Share of answer is the operating metric for this practice.

How we count it
Request an audit

Insight · 13 min

One question, several searches

Query fan-out is the step where an AI search engine turns one question into several related search queries, runs them, and writes its answer from the pages those queries return. Google, which named the technique, defines it as “a set of concurrent, related queries generated by the model to request more information.” As of November 2026, Google AI Overviews and AI Mode, ChatGPT, Perplexity, Claude, Gemini, and Copilot all document or show some version of it. The page that gets cited is often the best answer to one of those sub-queries, not the page that ranks for the prompt.

In short

  • Fan-out means the engine searches with its own queries, not yours. Google documents it for AI Overviews and AI Mode. Microsoft, Anthropic, and Google’s Gemini API document that the model writes the search query.
  • Most engines run few sub-queries per prompt. Independent studies in 2026 put ChatGPT at about 1.6 to 2.1, Perplexity at about 1.4, and Grok at about 6.8. Google’s Deep Search “can issue hundreds.”
  • How often ChatGPT searches at all depends on the sample: 14.4% of free logged-out responses in Ahrefs’ data, 47.5% of prompts in DataForSEO’s, and about 91% of queries in AirOps’ UI study.
  • Fan-out is why citations and rankings for the prompt have come apart. Ranking still matters, but for the sub-query: in AirOps’ data, the first search result was cited 58.4% of the time.
  • Cover sub-queries with one page per intent and a section per likely follow-up. A page per variant is scaled content abuse under Google’s policy, and full coverage did not beat partial coverage in AirOps’ data.

What is query fan-out?

Query fan-out is the expansion of one prompt into a set of search queries before the engine answers. The model reads the prompt, decides what it needs to look up, writes queries for each part, sends them to a search system, and reads the results.

Google’s example: for “how to fix a lawn that’s full of weeds,” the fan-out queries “might include ‘best herbicides for lawns’” and related queries about removal and prevention. Google’s May 2025 announcement of AI Mode said it “uses our query fan-out technique, breaking down your question into subtopics and issuing a multitude of queries simultaneously on your behalf.”

Other operators describe the same step in their own documentation without using the name:

  • Microsoft Copilot parses the prompt, identifies terms where web information would help, and sends Bing “a few words informed by the user’s prompt,” which Microsoft says “is different from the user’s original prompt.”
  • Claude decides when to search, and “this process can repeat multiple times throughout a single request.” Anthropic’s API documentation says “single-fact queries typically use 1–3 searches; comparative or multientity research can use 10 or more.”
  • Gemini, when grounded in Google Search, “automatically generates one or multiple search queries and executes them.”
  • ChatGPT search rewrites the prompt into one or more targeted queries for its search providers, according to OpenAI’s help center as quoted in third-party coverage. We could not load OpenAI’s help page directly to confirm the wording.

Fan-out is a retrieval step. It decides which pages the engine reads. Which of those pages it then cites, and what it says about them, is a separate step covered in what gets cited.

How does each engine fan out?

Each engine fans out differently, and only part of it is documented. The table separates what operators say from what independent studies measured. Measured figures come from one vendor each, with different samples and methods, so compare the direction, not the decimals.

EngineWhat the operator documentsWhat studies measured
Google AI Overviews“May use” fan-out, “issuing multiple related searches across subtopics and data sources.” Appears only when judged additive to classic resultsGoogle publishes no counts. Details in how AI Overviews choose a source
Google AI ModeFan-out breaks the question “into subtopics” and issues “a multitude of queries.” Deep Search “can issue hundreds of searches.” May use different models from AI OverviewsGoogle publishes no counts
ChatGPTPrompt rewritten into targeted queries (help center, reported secondhand)About 1.6 fan-outs per response that searched (Ahrefs, Aug–Sep 2026). 2.1 per prompt (Peec, April 2026). 88.6% of queries produced exactly two (AirOps, April 2026). Average length roughly doubled from about 6 to about 12 words between October 2025 and January 2026 (Peec)
PerplexitySearches the web in real time. Deep Research “performs dozens of searches automatically”1.4 fan-outs per prompt (Peec). Keeps about 88% of the prompt’s words, and 14% of its queries are unique across daily runs (Profound)
CopilotGenerates a short Bing query from the prompt. Copilot Chat shows “the exact web search queries” in the citation sectionKeeps about 50% of the prompt’s words. 47% of queries unique across runs (Profound)
ClaudeSearches when needed, may repeat within a request. 1–3 searches for a single fact, 10 or more for comparative research (API docs)No independent count found
Gemini (API grounding)Generates “one or multiple search queries.” Returns the queries in the responseNo independent count found
GrokNot documented in sources we reviewed6.8 fan-outs per prompt (Peec)

Two patterns hold across studies.

ChatGPT rewrites the most. Profound tracked 10,000 prompts daily for 14 days in March and April 2026. ChatGPT’s fan-out queries kept about 13% of the prompt’s words, and 91% of them were unique across runs. Perplexity tended to issue the same one or two queries every day. Profound’s conclusion: a single check on Perplexity is fairly stable, while “monitoring needs to be ongoing on ChatGPT.”

Sub-queries simplify. Profound found that fan-outs on all three engines it studied lean toward plain fact lookups, even when the prompt asked for advice or a comparison. Location almost always survived the rewrite. Price points, filters, and multi-brand comparisons were often dropped or reinterpreted. A buyer’s “best tool under $50 a seat for a 30-person team” may become “best [category] tools” and “[category] pricing.”

Why do the numbers on how often ChatGPT searches disagree?

Because each study sampled a different product and prompt set. Three 2026 figures:

StudySampleShare with fan-out
Ahrefs, October 202666.6 million free, logged-out ChatGPT responses, Aug–Sep 202614.4% of responses
DataForSEO, May 2026100,000 prompts from its LLM Responses API database47.5% of prompts
AirOps with Kevin Indig, April 202616,851 queries run three times each in the ChatGPT interfaceAbout 91% of queries

Ahrefs issued a correction in October 2026: the 14.4% describes responses that contained fan-out queries, not every response that used the web. It also notes that logged-out use searches less than paid or reasoning use. AirOps chose queries across ten categories, many with commercial intent, which search more often. Grow and Convert’s 100 buying-intent prompts triggered a search 83 times.

The useful reading: casual and conceptual prompts are often answered from training data, and commercial, current, and product-specific prompts usually trigger a search. Your buyers’ questions are mostly the second kind.

Why does fan-out decouple citation from ranking?

Because the engine retrieves for its sub-queries, and a page can rank well for a sub-query without ranking for the prompt. The citation then looks like it came from nowhere if you only track the prompt.

The measurements agree on the direction:

  • Ahrefs ran 118,931 ChatGPT fan-out queries through Google in August 2025. Only 6.82% of the URLs ChatGPT retrieved were in Google’s top 10 for the same query, and 83.39% were not in Google’s results at all.
  • Grow and Convert tested 100 buying-intent prompts in ChatGPT and found that only 27% of cited sources ranked on Google for the fan-out queries, and 23% on Bing. Counting any page on a cited domain raised the share to about 50%.
  • For Google’s own AI Overviews, Ahrefs found 38% of citations in the top 10 for the query in March 2026. That analysis is in how AI Overviews choose a source.

Ranking still matters. It matters for the query the engine ran, inside the engine’s own search system. In AirOps’ data, the first result returned for a fan-out query was cited 58.4% of the time, against 14.2% at position ten. The gap between Google rankings and ChatGPT citations is mostly a gap between Google’s index and the ones ChatGPT searches, plus the gap between your keyword and the engine’s sub-query. That is why the house view is one system scored on one query universe: the sub-queries belong in it. How that view treats SEO, AEO, and GEO as layers of one pipeline is set out in SEO vs AEO vs GEO.

What do engines add to the sub-queries?

They add words that point at official, current, and priced information. Ahrefs’ October 2026 study of about 15 million ChatGPT fan-out queries found:

  • “Official” appeared in 13.1% of English fan-outs in September 2026, up from 10.1% in August.
  • “Price” rose from 6.04% to 8.88% of fan-outs in one month.
  • 6.8% of fan-outs were site: searches, and 75.9% of the most common site: searches pointed at a brand’s own website.

Profound’s October 2025 launch post listed “best,” “top,” “reviews,” and “2025” among the words ChatGPT most often added to prompts. Taken together: when a buyer asks about your category, the engine often goes looking for your own pricing and documentation pages by name. If those pages are thin, blocked, or rendered only in JavaScript, the sub-query finds a reseller or a review site instead.

These figures describe one model over two months. Ahrefs says OpenAI changes retrieval “constantly,” and Peec notes that a GPT-5.6 release in August 2026 changed fan-out counts. Re-measure before you plan around a specific modifier.

How do you find the likely sub-queries for a topic?

Collect them from the engines that show them, then fill the gaps from buyer evidence. No engine publishes its fan-out for your topic, so the list is always an estimate.

Where sub-queries are visible as of November 2026:

SourceWhat it showsLimit
Microsoft Copilot ChatThe exact web search queries sent to Bing, in the response’s citation sectionAvailable for 24 hours in the chat. Not in the Copilot pane in Word or PowerPoint
Bing Webmaster Tools AI PerformanceA sample of grounding queries that led to citations of your pagesYour own cited pages only. A sample, not the full set
Claude API web searchThe query Claude ran, in each server_tool_use blockThe API, not the claude.ai app
Gemini API groundingThe generated queries in the search call stepThe API, which may differ from the Gemini app
ChatGPT interfaceSearch queries are visible in some interfaces and captured by third-party toolsVendors capture them by scraping, and samples vary
Vendor tools (Ahrefs, Profound, Peec, others)Logged fan-outs for tracked promptsOne vendor’s sample and prompt set

Then add the questions buyers actually ask, which the engines are approximating: People Also Ask boxes, sales call notes, support tickets, and the comparison criteria in RFPs. For each priority prompt, write down the follow-ups: cost, time to set up, limits, integrations, alternatives, risks, who it is not for. The sub-queries you capture will mostly be simpler versions of these, plus brand names, “official,” “price,” and the year.

Should each sub-query get its own page?

No. Each distinct intent gets one page, and each likely sub-query gets a section on the page that owns its intent. A sub-query earns its own URL only when its intent differs from every existing page and nothing on the site serves it. That is a content-mapping decision, made the same way you would make it without AI search.

Google is explicit about the alternative. Its generative AI guide says creating pages for fan-out variations “primarily to manipulate rankings or generative AI responses in Google Search” violates its scaled content abuse spam policy. It adds that its systems “can understand synonyms,” so you do not need “enough ‘long-tail’ keywords.”

The best evidence also says exhaustive coverage does not win on its own. In AirOps’ April 2026 data, among pages closely matched to the prompt, those covering 26% to 50% of fan-out subtopics were cited 38.2% of the time, against 34.0% for pages covering all of them and 35.5% for pages covering none. That is an observational comparison with small gaps, and it does not prove that partial coverage is better. It does undercut the idea that a page must answer every sub-query to be cited.

SituationWhere it goes
Follow-up to the main question (cost, setup time, limits)A section on the main page, heading phrased as the question
Same intent, different wording (“pricing” vs “cost”)Nowhere new. The existing section covers both
Comparison with a named competitorIts own versus page, because the intent is distinct
Official fact the engine searches for by name (pricing, specs, policies)The canonical page for that fact, in HTML text, kept current
Question nothing on the site answers, with a distinct intentA new page, mapped and linked from the related ones

Each section should answer its heading in the first sentence, so the passage works when the engine retrieves the page for that one sub-query. The writing rules are in what gets cited.

How do you measure fan-out coverage?

Measure it as a chain: the sub-queries you logged, the page that should answer each one, whether that page is retrieved for it, and whether it is cited across repeated runs.

  1. Log sub-queries per prompt. For each prompt in your query set, record the sub-queries from every source you can see, with the engine and the date.
  2. Map each to an owning page. If no page owns it, that is a gap. If two pages compete for it, that is a cannibalization problem to fix in SEO.
  3. Check retrieval for the sub-query. Where you can, check whether the owning page ranks for the sub-query in the engine’s search system. Bing results matter for Copilot. Google matters for Gemini and AI Overviews.
  4. Run each prompt several times. AirOps found only 2.3% of page-query pairs cited in all three runs. A single run cannot tell a gap from noise.
  5. Report coverage and citation separately. “We have a section for 9 of 12 logged sub-queries” is one number. “Our page was cited in 4 of 10 runs” is another. Do not blend them into one score.

The operating metric that combines the runs is share of answer.

Worked example: mapping sub-queries for one prompt

This example is hypothetical. It uses a made-up payroll software vendor, not a client, and the sub-queries are illustrative of the patterns above, not logged data.

Prompt: “best payroll software for a 40-person company in Texas.”

Sub-queries an engine might run, following the patterns studies report (simplification, brand names, “official,” “price,” location kept):

  1. best payroll software small business
  2. payroll software pricing per employee
  3. [Vendor] payroll pricing official
  4. Texas payroll tax requirements employers
  5. [Vendor] vs [Competitor] payroll
  6. payroll software reviews

The mapping:

Sub-queryOwning pageStatusAction
1Category guideExists, ranks on page 2Retrieval work: internal links, title, answer first
2Pricing pagePrice behind a formPublish the price model and starting price in HTML, dated
3Pricing pageSame pageSame fix. This is the page the engine searches for by name
4NoneGap, distinct intentNew page on Texas employer payroll obligations, if the vendor can write it with authority
5Versus pageExists, all checkmarksRewrite with facts, sources, and where the competitor fits better
6Third-party review sitesNot on the vendor’s siteOff-site work: see why AI cites what others say about you

Six sub-queries produced one new page, two rewrites, one retrieval fix, and one off-site task. They did not produce six new pages.

Common mistakes

  • Tracking only the prompt. The citation often comes from a sub-query. Log the sub-queries and measure them.
  • A page per sub-query. Near-duplicate pages compete with each other and fall under Google’s scaled content abuse policy. One page per intent.
  • Treating one vendor’s fan-out log as the truth. Each study samples one product tier and prompt set. Compare sources and date them.
  • Optimizing for the exact wording of a sub-query. ChatGPT’s sub-queries are mostly unique run to run. Cover the question, not the string.
  • Forgetting the official page. Engines search for pricing, specs, and “official” sources by name. Make those pages complete, current, and in HTML.
  • Assuming the Google rank carries over. Each engine searches its own index. Check retrieval where the engine searches.
  • Reading one run as a result. Fan-out and citations vary between runs. Report rates over repeated runs.

Fan-out changes what you measure more than what you write. The page still has to be the clearest answer to a real question. The difference is that the questions are now several, and you can see some of them. Mapping them across a query universe is part of our GEO work.

Sources

Bring us the domain.

We will tell you where you are absent, which page is in the way, and what has to be true before a citation is possible.

Request an audit