Query fan-out explained
Query fan-out is the step where an AI assistant rewrites one prompt into several searches and merges the results. Most of those searches have no keyword volume, so you find them in assistant interfaces, API responses, tracking tools and Bing's grounding query report.
In short
- Google's documentation says AI Overviews and AI Mode may use a "query fan-out" technique, issuing multiple related searches across subtopics and data sources.
- Research by Seer Interactive and Nectiv, cited by Ahrefs in March 2026, found an average of 9 to 11 fan-out queries per prompt, with 59% of prompts triggering 5 to 11 and 24% triggering 12 to 19.
- Over 95% of fan-out queries have no recurring human search volume, according to Ahrefs' March 2026 article, so keyword tools do not list them.
- Peec AI's 2026 analysis found ChatGPT added the word "best" to 24.3% of advice-style questions when the user had not typed it.
- Bing Webmaster Tools has shown the grounding queries behind Copilot citations since February 2026, which is a first-party view of fan-out queries.
Query fan-out is what happens between your prompt and the answer. The assistant does not search for what you typed. It writes several searches of its own, runs them, merges the results and composes one reply. A page gets cited because it ranked for one or more of those hidden searches, and not because it ranked for the prompt.
How AI search chooses sources places fan-out in the wider pipeline. This lesson goes into the step itself: what the rewrites look like, how to find them and what to do with them.
What is documented
Google’s documentation on AI features says AI Overviews and AI Mode may use a “query fan-out” technique, issuing multiple related searches across subtopics and data sources. Google presents this as a way to show a wider and more diverse set of links.
Google’s Gemini API documentation shows the same behaviour from the developer side. It says the model may decide to run several search queries for a single prompt, and gives the example of one question about a football final producing two searches, one for the winner and one for the score.
Beyond that, the detail comes from companies that sell tracking tools, so read the numbers as vendor data.
How many searches
| Source | Finding |
|---|---|
| Seer Interactive and Nectiv, as cited by Ahrefs in March 2026 | An average of 9 to 11 fan-out queries per prompt. 59% of prompts trigger 5 to 11, 24% trigger 12 to 19, and the highest seen was 28 |
| Ahrefs, March 2026 | One ChatGPT Deep Research run made 420 searches for “buy red phone case” |
| Peec AI, 2026, over 20 million fan-outs | Fan-outs per query were 6.8 for Grok, 2.1 for ChatGPT and 1.4 for Perplexity |
| Peec AI, 2026 | The average ChatGPT fan-out query grew from about 6 words in October 2025 to about 12 in January 2026 |
The figures Ahrefs cites and the Peec figures do not agree, and we cannot reconcile them from the published summaries. They differ in which assistants, modes and prompts were sampled. The safe reading is that a normal answer rests on a handful of searches and a research mode on dozens or hundreds.
What the rewrites look like
Ahrefs’ March 2026 article groups the reformulations into types. Peec’s data adds two more patterns.
- Recency. The model appends the current year, turning “best smartphones” into “best smartphones 2026”.
- Comparison. A question about one product produces searches that pit it against rivals, such as “Asana vs Monday”.
- Implied questions. A prompt about buying solar panels also triggers “how much do solar panels cost”.
- Context. Location or earlier conversation is folded into the search.
- “Best”. Peec found ChatGPT inserted the word “best” into 24.3% of advice-style questions even when the user had not used it.
- Language switching. Peec found that on non-English prompts, 43% of ChatGPT’s research steps ran on the English-language web.
These patterns explain two things link builders see. Third-party comparison lists and review pages are retrieved often, because the model keeps asking for “best” and “versus”. And a brand that is well covered in Greek or Polish can still be missing from answers, because part of the research ran in English. International link building covers the language question for links.
Why a page ranking sixth can win
After the searches run, the results are merged. Ahrefs describes the merge as reciprocal rank fusion, a method that rewards documents appearing across several result lists. A page that ranks around sixth for several fan-out queries can beat a page that ranks first for only one.
The effect shows in the data. Ahrefs found in March 2026 that 37.9% of URLs cited in AI Overviews ranked in the top 10 for the original query, down from about 76% in July 2025. Ahrefs attributes the change to heavier fan-out. That is the vendor’s explanation and Google has not confirmed it.
One further claim deserves a label. Analyst Mike King of iPullRank reads Google’s patents as showing that AI Mode retrieves passages and not whole pages, runs chains of reasoning and personalises results. That is an interpretation of patents, and Google has not confirmed it.
How to find the hidden queries
Over 95% of fan-out queries have no recurring human search volume, according to Ahrefs. Keyword tools will not show them. These are the places they do appear.
| Where | What you see | Limits |
|---|---|---|
| Bing Webmaster Tools, AI Performance | Grounding queries, defined by Microsoft as “the key phrases the AI used when retrieving content” that was cited | Your own site only, Copilot and Bing only, and Microsoft says the data is a sample |
| The assistant’s own interface | Some assistants list their searches in a thinking, activity or sources panel. Ahrefs notes that ChatGPT’s thinking view shows part of its reasoning | Not shown in every mode, and the display changes without notice |
| Model APIs with search | Google’s Gemini API documentation says the response contains the search queries the model executed | An API answer can differ from the consumer app |
| Tracking tools | Ahrefs says its Brand Radar product reports fan-out queries for ChatGPT and Perplexity. Other trackers offer similar reports | Vendor data, limited to the prompts they run |
A working method that uses these:
- Take 20 to 50 of the prompts from your AI visibility measurement set.
- Run each several times and collect every fan-out query you can see, from the interface, an API or a tool.
- Export your grounding queries from Bing Webmaster Tools and add them.
- Group the queries by theme, not by exact wording. The wording changes on every run, but the themes repeat: price, alternatives, reviews, a named competitor, a year.
- For each theme, note which pages are cited.
This is a working method, not a tested procedure, and you will only ever see a sample.
What to do with them
Check your own coverage. For each recurring theme, ask whether a page on your site answers it directly. Pricing and comparison themes are the usual gaps.
Check third-party coverage. Most cited pages will not be yours. List the review sites, comparison articles and forum threads that recur, and see whether they mention you. Getting onto those pages is the off-page task, and it is set out in link building for AI visibility.
Keep ranking for ordinary queries. Fan-out queries are still searches against an index. Pages with links and authority rank for more of them. Which pages need links still applies.
Use digital PR for the themes you cannot answer yourself. A model looking for “best” and “reviews” wants independent sources. Digital PR is how you earn them.
What not to do
Do not publish a page for each fan-out query. They are synthetic and unstable. On our reading, mass-produced pages aimed at them risk falling under Google’s scaled content abuse policy, which covers many pages generated mainly to manipulate rankings. Google’s policy page does not mention fan-out queries. Do not fill your own blog with “best” lists that rank you first. Lily Ray reported in February 2026 that several software sites with many such pages lost 29% to 49% of their Google organic visibility, and she noted it was one factor among several.
Do not treat any list of fan-out queries as complete. No platform publishes them in full.
Where to go next
The searches run against different indexes on each assistant. Which search index each AI assistant uses covers that. Terms such as citation and co-occurrence are in the glossary.
Common questions
What is query fan-out in simple terms?
An AI assistant takes your question, writes several search queries of its own from it, runs them all and builds one answer from the combined results. You see the answer and usually not the searches.
How many fan-out queries does a prompt trigger?
Research by Seer Interactive and Nectiv, cited by Ahrefs in March 2026, found an average of 9 to 11 per prompt. Peec AI reported lower figures per assistant, such as 2.1 for ChatGPT and 6.8 for Grok. The two used different methods, and research modes run far more.
Can I see the fan-out queries for my own prompts?
Sometimes. Some assistants show their searches in a thinking or sources panel, Google's Gemini API returns the queries it ran, several tracking tools report them, and Bing Webmaster Tools lists grounding queries that led to your pages being cited.
Do fan-out queries have search volume?
Mostly not. Ahrefs' March 2026 article says over 95% receive no recurring human searches. They are written by the model for one answer, so they are not keywords in the usual sense.
Should I create a page for every fan-out query?
No. They change with each run. Group them into recurring themes and check whether your pages, and the third-party pages that mention you, cover those themes.
