Query Fan-Out: What AI Engines Actually Search, and How to Optimize for It

What Query Fan-Out Is

Query fan-out is the step between a question and an AI answer. The engine works out what it needs to know, writes search queries of its own, runs them, reads the pages that come back and writes the answer from those pages. The searches belong to the engine, not to the person asking. Microsoft says the query Copilot sends to Bing is different from the user's prompt and consists of a few words informed by it. Fan-out is one of the mechanics generative engine optimization works with: you are optimizing for searches you did not write.

How ChatGPT fanned out one buyer question

Every search ChatGPT ran for one question in the first of three runs, asked plainly. Highlights mark the kinds of addition the test counted; other new words are left plain.

You ask

“best running shoes for flat feet”

ChatGPT6 searches in 2 requests

1. 1best stability running shoes flat feet 2026 reviews Brooks Adrenaline ASICS Kayano Saucony Guide Hoka Arahi
2. 2flat feet running shoes stability vs neutral evidence guidance podiatry
3. 3site:brooksrunning.com Adrenaline GTS 25 running shoe
4. 4site:asics.com GEL-KAYANO 33 running shoe
5. 5site:newbalance.com Fresh Foam X 860v15 running shoe
6. 6site:saucony.com Guide 19 running shoe Hoka Arahi 9 official

One answer

Written from the pages these searches return.

Added to the questionAaBrand or productAaYearAaOfficial site or site: searchAaComparison wordAaReview word

Searches as the API reported them. Which pages they returned was not recorded.

Google's term

Google used the term when it introduced AI Mode in March 2025, as a limited, opt-in experiment in Search Labs, describing a technique that issues multiple related searches at once across subtopics and data sources. Its documentation for site owners says both AI Overviews and AI Mode may use it. At I/O in May 2025, Google's head of Search described AI Mode breaking a question into subtopics and issuing a multitude of queries on the user's behalf. We covered what that means for Google's own surfaces in our AI Mode guide.

The other AI engines do it too

The term is Google's, but the mechanism is everywhere. OpenAI says its model can search as part of its reasoning and decide whether to keep searching. Anthropic says Claude's searching can repeat multiple times in a single request. Google's grounding documentation for Gemini notes that one prompt can lead to one or more searches, Perplexity's Pro Search runs multiple searches across the web, and Copilot Chat shows users the exact queries it sent to Bing.

Why engines split a question

A question bundles several needs: the product, the price, who it suits, what owners say about it. One search for the whole sentence returns pages that cover part of it, so the engine fetches pages for each need instead. Research backs the design. In a 2025 ACL workshop paper, a retrieval pipeline that split questions into sub-questions and then reranked the passages it found improved retrieval by 36.7% (MRR@10) and answer accuracy by 11.6% (F1) over standard retrieval-augmented generation. The earlier self-ask method had a model ask itself follow-up questions before answering, and showed that plugging in a search engine to answer them improved accuracy further.

Deep research is a different mode

The large numbers you will see quoted belong to deep research modes. Google says Deep Search can issue hundreds of searches. An ordinary AI Mode answer is smaller: a Google Search engineering director described it as doing a dozen searches in the time it takes to do one. Anthropic's web search documentation says a simple factual question typically takes 1–3 searches, and comparative research can take 10 or more.

How We Recorded the Searches

Google does not publish AI Mode's searches. Its own guide says no third-party tool has access to its internal ranking or AI systems. ChatGPT, Gemini and Claude do report their searches through their APIs, so we recorded them there.

We wrote 30 everyday buyer questions, phrased the way people type them, and asked each one three times on each engine, in two setups: plainly, with no system prompt, and with an instruction added as a system prompt. Among other things, that instruction asks for community discussions such as Reddit threads, review sites such as G2 and Trustpilot, editorial coverage and official documentation, prefers sources from the last 12 months for current products and trends, and asks for the URLs used to be cited. Comparing the two setups, run hours apart on the same day, shows how far an instruction moves the searches.

These are the engines' APIs, not their apps. The apps add instructions of their own: Anthropic publishes the system prompt its apps send at the start of every conversation, and the API does not send it. So read the numbers as how each engine searches when asked directly; an app can search more or less. And 30 questions show patterns, not market-wide rates.

How Many Searches Each Engine Runs

It depends on the engine, and sometimes the answer is none. Asked plainly, Claude ran one search for 83.3% of answers and none for 10%. Gemini ran two for 55.6% and none at all for 26.7%, answering about a quarter of the time from what it already knew. ChatGPT, which we set to always search, averaged 5.2 searches, and 31.1% of its answers ran seven or more.

ChatGPT packs several searches into one request

ChatGPT's 5.2 searches came in 1.6 search requests per answer: it sends a batch of queries at once, so counting its requests understates how much it searches. Each search was about as long as the question: a median of 8 words for ChatGPT, 7 for Gemini and 7 for Claude, against 7 in the questions.

Not hundreds

Our test cannot see AI Mode, and Google's own description of it, about a dozen searches, is above the averages the APIs ran here. What holds across all of them is the order of magnitude: an ordinary answer rests on a dozen searches or fewer, rarely more. The hundreds belong to deep research.

What Engines Add to Your Question

The searches are rarely the question as typed: each engine rewrites it, in its own way. Here is every search the three engines ran for three of the questions, in the first run:

Across all the plain answers, these are the things the engines added that the question did not have:

What the engines added to their searches

Share of answers in which at least one search added this and the question did not have it. Asked plainly, no system prompt.

Added to a search	ChatGPT	Gemini	Claude
“official” or a site: search	88.9%	0%	0%
A brand or product name	66.7%	8.9%	6.7%
A year	53.3%	20%	73.3%
“best” or “top”	14.4%	32.2%	33.3%
Price words	23.3%	6.7%	11.1%
Review words	21.1%	4.4%	3.3%
A review site by name	12.2%	15.6%	0%
Comparison words	11.1%	3.3%	6.7%
“reddit”	0%	0%	0%

These count words in the searches, not the pages the searches returned. A year means 2024 or later. Brand and product names come from a hand-tagged list of the companies, products and businesses the searches named, including those reached through a site: address; review sites count separately, and places, months, units, government bodies and associations do not count. Review words are review, reviews, reviewed, rated, rating, ratings, and tested. Comparison words are vs, versus, compare, compared, comparison, alternative, and alternatives. Price words include price, cost, cheapest, fee, budget, a free plan or trial, and a dollar amount.

ChatGPT goes to the source

In 88.9% of its answers ChatGPT aimed its searches at official sources: a search added the word “official” or used site: to search a single website, usually a brand's own or a government agency's. In 66.7% it named brands or products the question did not. For the CRM question it searched HubSpot's, Pipedrive's and Zoho's pricing pages directly. It also searched for authorities: the EPA and the Association of Home Appliance Manufacturers (AHAM) for the air purifier question, and the Consumer Financial Protection Bureau for money transfers.

Claude adds the year

In 66.7% of its answers Claude ran a single search with the year in it, such as “best running shoes for flat feet 2026”, which it ran in all three runs. Claude added a year in 73.3% of answers, always the current one; ChatGPT added one in 53.3%, nearly always including the current one. Gemini added a year in 20%, and 94.4% of those answers searched a past year, 2024 or 2025, often beside the current one.

Gemini and Claude add “best”

Gemini and Claude both reached for ranking words: Gemini added “best” or “top” in 32.2% of answers and Claude in 33.3%. Gemini also named a review site, such as Wirecutter, in 15.6%.

What no engine added on its own

Not one plain answer, on any engine, had a search containing “reddit”. That does not mean Reddit threads go uncited: a thread can come up for an ordinary search. We recorded the searches, not the pages they returned, so this test cannot say how often Reddit was cited. Comparisons were rare too. ChatGPT added a comparison word such as “vs”, “compare” or “alternatives” in 11.1% of answers, Claude in 6.7% and Gemini in 3.3%.

Most Searches Change When You Ask Again

Ask the same question again and you often get different searches. Only 5.1% of ChatGPT's searches reappeared word for word in another run of the same question, and 19.5% of Gemini's. Claude was the exception at 73.6%: it usually ran a single search, and asked again it often ran the same one.

Across engines the overlap is almost nil. Of the 74 cases where two engines both searched the same question, they shared an identical search once.

So a list of sub-queries is a snapshot of one run on one engine. What stays steady is what each engine adds most. In each of the three runs, ChatGPT aimed its searches at official sources in 86.7% to 93.3% of answers, Claude added the year in 70% to 76.7%, and Gemini added “best” or “top”, its most common addition, in 30% to 33.3%. Those are the targets that hold still.

An Instruction Changes the Searches

The apps people use add instructions of their own, as the Claude app prompt Anthropic publishes shows. To measure how much an instruction matters, we also asked the questions with an instruction added as a system prompt. Among other things, it asks for community discussions such as Reddit threads, review sites such as G2 and Trustpilot, editorial coverage and official documentation, prefers sources from the last 12 months for current products and trends, and asks for the URLs used to be cited.

Answers with “reddit” in a search went from none to 74.4% on ChatGPT, 88.9% on Gemini and 91.1% on Claude. Without the instruction, Reddit pages could still come up for an ordinary search; what changed is that the engines asked for Reddit by name. The engines also searched more: Claude from 1 to 2 searches per answer, Gemini from 1.4 to 2.3 and ChatGPT from 5.2 to 6.5. Gemini all but stopped answering from memory (no search in 1.1% of answers, down from 26.7%), and ChatGPT's searches with review words went from 21.1% of answers to 63.3%. For Claude, part of this is documented: Anthropic says whether Claude searches can be steered through the system prompt.

Answers with “reddit” in a search, once the instruction was added

Share of answers with at least one search containing “reddit”. Asked plainly, it was 0% on every engine. The instruction, added as a system prompt, names Reddit threads among the sources to draw on.

Measure	ChatGPT	Gemini	Claude
Answers with “reddit” in a search, plainly	0%	0%	0%
Answers with “reddit” in a search, instructed	74.4%	88.9%	91.1%
Searches per answer, plainly	5.2	1.4	1
Searches per answer, instructed	6.5	2.3	2
Answers with no search, plainly	0%	26.7%	10%
Answers with no search, instructed	0%	1.1%	0%
Answers adding a year, plainly	53.3%	20%	73.3%
Answers adding a year, instructed	88.9%	28.9%	76.7%
Answers adding review words, plainly	21.1%	4.4%	3.3%
Answers adding review words, instructed	63.3%	7.8%	2.2%

Reddit pages can still come up for searches that do not name it; those are not counted.

Fan-out belongs to the setup, not the question

The same question produces different searches on different engines, on different runs, and under different instructions, and the instructions differ from one app to the next. A list of “the” sub-queries for a question shows one of many. That is why the advice in the rest of this guide aims at what engines consistently look for rather than at any one search.

Why Fan-Out Decides Who Gets Cited

When an engine searches, its answer cites pages those searches found. If a narrower search finds your page, you can be cited for a question you never ranked for. If every search lands somewhere else, ranking for the question itself does not help. The patterns above point at four kinds of page.

Your official pages

ChatGPT aimed its searches at official sources in 88.9% of answers, sometimes with a site: search for a brand's pricing or product page. If your prices, plans and specs sit in an image or behind a “contact sales” button, that search finds nothing to quote.

Pages that name the alternatives

ChatGPT named brands or products the question did not in 66.7% of answers, sometimes several in one search: its first running-shoe search above named Brooks, ASICS, Saucony, and Hoka, with a model of each. Pages that compare you with named alternatives, and the “best X” lists that name you beside them, match searches like that.

Pages that carry the date

Claude put a year in its searches for 73.3% of answers and ChatGPT for 53.3%. A page that says which year its facts apply to, and is kept current, matches those searches; our refresh checklist covers the upkeep.

Ranked lists

Gemini and Claude added “best” or “top” to the searches for about a third of their answers. A “best” search asks for a ranking, and ranked lists are the format AI engines cite most often, as our page-type study found.

How to Optimize for Query Fan-Out

You cannot write for searches you cannot see, and the ones you can see often change from run to run. Six steps work anyway.

Step 1: Start from the questions buyers ask

Fan-out starts from a question, not a keyword. List the questions a buyer would put to an AI engine before choosing you, phrased the way they would type them, and favour ones whose answer can name products; our guide to choosing prompts explains why.

Step 2: Plan the angles, then check real searches

Our free Query Fan-Out Generator suggests five to seven keyword variations of a question: the bare topic, a “best” version, a longer phrasing with features or audience, a comparison and a buying-intent phrasing. Use it as a checklist of angles a page should answer, not as a record of what any engine searched. To see real searches, Copilot Chat shows the queries it sent, and the ChatGPT, Gemini and Claude APIs report theirs, as in our test. Treat any list as a sample; the next run will often differ.

Step 3: Put the facts engines look for on the page

Write prices, plans and specs as text, with the date they apply to. Name the alternatives a buyer compares and say how you differ. Say who the product suits and who it does not. These are the facts the narrower searches go looking for, and a page that states them gives the engine something to quote.

Step 4: Answer the sub-questions on one strong page

Put the likely follow-up questions on the main page under clear headings rather than spinning each into a thin page of its own. Google's AI optimization guide says there is no requirement to break content into tiny pieces, and that creating separate content for every variation of a search primarily to manipulate rankings or AI responses violates its scaled content abuse policy.

Step 5: Be on the pages the searches land on

Many searches look for rankings, reviews and named alternatives rather than for you. Being named, accurately, on the review sites, “best” lists and comparison pages in your category puts you in front of searches that never name you.

Step 6: Measure the answers, not the searches

Because the searches change from run to run, measure the answer instead: whether it names you, where, and which pages it cites. Track that for the same questions every week, and you see the effect of fan-out without needing to see the fan-out.

Four Fan-Out Myths

Myth 1: AI runs hundreds of searches per question

Deep research modes do. Ordinary answers do not: asked plainly, the engines in our test ran between 1 and 5.2 searches per answer on average.

Myth 2: You need a page for every sub-query

Google says separate content for every variation, made primarily to manipulate rankings or AI responses, violates its scaled content abuse policy, and the variations change from run to run anyway. One page that answers the follow-ups does the job.

Myth 3: Fan-out queries are keywords you can target

Only 5.1% of ChatGPT's searches repeated word for word, and engines shared an identical search for the same question once in 74 chances. Almost none were short keywords either: only 0.4% of the plain searches had 3 words or fewer. They are rewrites carrying facts, not a keyword list.

Myth 4: You can see Google's fan-out

You cannot. Google does not publish AI Mode's searches, and it says no third-party tool has access to its AI systems. What you can see are the searches the ChatGPT, Gemini and Claude APIs report, which is why our test used them, and the queries Copilot Chat shows.

The Query Fan-Out Checklist

Run this against the page you most want AI engines to cite for one buyer question. Every no is a gap a narrower search can expose.

Query fan-out checklist (one page, one buyer question)

1. The page answers the question in its first two sentences.
2. It names the products, brands or providers a buyer compares, on the page.
3. Prices, plans and fees are written as text, with the date they apply to.
4. The page says what year its facts are for, and it is current.
5. Specs and limits are numbers (size, capacity, battery life, fees),
   not adjectives.
6. It says who the product suits and who it does not.
7. It compares you with the main alternatives, in a table if you can.
8. Your official pricing and product pages say the same as this page.
9. Likely follow-up questions are answered here under clear headings,
   not spread across thin pages of their own.
10. AI crawlers can read it: allowed in robots.txt, no login, text
    present without running scripts.
11. You appear on the review sites, "best" lists and comparisons that
    answer the same question.
12. You check every week whether AI answers to this question name you
    and which pages they cite.

How Ranqo Fits

Fan-out happens inside the engine, so the thing to measure is the answer. Ranqo asks your buyers' questions every week on the AI engines your plan covers and records whether each answer names you, where you rank and which pages it cites. The Sources page shows which domains and pages each engine cites for your questions.

To plan the pages, try the free Query Fan-Out Generator: enter a buyer question and get five to seven keyword variations of it, such as a “best” version, a comparison and a buying-intent phrasing, to check your page against.