Query fan-out is the technique an AI search engine uses to split the user's question into several searches, run them in parallel and use the results to build a single answer. The user types one sentence. The engine searches for two, ten or, at the extreme, hundreds.
Google explains it like this in its documentation AI features and your website: AI Overviews and AI Mode may use a fan-out technique, "issuing multiple related searches across subtopics and data sources", to develop a response. In the post announcing AI Mode in the UK, on 28 July 2025, the wording is that AI Mode works by "breaking down your question into subtopics and issuing a multitude of queries simultaneously on your behalf".
For anyone doing SEO, the consequence is simple. The page that gets cited doesn't have to rank for the user's question. It has to rank for one of the searches the engine generates.
How AI Mode and AI Overviews use it
Google introduced fan-out at Google I/O in May 2025: "AI Mode uses our query fan-out technique, breaking down your question into subtopics and issuing a multitude of queries simultaneously". Deep Search, the in-depth research mode, uses the same technique and "can issue hundreds of searches".
Two figures help show what this changes for the results:
- according to Ahrefs, on the same sample of queries only 13.7% of AI Mode citations match those in AI Overviews, even though 86% of the cited content is semantically similar;
- according to Semrush (July 2025), in AI Mode the overlap between cited sources and the organic top 10 is around 54% by domain and 35% by URL. In AI Overviews it rises to about 86% and 67%.
The more the engine breaks a question down, the further the sources drift from page one of the original query. A seoClarity analysis of 1,000 US transactional queries found that only 19% of the 12,011 AI Mode citations came from the organic top 20.
How ChatGPT uses it
OpenAI doesn't use the term fan-out, but it describes the same mechanism in its ChatGPT search guide: ChatGPT rewrites the question into more targeted searches. The official example: "CCR8 for cancer" becomes "CCR8 immunotherapy drug development 2025". For local searches it adds the location it has inferred.
The most widely cited study of how ChatGPT searches is by Nectiv, published in Search Engine Land in October 2025 and based on more than 8,500 prompts:
- 31% of prompts trigger at least one web search;
- when it searches, ChatGPT runs an average of 2.17 searches per prompt, with a maximum of 4;
- the average search is 5.48 words long and 77% are longer than 5 words;
- the most frequent modifiers are "reviews", the current year, "features" and "comparison".
A more recent study by Peec AI, covering 20 million fan-out queries between October 2025 and January 2026, saw the average length roughly double, from about 6 to about 12 words, while fan-outs per prompt stayed between 2.3 and 2.8. The UK was one of the five countries in the sample, along with Germany, the US, Singapore and Thailand, and the trend was almost identical in all of them.
Our test on five UK questions
Almost all the published data comes from the US or northern Europe. We first ran a small test in Italian on 27 September 2026, and on 10 October 2026 we repeated it for the UK, using the DataForSEO APIs, which return the searches the model generates. Five questions from sectors we work in, rewritten for British businesses and passed to Gemini 3.6 Flash with web search switched on. We also passed the first question to Claude Sonnet 4.6.
| User question | Searches generated by the model |
|---|---|
| which SEO agency should a manufacturing company in the Midlands choose? (Gemini) | best SEO agency industrial manufacturing West Midlands East Midlands · top SEO agency manufacturing companies Midlands UK |
| same question (Claude) | best SEO agencies for manufacturing companies Midlands UK |
| how much does a 200 kW commercial solar PV system cost in the UK? | 200 kW commercial solar PV system cost UK |
| best outdoor LED walls for corporate events | best outdoor LED screen manufacturers corporate events 2025 2026 · top outdoor LED display brands corporate production |
| difference between SEO and GEO for a B2B company | no search: Gemini answered from its own knowledge |
| software to calculate the carbon footprint of a fashion brand | fashion brand carbon footprint calculation tool · fashion carbon footprint software options |
What we saw:
- Between zero and two searches per question, 8 in total. Gemini ran two for three of the questions, one for the solar question and none at all for SEO versus GEO. Claude ran one. These are simple questions: with more complex ones the numbers go up, as the studies above show. In the Italian run every question produced exactly two.
- An average length of just over 7 words (7.4) across the 8 searches. That's close to Nectiv's 5.48 and well short of the 12 words Peec AI measured in January 2026, probably because our questions were short.
- The first search condenses the question and the second widens it with a synonym or a neighbouring intent: "brands" next to "manufacturers", "software options" next to "calculation tool".
- Gemini added the year on its own: "best outdoor LED screen manufacturers corporate events 2025 2026". It's the same "current year" modifier Nectiv found in ChatGPT's searches.
- The vendors wrote the sources. In the carbon footprint answer, 4 of the 7 cited domains belonged to software companies that then appeared in the list (Carbonfact, Vaayu, GreenStitch and Carbon Trail). In the SEO agency answer, 4 of the 6 domains were the agencies' own service pages.
The result that struck us most comes from the Italian run and involves a client. In the answer about outdoor LED walls, Gemini cited two pages from Macropix, an Italian LED display company we work with: the page on outdoor LED walls and the one on applications for events. The second is devoted to events, which was exactly the intent of the second search the model generated. In the UK run the sources for the same question were "top LED screen manufacturers" lists, mostly published on the blogs of LED makers themselves: the page that matches the generated search gets cited, whoever wrote it.
What changes for keyword research
With fan-out, keyword research is no longer just about finding "the" keyword for a page. You need to work out which variants and sub-questions an engine might generate from the questions your customers ask.
In practice:
- Start from questions, not keywords. Collect the questions customers really ask: emails, phone calls, quote requests, People Also Ask.
- For each question, write down the two or three searches it would turn into. They're usually the condensed question plus a variant: price and cost, best and hire, what is and difference.
- Check who ranks for those searches. If it's always the same sites, those are your competitors in AI answers.
- Cover the sub-questions on the same page or on linked pages, with clear internal links.
Our guide to keyword research explains how to start from the data you already have in Search Console.
How to write for the sub-searches
Google says no special optimisation is needed. Its documentation gives the usual advice: helpful content that can be indexed and shown with a snippet. That said, the tests and studies point to a few habits that help:
- one section per sub-question, with an H2 that states it and the answer in the first two sentences. The engine extracts passages, not whole pages;
- explicit numbers and facts: prices, measurements, dates. In the solar test Gemini built its answer from installer pages that published price ranges, and quoted £135,000 to £200,000 for 200 kW, or £700 to £950 per kW installed;
- natural wording variants: if customers say both "cost" and "price", use both where they make sense;
- comparison and pricing pages, because the modifiers "comparison" and "reviews" are among the most frequent in ChatGPT's searches.
The Princeton paper on GEO reaches similar conclusions: adding statistics, quotations and sources increases visibility in generated answers by up to 40%. We cover this in what GEO is and in the article on AI Overviews ranking factors.
Limits of the method
Fan-out doesn't show up in Search Console, neither for AI Mode nor for AI Overviews. You can only observe it indirectly, with APIs like the one we used or with tools that simulate the searches. The generated searches change from one request to the next and from one model to another: our test is a snapshot of one day, not a rule.
There's also a more basic limit. Searches made through the API aren't identical to those in the app people use, which knows the user's history and location. The test helps you understand the logic, not predict the exact answer a customer will see.
Further reading: AI search engines: which ones exist and how they choose their sources.
FAQs
It's the technique an AI search engine uses to split the user's question into several related searches, run them in parallel and combine the results into a single answer. Google uses it in AI Mode and AI Overviews, and ChatGPT does something similar when it searches the web.
It depends on how complex the question is. Google talks of "a multitude" of searches for AI Mode and hundreds for Deep Search. Studies of ChatGPT measure an average of 2 to 3 searches per prompt. In our test on five UK questions, Gemini and Claude generated between zero and two per question.
No. Search Console shows impressions for pages in AI Overviews and AI Mode, but not the internal searches the model generates. You can only observe them with APIs or monitoring tools that expose them.
Google says no special optimisation is needed beyond SEO. In practice it helps to cover the sub-questions a customer would ask, with clear sections, explicit data and internal links between related pages.
