Short answer
Query fan-out is how AI search engines research a question: they split it into several related searches, run them at once, and write the answer from the pages those searches return. Google uses it in AI Overviews and AI Mode, and ChatGPT and Gemini write their own searches. A quick ChatGPT answer runs one or two; a reasoning mode can run twenty or more, and Google’s Deep Search hundreds. A brand is cited when its pages answer those searches.
Key takeaways
- An assistant seldom searches for the question as typed. It writes its own searches and adds words such as “best”, “price”, “reviews”, “official” and the current year.
- The count depends on the engine and the mode: about 1.6 per ChatGPT answer with fan-out queries in Ahrefs’ October 2026 data, 6.8 per prompt for Grok in Peec AI’s, and hundreds in Google’s Deep Search.
- A page competes in the searches behind a question as well as in the question itself: 38% of the pages cited in AI Overviews rank in the top 10 for the query typed.
- In a September 2026 study, brands that were retrieved and also searched for by name were mentioned in 91.4% of GPT answers, against 2.8% for brands absent from both.
- Planning for fan-out is topical authority work: collect the searches behind each priority question, group them by intent, and give each group a page that answers it.
What query fan-out is
Query fan-out is the technique AI search engines use to research a question before they answer it. The engine reads the question, writes several narrower searches that cover its parts, runs them at the same time, and builds one answer from what they return. Each of those searches is a fan-out query.
When Google rolled AI Mode out across the US in May 2025, it described the technique this way: AI Mode breaks a question into subtopics and issues “a multitude of queries simultaneously” on the searcher’s behalf [1]. Its documentation says AI Overviews and AI Mode may both use it, issuing related searches across subtopics and data sources [2]. OpenAI describes the same step in ChatGPT, which typically rewrites a prompt into one or more targeted queries for the search providers it works with [3]. In Google’s Gemini API, the model decides whether a search would improve its answer, then generates one or more queries and runs them [4].
Fan-out is the first stage of the pipeline I describe in how AI assistants choose sources: the searches decide which pages become candidates, and only candidates can be cited. Covering those searches is central to generative engine optimization (GEO) and answer engine optimization (AEO) alike.
How many searches one question becomes
The count depends on the engine, the mode and the question. These are the measurements published so far:
| Engine and mode | Searches per question | Measured by |
|---|---|---|
| ChatGPT, free and signed out | About 1.6, in the 14.4% of answers that had fan-out queries | Ahrefs, 66.6 million answers, August and September 2026 [5] |
| ChatGPT | 2.1 | Peec AI, 5 million fan-out queries across three engines, April 2026 [6] |
| Perplexity | 1.4 | Peec AI, April 2026 [6] |
| Grok | 6.8 | Peec AI, April 2026 [6] |
| ChatGPT through the API | 7.61 | Nectiv, about 4,000 prompts, as compiled by Ahrefs [5] |
| ChatGPT Thinking, signed in | 15 to 40 | Suganthan Mohanadasan, 57 conversations, as compiled by Ahrefs [5] |
| Google Deep Search in AI Mode | Hundreds | Google [1] |
So a quick answer runs one or two searches, and a reasoning mode on a buying question can run twenty or more. Ahrefs puts the gap down to the prompts and the plan: its study used broad prompts from real search demand, many other studies used commercial “best X” prompts, and free, signed-out sessions search less than paid or reasoning models [5]. The higher counts come from those commercial questions, the ones that decide which brand an assistant recommends.
What fan-out queries add to the question
Assistants seldom search for the question as it was typed. They add words that point at the pages they want to read, and Ahrefs and Peec AI found the same patterns across millions of fan-out queries:
- The current year: “2026” appeared in 22.9% of ChatGPT’s fan-out queries in September, up from 17.6% in August [5].
- Official sources: “official” appeared in 13.1% of English fan-out queries in September, 6.8% of all fan-out queries searched a single site with the site: operator, and three in four of the most common site: searches pointed at a brand’s own site [5].
- Prices and reviews: “price” grew from 6.04% to 8.88% of fan-out queries between August and September, a rise of 47%, and “reviews” grew by a third [5]. ChatGPT adds “reviews” even when the prompt never asked for them [6].
- Best-of lists: for advice-style questions such as “should I use X?”, ChatGPT added “best” to 24.3% of its fan-out queries [6]. In Peec AI’s data, “Which Samsung Galaxy should I get?” became “best Samsung Galaxy phones comparison 2026” [6].
- Brands and comparisons: ChatGPT writes named brands and head-to-head comparisons into its searches [6].
Each pattern names a page type. A search with the year looks for a page that is kept current. A site: search looks for the facts on the brand’s own site, in plain text. A price or review search looks for a pricing page and for reviews on the sites the engines read. A best-of or comparison search looks for a page that compares the options by name.
Why fan-out decides which brands get cited
An assistant can only cite a page that one of its searches returns. With one or two searches behind a ChatGPT answer, a page has to match one of them, or ChatGPT never sees it [5].
That makes the searches behind a question worth more than the question itself. In Ahrefs’ study of AI Overviews, 38% of the cited pages ranked in the top 10 for the query that was typed [7]; the rest came through the other searches. One researcher, Metehan Yesilyurt, found that ChatGPT merges the results of its searches with reciprocal rank fusion, a method that scores a page higher when it appears in several of them [6]. A site that covers a subject in depth turns up in search after search, and that is how topical authority works in AI search.
The strongest position is a brand the assistant already connects with the category, so that it writes the brand’s name into its own searches. A September 2026 study of 34,960 runs of unbranded prompts on GPT and Gemini measured this [8]. When neither the brand nor its site appeared among the retrieved sources, the brand was mentioned in 2.8% of GPT answers and 3.8% of Gemini answers. When the brand’s own site appeared among the sources, that rose to 49.0% and 58.4%. When the site appeared and the assistant also searched for the brand by name, it reached 91.4% and 100%. The study is observational, so it shows association rather than cause, and its direction is clear: be retrieved, and be the brand the assistant thinks to search for.
How to see the fan-out queries behind your prompts
Fan-out queries run out of sight, and four sources show them:
- Gemini’s API: grounding with Google Search returns the search queries the model ran with each answer [4], the closest view of how a Google model searches.
- Ahrefs Brand Radar: it logs the fan-out queries for ChatGPT and Perplexity next to each answer, for its own prompt index and for custom prompts, most often two per prompt [9].
- Peec AI: its prompt reports list the latest fan-out queries ChatGPT, Perplexity and Grok ran for each tracked prompt [6].
- The browser: the network tab of a browser’s developer tools shows the queries inside the requests as an answer loads, one prompt at a time [6].
AI Overviews and AI Mode show their supporting links [2], so for Google I read the cited pages next to the queries Gemini runs. Read fan-out queries over several runs and several weeks. They vary from one run to the next and shift with each model update: price searches in ChatGPT rose by 47% in one month [5]. I collect them across engines and runs with tools I build, and group them before anything is written.
How I plan content for fan-out queries
Fan-out is a content strategy question, and I plan for it in six steps:
- Collect the questions buyers ask an assistant in your category, then the fan-out queries each engine runs for them, over several runs.
- Group the searches by intent: main keywords for the core intent of each question, satellite keywords for the context the assistant adds, such as price, reviews, comparisons and the year.
- Map your pages and the top 20 competitors against those groups to find the gaps.
- Give every main keyword one page and a brief, and answer its satellite searches as sections and passages on that page, each with the answer in its first lines. Link the pages as hubs and spokes, so the site covers the whole subject.
- Put the facts the searches look for on your own site in plain text: prices, specifications, dates and named comparisons, kept current.
- Make the brand an entity the engines connect with the category: the same name, description and category everywhere, and a place in the reviews and comparisons the fan-out queries reach.
Step 4 is where topical authority comes from. One strong page per intent, with its sub-questions answered inside it, gives the engines a page that matches search after search, where a page per search spreads the subject thin. I measure the result as fan-out coverage, the share of the searches behind each priority question that a page on the site answers, next to citations in each engine.
On one program for a FinTech comparison brand across 26 markets, AI Overview and assistant citations rose from 14 answers to more than 160, and AI-assistant referral sessions from under 500 to 18,000 a month. This planning is month two of my programs: keyword research and content planning.
A worked example: one question, twenty searches
The company and its searches here are illustrative, built on the patterns above. A company that sells accounting software for builders tracks one prompt: “What is the best accounting software for a small construction company?” Over a month of runs in ChatGPT, Perplexity, Grok and Gemini’s API, that prompt produces twenty distinct searches:
- Best-of and comparisons (4): best accounting software for construction companies 2026; best accounting software for small contractors; construction accounting software comparison; top job costing software for builders.
- Features (5): job costing accounting software construction; accounting software with progress billing; retainage tracking accounting software; certified payroll accounting software; construction accounting mobile app.
- Prices (3): construction accounting software price per user; site:yourbrand.com pricing; construction accounting software free trial.
- Reviews (3): construction accounting software reviews; yourbrand reviews contractors; site:reddit.com accounting software for contractors.
- Integrations (2): accounting software that integrates with estimating software; construction accounting software bank feeds.
- Official facts (2): site:yourbrand.com job costing; yourbrand official site.
- Recency (1): construction accounting software new features 2026.
The company ranks well for “construction accounting software”, and its pricing page shows prices only after a sign-up. Its pages answer five of the twenty searches.
The plan follows from the groups: a job costing page and a progress billing and retainage page, each with its sub-questions answered inside it; a comparison page that names the main competitors with real prices; prices per user in plain text on the pricing page; an integrations page; and outreach to the two review sites the engines retrieved most. Next month the same prompt runs again, and fan-out coverage and citations are compared engine by engine.
Frequently asked questions
What is a fan-out query?
A fan-out query is one of the searches an AI assistant writes and runs while it answers a question. Query fan-out is the technique, the fan-out queries are the searches it produces, and the pages they return are the candidates for citation.
How can I see the fan-out queries for my prompts?
Gemini’s API returns the searches the model ran with each grounded answer. For ChatGPT and Perplexity, tracking tools such as Ahrefs Brand Radar and Peec AI log them per prompt, and a browser’s developer tools show them one answer at a time.
Are fan-out queries the same as People Also Ask?
They come from different places. People Also Ask shows questions many people search for, so their search volume is known. Fan-out queries are what the assistant needs to answer one prompt, so many have no measured volume. People Also Ask suits keyword research; fan-out queries show what a page has to contain to be cited.
Should I write a page for every fan-out query?
No. Group them by intent. Each main keyword gets one page, and most fan-out queries are answered by a section or a passage on it. One strong page per intent builds topical authority, and a page per search spreads it thin.
How often do fan-out queries change?
Often. They vary between runs of the same prompt and shift with model updates: in Ahrefs’ data, price searches in ChatGPT grew by 47% in one month. Read them monthly on a fixed prompt set, and plan from the patterns that hold across runs.
Sources
- Google, Elizabeth Reid, “AI in Search: Going beyond information to intelligence”, 20 May 2025. blog.google
- Google Search Central, “AI features and your website”, last updated 10 December 2025. developers.google.com
- OpenAI Help Center, “Searching the web with ChatGPT”, updated October 2026. help.openai.com
- Google AI for Developers, “Grounding with Google Search”, last updated 13 October 2026. ai.google.dev
- Ahrefs, Louise Linehan, “How ChatGPT Does Query Fan-Out (15 Million Queries Studied)”, 12 October 2026, corrected October 2026. ahrefs.com
- Peec AI, Tomek Rudzki, “Patterns we see in ChatGPT query fanouts”, 5 May 2026. peec.ai
- Ahrefs, “Update: 38% of AI Overview Citations Pull From The Top 10”, 2 March 2026. ahrefs.com
- Benjamin Tannenbaum, “From Prompt to Recommendation: A Fitted Stage Model of Brand Visibility in AI Search”, arXiv, 19 September 2026. arxiv.org
- Ahrefs Help Center, “How to view fanout queries generated by AI”, 3 September 2026. help.ahrefs.com