits. Its guide says AI systems can judge relevance without an exact query match, but that separate content for every variation, mainly to manipulate rankings or AI answers, violates its scaled content abuse policy, and that content need not be cut into tiny pieces for AI ⁴.

What does a page plan look like for a real prompt?

Group the searches by need, then decide where each need is answered. The table does that for the seven searches from our test: seven searches became six needs. Three are answered on this page (definition, Google's account and evidence), one is background covered in the terms table (the older angle), and advice and tools get their own pages because readers want steps and a working tool there.

From seven searches to a page plan
Need	Searches that share it	Where the answer should live
Definition	what is query fan-out in AI search; query fan-out AI search	One definition page, which is the page you are reading
Older angle	query fan-out in information retrieval	The terms table below (query expansion); no separate page needed
Google's account	Google AI Mode query fan-out technique explained	The Google section above, built on Google's own documents; see also Google AI Mode
Advice	how to optimize content for query fan-out	A how-to guide: How to do query fan-out research
Evidence	ChatGPT query fan-out sub-queries study how many searches	A sourced data section, updated as new studies appear (the count table above)
Tools	tools to see query fan-out queries Perplexity ChatGPT	A tool page: Query Fan-Out Explorer

What do people get wrong about query fan-out?

One prompt means one search. The vendors' own documents describe tools that run none, one, several or, in Google's Deep Search, hundreds of searches, depending on the question and mode ¹⁰⁶¹³.

Only Google does this. Google uses the phrase for its own products, but Anthropic, OpenAI and the Gemini API each document models that write several searches for one request ¹⁰⁶⁷.

It is brand new. Automatic query expansion was first proposed in 1960 ¹⁷. What differs now is that a language model plans separate searches and reads their results itself ⁶⁷.

How can you use query fan-out this week?

Pick your five most valuable customer questions, see what AI tools search for, and check that you have one clear page for each need. For finding the questions, see how to find what people ask ChatGPT and how to build a prompt list.

List five questions customers ask before they buy. Use their words, taken from sales calls and emails.
Run each one in two or three assistants. The free Query Fan-Out Explorer lists our recorded examples and, when live runs are available, shows the searches Gemini runs for your prompt. With developer access, you can also read them from the API response (see the FAQ below).
Group the searches by need, not by wording, using the merge-or-split rule above.
Search your own site for each need. If no page opens with a direct answer, you have found a gap.
Write or fix the page answer first. State the facts these searches often add: price, the current year, and official details such as specifications and support details ².
Repeat every month. The searches change between runs, and the models change over time ²⁵.

Fan-out worksheet (copy and fill in)

Customer question: ______
Tool and date: ______ (for example Gemini, 2026-10-20)
Searches the AI ran, copied exactly: 1. ______ 2. ______ 3. ______
Need behind each search (definition, how-to, price, comparison, evidence, tool, local): ______
Do we have a page that answers this need in its first sentence? yes / partly / no
Merge or split? Same answer: merge. Different answer or format: split.
Page to write or fix, and who owns it: ______

How is query fan-out different from similar terms?

Query fan-out names one specific thing: the extra searches an AI writes and runs for a single prompt. It overlaps with several older and newer terms, which the table separates.

Terms often mixed up with query fan-out
Term	What it means	How it differs from fan-out
Query expansion	A search engine reformulates a query to improve retrieval. Automatic versions date to 1960 ¹⁷.	Expansion adds terms to one query. Fan-out sends separate searches and reads each set of results.
Grounding	Google's word, in its Gemini API, for answering with the help of live Google Search results ⁷.	Grounding is the whole feature of answering from live search results. Fan-out is the part where the model writes several searches.
Retrieval-augmented generation (RAG)	A model fetches outside text before it answers; see what RAG is.	Fan-out is one way of deciding what to fetch.
Deep research	A mode that runs many searches over minutes and writes a long cited report ¹³¹⁰.	Fan-out at larger scale, with the model deciding what to search next.
Passage ranking	Scoring parts of a page instead of the whole page; see passage ranking and chunking.	Fan-out decides which searches run. Passage ranking decides which parts of the results get used.

What an AI assistant searches when it answers this

We gave the question below to AI search tools and recorded the web searches they ran behind the scenes. This is called query fan-out. Each phrase is a page you could publish to be found and cited.