What query fan-out searches for and how to rank for it

When ChatGPT searches the web for your question, it usually doesn’t search for exactly what you typed. OpenAI’s ChatGPT search help page says it “typically rewrites your query into one or more targeted queries”, and it picks its sources from what those searches return. Google AI Mode does the same thing, splitting a question into several searches at once.

That process is called query fan-out, and it decides which pages get cited. I wanted to see what the searches actually look like, so I asked ChatGPT 27 questions, from passport fees to note-taking apps, and logged every search it ran. They look nothing like the keywords in a keyword tool.

What query fan-out is

Query fan-out is an AI search engine turning one question into one or more searches of its own, then reading the results before it answers. Google’s guide to generative AI features defines it as “a set of concurrent, related queries generated by the model to request more information and fetch additional relevant search results to address the user’s query.”

Google’s example is someone asking “how to fix a lawn that’s full of weeds”. AI Mode might search for “best herbicides for lawns”, “remove weeds without chemicals” and “how to prevent weeds in lawn”, and cite pages from all three.

So a page doesn’t need to rank for the question a person typed. It usually needs to rank for one of the searches the engine wrote. You can’t see those searches in Search Console or a keyword tool, so I went and looked at them directly.

What ChatGPT searched for

ChatGPT’s searches were long and usually dated. In a first set of 20 questions, on topics my own sites cover, it ran 28 searches, averaging 13 words each, and 19 of the 28 included a year. Ten were simple questions, and the other ten were harder versions of the same topics: comparisons, two-part questions and more specific cases.

I then asked a few everyday questions to show the pattern in public. These are some of them, with the search ChatGPT wrote for each:

Three things stand out across all the questions.

ChatGPT names the publisher it wants. Many searches named a source, such as the State Department, the Department of Energy or a national statistics office, and seven of the 28 on my topics included the word “official”. Suganthan Mohanadasan, who reads ChatGPT’s network traffic, saw the same “official” suffix 17 times in his own captures.

ChatGPT adds a place and a date. The Austin rent question became a search for November 2026 rent data. OpenAI’s help page says ChatGPT estimates a general location from the person’s IP address and can add it to the search, so “good restaurants near me” becomes a search for restaurants in a named city. A local page competes for the version of the question with the town in it.

ChatGPT splits comparisons apart. The heat pump question became two searches, and comparisons on my own topics became up to three: one for the comparison and one for each option. Simple questions mostly got one search each.

Those counts match ChatGPT’s quick answers. Its thinking mode searches far more: Suganthan logged 15 to 40 searches for one product comparison, many of them site: lookups aimed at vendors’ pricing pages.

Why ChatGPT cites the official page over yours

In my test, when ChatGPT put “official” or a publisher’s name in its search, the publisher won. The passport question went to the State Department, and nothing else was cited.

I wouldn’t try to beat an official source on a search that asks for one. If it answers the question well, ChatGPT is right to cite it.

The publisher names come from what the model already knows, before any result comes back. Suganthan found that ChatGPT’s first search contained brands the person never typed in 21 of 27 conversations. A site the model has never heard of won’t be written into the search, which is why being written about elsewhere matters as much as the page itself.

Other sites win where no official source answers the exact question. No official body publishes a monthly average rent for one bedroom apartments in Austin, so ChatGPT cited five rental listing sites. That’s the useful lesson. Look at the searches ChatGPT runs for your topic, find the part the official source doesn’t cover, and make that part the page.

Comparison searches need a comparison page

A comparison question gets a comparison search, and ChatGPT cites pages that compare. When I asked “notion vs obsidian for personal notes”, ChatGPT ran one search: “Notion vs Obsidian personal notes comparison 2026 pricing offline privacy features”. All 12 pages it read were “Notion vs Obsidian” articles, and it cited four of them, including Zapier’s. Neither notion.so nor obsidian.md came back at all.

Ask about one app and the picture flips. For “is obsidian good for personal notes”, ChatGPT searched for “Obsidian notes app personal knowledge management features sync privacy markdown mobile official” and cited three pages on obsidian.md.

So a product’s own pages win the single-product question and don’t even get read for the comparison. If people compare you with a rival, the page that answers that question has to compare the two, and it can be yours as long as the comparison is fair.

How AI Mode fans out, and why Search Console can’t show it

Google AI Mode runs several searches at once, and Search Console doesn’t show them. Google’s Search Console help says AI Mode “groups the user’s question into subtopics and searches for each one simultaneously”. Search Console reports AI Mode under the question the person asked, and a follow-up question counts as a new query.

So a page can show up in Search Console for a question it was never written for, because it ranked for one of AI Mode’s hidden searches. The Generative AI performance report doesn’t help here either, because it shows pages, countries and devices but no queries. And many conversational questions are rare enough that Google hides them, so whatever you find is a floor.

Google hasn’t published what AI Mode’s searches look like, but Gemini’s are a fair guide. Moz collected 5,333 grounding queries from Gemini and found they averaged 6.56 words, and a third mentioned a year. ChatGPT’s searches in my test were about twice that length. Both add years and specifics that people leave out.

ChatGPT cited pages that barely show in Google

A weak Google ranking didn’t stop ChatGPT citing a page. In a separate ChatGPT test with questions about one of my own sites, ChatGPT cited a page that had 2 Google impressions over three months, at an average position of 32.5.

Another page ChatGPT cited in that test does better in Bing than in Google. Over four weeks it averaged position 8 in Google and 3.6 in Bing, with more than twice as many Bing impressions. OpenAI doesn’t say how much each search provider is used. Its help page says ChatGPT sends searches to third-party providers and links Microsoft’s and Shopify’s privacy policies for them (until August 2026 the same page listed Bing and Shopify by name), and Suganthan’s traffic captures showed results arriving through several retrieval services, including one labelled bing on some accounts. So a page that ranks well in Bing has one more way into ChatGPT’s answers, even when Google buries it.

How to see the fan-out searches for your own questions

SEO Skill shows you the searches ChatGPT runs for the questions you care about. Give it up to 10 questions. Any domain works as the target, so you can check a competitor’s questions too:

Each ChatGPT answer lists the searches it ran and the pages it cited, and says whether it cited or mentioned your site. The notes rank the searches that came up most often across all your questions, which is the list to write for. The same call checks whether Google AI Mode and AI Overviews cite you, but it can’t show their searches, because DataForSEO returns none for Google’s engines. It runs on your own DataForSEO key, for about $0.007 a question across the three engines.

Pick questions your customers really ask, including the comparisons and two-part questions. In my first set, the harder questions produced 17 of the 28 searches.

ChatGPT decides first whether a question needs a search at all. When an answer shows no searches, ChatGPT answered from the model’s training, as it would for a person asking the same question. How-tos and definitions often go that way. Those answers cite no page, and the way into them is the model’s training: being named next to your topic on pages other sites quote, which How to get cited by ChatGPT covers. (My own test above forced a web search on every question, so it shows what ChatGPT searches when it does search.)

Google AI Mode’s searches aren’t visible anywhere, but the questions that led to them turn up in Search Console. The AI queries section of seo report finds the conversational questions and follow-ups, such as “how much is it” or “what about the pro plan”, and shows which page Google showed for each one:

How to rank for fan-out searches

Ranking for a fan-out search is ordinary SEO aimed at a longer, more specific query. These are the changes I’d make, based on which pages got cited in my test.

1. Put the words the engine adds into your title and first paragraph. ChatGPT searched for years, places, months and data sources, such as “November 2026” in the Austin rent search. Put the year in the title only when the page covers something that changes each year, and update it when the year turns. Keep it out of the URL either way, so the page keeps its address when you update it.
2. Say where your numbers come from. The heat pump search asked for the Department of Energy, and ChatGPT cited a Department of Energy report. Name your source in plain text near the number.
3. Answer the part the official source doesn’t. A regional split, one location, one month or one model year is often missing from the official page.
4. Write the comparison page. If you have pages for A and B, people will ask about A versus B, and the engine will search for both together.
5. Cover the sub-questions on the page that already ranks. Add a section for each related question rather than a new page for each. ChatGPT groups results by domain, so a pile of thin pages collapses into one entry anyway. Google’s guide warns that making a page for every fan-out query “primarily to manipulate rankings or generative AI responses in Google Search violates Google’s scaled content abuse spam policy”.
6. Get indexed in Bing. Submit your sitemap in Bing Webmaster Tools and turn on IndexNow, so one more of the indexes ChatGPT may search has your pages.

To see whether the changes work, ask the same questions again a month later with the same command, and compare which pages get cited. The AI visibility checker page has a worked example.

What fan-out means for keyword research

Fan-out makes long, specific queries more valuable than they look in a keyword tool. A search like “Austin Texas average rent one-bedroom apartment November 2026 current rent data” will never show volume, because no person types it. The engine types it, on behalf of people who asked a shorter question.

I’d stop judging a page only by the head term it targets. A page that answers one specific thing better than anyone else, with the year, the place and the source in plain sight, is the page fan-out searches find.

How this was checked

• I asked ChatGPT a main set of 20 questions, then 7 more for the examples, through DataForSEO’s ChatGPT scraper, US location, English, with web search forced on, in November 2026. It cost about $0.03 in total. DataForSEO reported the model as gpt-6, and each answer ran one to three searches, which matches ChatGPT’s quick answers rather than its thinking mode. Each question was asked once, so a second run could search differently.
• “Cited” means the page was in ChatGPT’s sources for the answer. “Retrieved” means it was in the search results ChatGPT read but not in its sources.
• The Search Console and Bing numbers come from sites I run, through SEO Skill. Bing only lists its top pages, so a page missing from Bing’s list is unknown, not zero.
• Google AI Mode’s searches aren’t visible to site owners, so the AI Mode section relies on Google’s documentation and Moz’s study rather than my own test.