Skip to content
Chalk Labs / AI SEO
How-to guide

How to Do Query Fan-Out Research for AI SEO

By Rahil Jain, founder of Chalk LabsUpdated How we research

The short answer

To do query fan-out research, take one real buyer question, run it in an AI assistant that searches the web, and record every sub-search it runs. Repeat the run several times, group the sub-queries into themes, then map each theme to a page you have or need. Re-run monthly, because results change.

  • Fan-out means the AI splits one question into several smaller searches. Google says AI Overviews and AI Mode may do this.
  • Google does not expose its fan-out queries, so you can only infer them. Some assistants, such as Claude with web access, list their searches; others do not.
  • One run is an anecdote. Run each prompt three to five times per engine and keep the queries that repeat.
  • Map clusters to pages. Do not write a separate page for every variation, which Google calls scaled content abuse when done mainly to manipulate rankings or AI responses.

What is query fan-out research?

Query fan-out research is the practice of finding out which smaller searches an AI assistant runs to answer one question, then making sure your site answers those smaller searches. The searches the AI runs on your behalf are called sub-queries.

Google describes the technique as running multiple related searches across subtopics and data sources to build a response in AI Overviews and AI Mode 1. Its AI optimization guide defines fan-out as a set of concurrent, related queries generated by the model to request more information 3. For a plain definition, read what is query fan-out.

Do it step by step

0 of 8 done0%

  1. Write the question a customer would type, such as 'how do I choose a payroll provider for 20 staff'. Avoid your own jargon. Choose three seeds: one early-stage, one comparing, one ready to buy.

    Tip: Take wording from real sales calls, not from your product pages.

  2. Use ChatGPT with search on, Perplexity, Claude with web access or Google AI Mode. Open a fresh logged-out or new chat so past memory does not colour the result.

  3. Copy the searches shown by the assistant, plus the sources it cites. Paste each into the capture sheet with the engine name, date and run number.

    Tip: Some Google features do not show their searches. For those, record the cited URLs and infer the themes.

  4. Run the identical prompt again in new chats. Mark which sub-queries repeat. A query seen in three of five runs is a stable theme; one seen once is noise.

  5. Lowercase the queries, remove the year if it is the only difference, and count duplicates with the bash one-liner in the templates. Sort by count.

  6. Group queries by the page type that would answer them: definition, how-to, price, comparison, review, official facts, evidence. Aim for five to eight themes per seed.

  7. For each theme write one of three verdicts: have a strong page, have a weak page to improve, or need a new page. Improve before you create. Check for an overlap with an existing page first.

  8. Write or fix the top three pages, using the answer first and one original element on each. Re-run the same seeds next month, add the new data to the sheet and compare.

Your ticks are saved in this browser only.

Copy-ready templates

Fan-out capture sheet columns (tab-separated)

text
seed_prompt	engine	date	run_no	sub_query	cited_url	theme	my_page_or_gap

Paste into the first row of a spreadsheet. One row per sub-query per run.

Count repeated sub-queries from a text file

bash
tr 'A-Z' 'a-z' < queries.txt | sed -E 's/ 20[0-9]{2}$//; s/[[:space:]]*$//' | sort | uniq -c | sort -rn

queries.txt holds one sub-query per line. The output lists each unique query with its count, highest first.

Prompt to cluster captured sub-queries

text
Here is a list of search queries an AI assistant ran for the question: [SEED].
Group them into 5 to 8 themes. For each theme give: a short name, the queries in it, and the page type that would answer it best (definition, how-to, price, comparison, reviews, official facts, evidence).
Do not invent queries that are not in the list.

[PASTE QUERIES]

Check the clusters yourself. Assistants sometimes merge themes that need separate pages.

Why does fan-out change how you plan content?

Your page may never rank for the question the user typed. It only has to be a strong answer to one of the sub-queries. Google said at its I/O event on 20 May 2025 that AI Mode breaks a question into subtopics and issues a multitude of queries simultaneously, and that Deep Search can issue hundreds of searches 2.

Ahrefs links this to citations. Its March 2026 study of 863,000 keyword results found about 38% of AI Overview citations came from the top 10 for the original query, against about 76% in an earlier July 2025 analysis 5. Ahrefs suspects fan-out explains the shift, but says Google does not expose its fan-out queries, so the link is an interpretation and the two analyses used different methods 5.

The takeaway: ranking for the head term is no longer the only way in. Mapping the sub-queries shows the pages that decide whether you appear.

What do you need before you start?

You need one to three seed prompts, access to at least two AI assistants that search the web, and a spreadsheet. Allow about two hours for your first seed. That time is our estimate.

Choose engines that show their searches. In our own check on 20 October 2026 we ran the seed through Claude with web access, which listed the searches it ran. For engines that hide them, use the Query Fan-Out Explorer or the Bing AI Performance report (see 'How do you check it worked?' below).

What does a real fan-out look like?

We ran the seed 'how to do query fan-out research for AI SEO' once through Claude with web search on 20 October 2026. It ran six sub-queries, listed below. This is one run on one engine, so treat it as an illustration of shape, not a statistic.

Notice the pattern: each sub-query asks for a different type of page. One wants a definition, one a method, one tools, one an original source, one examples and one a study. A single page answering only 'what is it' would meet one of six.

Sub-queries observed in one Claude run, 20 October 2026
Sub-query Claude ranWhat the searcher wantsPage type that answers it
query fan-out AI search explainedA definitionWhat-is page
how to research query fan-out for SEOA methodHow-to guide
query fan-out tools freeTools to tryTool list or free tool
Google AI Mode query fan-out patent sub-queriesThe original sourcePage citing primary documents
optimize content for fan-out queries examplesWorked examplesExample or case page
ChatGPT search queries behind the scenes studyEvidence from a studyResearch summary

What words show up in fan-out queries?

Ahrefs studied 66.6 million logged-out ChatGPT responses from August and September 2026. Only 14.4% included a web search, and those averaged about 1.6 fan-out queries each 4. So expect some prompts to produce no sub-queries at all.

Where ChatGPT does search, the words are practical. In the same study, 'price' rose from 6.0% to 8.9% of fan-outs, and 'official' appeared in 13.1% of English ones in September 4. In a sample of 1,474 running-shoe responses, 'best' appeared 206 times, 'review' 117 times and 'vs' 105 times, for a retail product category; B2B terms may differ 4.

Reuse these as a checklist for your own seed: does a result exist for 'best', 'price', 'review', 'vs' and 'official' versions of your topic?

Where does fan-out research go wrong?

The biggest mistake is treating one run as the truth. Ahrefs notes that AI Overview results are probabilistic and change with each query, so any capture is a snapshot 5. Repeat the run.

The second is a page for every variation. Google says creating separate content for every possible way people might search, including fan-out queries, mainly to manipulate rankings or AI responses violates its scaled content abuse policy 3. Group queries by the page that would answer them.

The third is chasing queries you cannot serve with something distinctive. Google's guide asks for non-commodity content with unique expert or experienced takes 3.

Does matching a fan-out query help a page get cited?

It appears to help, though the evidence is correlational. In an Ahrefs study of 1.4 million ChatGPT prompts, cited page titles matched the prompt more closely than non-cited titles (0.602 against 0.484 on its similarity score), and cited titles matched the fan-out queries well too 6.

Ahrefs warns that comparing cited and uncited pages without splitting by source type can mislead, so it analysed each type separately 6. Use this as support for clear, query-shaped titles, not as a guaranteed ranking factor.

How do you check it worked?

Re-run the same seed set one month after you publish. Count how many cluster themes now have a strong page, and whether your pages appear as cited sources in the engines that show citations.

Bing's AI Performance report helps for Copilot. It lists grounding queries, the key phrases the AI used when retrieving content it later cited, and Microsoft notes they are a sample of overall citation activity 7. Compare them with your clusters to see which sub-queries your pages already serve.

When is fan-out research not worth it?

Skip it if you have fewer than ten pages and no clear offer, because the research will outrun your ability to publish. Write your core service and pricing pages first.

Skip it also for questions where the answer is a quick fact. The AI answers those without searching in most cases, as the 14.4% figure above suggests 4.

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.

The prompt

“how to do query fan-out research for AI SEO”

The searches it ran
  • query fan-out AI search explained
  • how to research query fan-out for SEO
  • query fan-out tools free
  • Google AI Mode query fan-out patent sub-queries
  • optimize content for fan-out queries examples
  • ChatGPT search queries behind the scenes study

Observed October 20, 2026 using Claude web search. Claude agents were told to run 6 to 10 searches, so that count is our setting, not Claude's own choice. Assistants differ and change often, so treat this as a sample, not a rule. See how we collect it.

Frequently asked questions

Can I see Google's fan-out queries?

Not directly. Google confirms that AI Overviews and AI Mode may use query fan-out 1, but Ahrefs notes that Google does not expose the queries it simulates 5. You can infer themes from the sources cited, or use reports from Bing for Copilot.

How many times should I run each prompt?

Three to five runs per engine is our recommended minimum, not a published standard. AI answers vary between runs, so queries that repeat across runs are the ones worth planning pages around.

Is fan-out research the same as keyword research?

No. Keyword research starts from search volume for typed terms. Fan-out research starts from one question and records the smaller searches an AI generates. The two overlap, and the fan-out list is often more specific and varied.

Do I need a paid tool?

No. The free Query Fan-Out Explorer, a spreadsheet and manual runs are enough to start. Paid trackers help once you track many prompts every month, but they do not replace choosing good seed prompts.

How long before changes show up?

There is no published timeline we can cite. AI engines re-fetch pages at different speeds, so measure monthly and compare against your baseline run instead of expecting results in days.

About the author

Rahil Jain

Rahil Jain is an AI SEO (answer engine and generative engine optimisation) consultant and co-founder of Chalk Labs.

LinkedIn · Method and credentials

Sources

  1. AI features and your website · Google Search CentralOpened October 20, 2026. Supports: AI Overviews and AI Mode may use query fan-out, issuing multiple related searches across subtopics and data sources; last updated 2025-12-10.
  2. AI in Search: Going beyond information to intelligence · Google (The Keyword blog)Opened October 20, 2026. Supports: Post dated 20 May 2025: AI Mode breaks questions into subtopics and issues a multitude of queries simultaneously; Deep Search can issue hundreds of searches.
  3. Optimizing your website for generative AI features on Google Search · Google Search CentralOpened October 20, 2026. Supports: Defines query fan-out; separate pages per query variation made to manipulate rankings or AI responses violate scaled content abuse policy; asks for non-commodity content; updated 2026-07-10.
  4. How ChatGPT Does Query Fan-Out (15 Million Queries Studied) · AhrefsOpened October 20, 2026. Supports: 66.6 million logged-out ChatGPT responses, 14.4% searched at about 1.6 fan-outs each; 'price' 6.0% to 8.9%; 'official' 13.1% of English fan-outs; running shoes sample of 1,474 responses (best 206, review 117, vs 105).
  5. Update: 38% of AI Overview Citations Pull From The Top 10 · AhrefsOpened October 20, 2026. Supports: March 2026 study of 863K keyword SERPs: about 38% of AIO citations from top 10 vs about 76% in July 2025; fan-out explanation is a hypothesis; Google does not expose fan-out queries; results are snapshots.
  6. Why ChatGPT Cites One Page Over Another (Study of 1.4M Prompts) · AhrefsOpened October 20, 2026. Supports: Cited titles matched the prompt more closely (0.602 vs 0.484) and matched fan-out queries (0.656); authors caution that pooling source types can mislead.
  7. Introducing AI Performance in Bing Webmaster Tools Public Preview · Microsoft Bing Webmaster BlogOpened October 20, 2026. Supports: Grounding queries are key phrases the AI used when retrieving content it later cited; Microsoft calls them a sample of overall citation activity.
How this page was made. Built on Rahil Jain's AI SEO method and drafted with AI research assistance. Every statistic links to the page we read it on (opened October 20, 2026 or nearby). Published October 20, 2026; last updated October 20, 2026. Chalk Labs sells AI SEO services, so we have a commercial interest in this topic. Spotted an error? Tell us and we will fix it.