Fan-out queries: what ChatGPT, Claude and Gemini actually search

We collected the fan-out queries by sending 100 prompts to three engines on October 7, 2026, with web search switched on, using the same DataForSEO calls our fan-out tool makes:

• ChatGPT: the ChatGPT app, read by DataForSEO's scraper, with search forced on (the model reported itself as gpt-5-6).
• Claude: Claude Haiku 4.5 through Anthropic's API, with web search forced on.
• Gemini: Gemini 3.6 Flash through Google's API with Google Search grounding, left to decide for itself when to search.

The prompts are ten everyday questions in each of ten topics (software, personal finance, health, home, travel, shopping, local services, marketing, tech help and careers), written the way people type them, in US English. We ran 20 of the prompts two more times on every engine to test repeatability, and pulled Google's US top 10 for the 122 searches the engines ran on 30 of the prompts. Every prompt, search and cited domain is in the CSV listed under References.

Three limits are worth stating up front. These are API and scraper results, not a logged-in person's app with memory and history. Google AI Mode's own fan-out can't be seen from outside Google, so it isn't here. And one day of runs is a snapshot, which is exactly why we repeated a subset.

How many fan-out queries an AI runs depends heavily on the engine: ChatGPT ran 2.2 per prompt on average, Claude ran exactly 1 every time, and Gemini searched for only 54 of the 100 prompts.

The table puts the fan-out query counts for all three side by side.

ChatGPT ranged from 0 to 6 searches. By topic, software drew the most ChatGPT searches (3.9 a prompt) and shopping the fewest (1.1), since for product questions it usually crammed every model name into a single long search. Claude ran one short search on every prompt, even though Anthropic's documentation says its models can search several times in one answer; that's the setup our tool uses, and bigger Claude models may search more. Gemini's 46 silent answers fit Google's own description of grounding: the model first decides whether a Google search would improve the answer at all, and on those prompts it answered from what it knew and cited nothing.

Fan-out queries barely repeat word for word: we asked 20 of the prompts three times on each engine, and not one of ChatGPT's 122 different searches came back in all three runs.

97% of them appeared in only one run. Even counting two searches as the same when they shared 70% of their words, only 11% of ChatGPT's searches recurred in all three runs, and its first search was never the same in all three runs. Gemini was similar (2% exact repeats). Claude was the steady one: 21% of its searches repeated word for word, and its first search was identical in all three runs for 8 of the 20 prompts.

What did repeat was the substance. In the four product questions we repeated, ChatGPT named 18 brands and products across the three runs, and 14 of them showed up every time. For "best CRM for a 10-person sales team" the sentence around them changed on every run, but HubSpot, Pipedrive, Zoho and Salesforce were in all three.

The sources were less steady. Only 25% of the domains ChatGPT cited across three runs were cited in all three, and 58% appeared once. Claude's figure was 23% and Gemini's 14%. So if you're tracking AI visibility by checking one prompt once, you're reading a single roll of the dice.

Fan-out queries differ between ChatGPT, Claude and Gemini so much that the three never ran the same search for the same prompt: across 100 prompts, no search appeared in all three engines' lists, ChatGPT never matched Claude or Gemini once, and Claude and Gemini matched on a single prompt.

Their sources barely overlapped either. On a typical prompt, any two engines had fewer than 1 in 10 of their cited sites in common.

The styles are easiest to see on one prompt. Asked for the best noise cancelling headphones under 300 dollars, ChatGPT searched "best noise cancelling headphones under $300 2026 Sony WH-1000XM4 Bose QuietComfort Momentum 4 prices", Claude searched "best noise cancelling headphones under 300 dollars 2026", and Gemini ran "best active noise cancelling headphones under 300" plus a version with "2025 2026" on the end.

ChatGPT names brands and checks official pages. When a prompt asked it to recommend a product or service without naming any, its searches named specific brands anyway, in 21 of 24 cases; 16% of its searches included the word "official" and 12% used a site: operator. It cited a government site on 25 of the 100 prompts (the IRS for the Roth IRA question, the CDC, NIH and ENERGY STAR elsewhere), and never once cited YouTube or Reddit. More in what search engine ChatGPT uses.

Claude searches the plain question. Its searches were short, literal and often built on "best" plus a year, it named a brand in none of those 24 product prompts, and it cited roundups and publishers, Forbes and NerdWallet most often. Its index is a different one, too: Claude's web search runs on Brave.

Gemini searches for reviewers. In 9 of the 24 product prompts it put a review publisher into the search (Wirecutter in 6 of them, plus RTINGS, Consumer Reports, Edmunds and Car and Driver, Bankrate, Sleep Foundation and The Verge), and it cited YouTube on 22 and Reddit on 20 of the 54 prompts where it searched. Neither ChatGPT nor Claude cited either site at all.

Fan-out queries led to Google's top 10 less often than most GEO advice assumes.

On 30 prompts, only 8% of the pages ChatGPT cited, 3% of Claude's and 14% of Gemini's were in Google's US top 10 for the search that engine had just run. At the level of whole sites it was 31%, 9% and 24%: the engines often cited a site that ranked, just not the page that ranked.

Read those as rough floors. Our Google results came through a SERP feed that left the brand's own site out of 2 of 6 brand pricing searches we spot-checked, which hurts ChatGPT most because it cites official pages so often. And ChatGPT and Claude aren't searching Google in the first place, as far as anyone can tell, so low overlap for them is expected; Gemini is the one that grounds in Google Search.

The useful conclusion is that page one for the exact string isn't the finish line. ChatGPT's search returned about 22 results per prompt and it cited about 4; Gemini cited 8 pages per prompt when it searched. The engines read past the top few results and then choose. Ranking gets you into the pile, and being the clearest, most authoritative answer on the topic gets you picked, which is Rank = Authority + Relevance applied twice: once by the search engine and once by the model choosing from what came back.

To find fan-out queries for your prompts, you need the searches the engine actually ran, not a list of guesses, and there are four places to get them.

1. Our free query fan-out tool runs a prompt through ChatGPT, Claude and Gemini with search switched on and shows every search each one ran, once or three times so you can see which come back, with Google's top 10 for each search and your site marked where it appears.
2. The APIs report them. Claude's web search tool returns each query it ran in the response, and Gemini's grounding with Google Search returns the queries the model executed.
3. Bing Webmaster Tools' AI Performance report lists grounding queries, the searches Microsoft's AI ran before citing your pages.
4. Google AI Mode isn't an option. Google hasn't exposed AI Mode's fan-out queries, so anyone selling you "AI Mode fan-out" data is selling an estimate.

To use fan-out queries in SEO, treat them as a map of what the engine needs to find, not as a list of pages to write.

1. Collect the searches for your 10 to 20 money prompts, on each engine, over three runs.
2. Group them by entity and intent, not by exact wording, because the wording won't repeat and the brands and topics will.
3. Map each group to one page that answers it by name, in the title, H1 and slug. Google's guide is blunt here: creating separate content for every fan-out variation mainly to manipulate rankings or AI answers violates its scaled content abuse policy, and a pile of thin pages doesn't make a site more relevant anyway.
4. Check where those pages rank in the index each engine reads: Google for Gemini, Bing for ChatGPT, Brave for Claude.
5. Build Authority to the pages that should be retrieved, and get onto the shortlist the engines start from: reviews, comparisons and mentions in the publications they cite.

None of that is a new discipline. It's keyword research where the searcher happens to be a model, followed by the same SEO you'd do anyway, and Google's own guide says as much: from Google Search's perspective, optimizing for generative AI search is "thus still SEO".

What most people overlook about fan-out queries is that the unit worth tracking is the entity, not the query.

The engines rarely search the same words twice, they never search the same words as each other, and yet the brands, publishers and topics inside their searches are stable enough to plan around: ChatGPT kept 14 of its 18 named products across three runs, and Gemini kept naming the same handful of reviewers from one prompt to the next. To check whether Google files your own brand as an entity, and under what type and ID, run its name through our knowledge panel checker.

There's a calendar wrinkle too. About 4 in 10 searches on every engine carried a year, but not always the right one: Claude put 2024 into 13 of its 100 searches, and Gemini often searched "2025 2026" together. A page that only answers the question for this year can miss a search pinned to last year, so the pages you most want cited should answer the question in their title and opening lines without leaning on the year.

These are the fan-out queries questions we get most often.

What is the fan-out approach?

The fan-out approach is splitting one question into several related searches, running them, and combining what comes back into one answer. Google describes AI Overviews and AI Mode as using it, and ChatGPT, Claude and Gemini all did some version of it in our test.

Can you see Google AI Mode's fan-out queries?

You can't see Google AI Mode's fan-out queries, because Google hasn't exposed them anywhere public. The nearest public window is Gemini with Google Search grounding, which is what our tool uses, and it's labeled as Gemini, not AI Mode.

Are fan-out queries the same every time?

Fan-out queries are not the same every time: none of ChatGPT's 122 different searches came back word for word in all three of our runs. The brands and topics inside them were far more stable than the wording.

How many fan-out queries does ChatGPT run?

ChatGPT ran 2.2 fan-out queries per prompt in our test, between 0 and 6, with comparison and software questions at the high end.

Do AI assistants ever search the prompt as typed?

AI assistants rarely search the prompt as typed: ChatGPT did it 0 times in 216 searches, Claude 3 times in 100 and Gemini once in 91. Optimizing a page for the exact prompt wording is optimizing for a search that mostly never runs.

Should I write a page for every fan-out query?

You shouldn't write a page for every fan-out query. Google's guide warns that doing so mainly to manipulate rankings or AI answers counts as scaled content abuse, and the queries change from run to run, so map them to the pages you already have and strengthen those.