When you type one question into Google's AI Mode or ChatGPT, the AI often does not run just one search. It may split your question into several smaller questions, search for each one, and then combine what it finds into a single answer. This process is called query fan-out.
This matters for anyone who publishes content. Your page might never rank for the question a person typed. But it could still be found and cited because it answers one of the hidden smaller searches. This guide explains what is known about fan-out and how to prepare your content for it.
What query fan-out means, in plain words
A "query" is just a search. "Fan-out" means spreading out from one point into many, like a hand fan. Here is the basic flow:
- A person asks a question, often a long and detailed one.
- The AI works out what the person really needs to know. It breaks the question into smaller parts, sometimes called subtopics.
- It runs a separate search for each part. These can happen at the same time.
- It reads the pages that come back from all of those searches.
- It writes one answer and links to some of the pages it used. Those links are called citations.
Google introduced AI Mode in March 2025 and described this exact approach. Its launch post said AI Mode uses a "query fan-out" technique, "issuing multiple related searches concurrently across subtopics and multiple data sources and then brings those results together to provide an easy-to-understand response."1 "Concurrently" just means at the same time.
Two months later, at its I/O developer event in May 2025, Google repeated the idea in simpler terms: "Under the hood, AI Mode uses our query fan-out technique, breaking down your question into subtopics and issuing a multitude of queries simultaneously on your behalf."2
What Google has said about fan-out so far
Google has described fan-out several times since early 2025, including in its official help page for website owners.
| Date | Where | What Google said |
|---|---|---|
| March 5, 2025 | AI Mode launch post (Robby Stein, VP of Product, Google Search) | AI Mode issues "multiple related searches concurrently across subtopics and multiple data sources."1 |
| May 20, 2025 | Google I/O post (Elizabeth Reid, VP, Head of Search) | AI Mode breaks a question "into subtopics." A new feature, Deep Search, "uses the same query fan-out technique but taken to the next level" and "can issue hundreds of searches."2 |
| July 30, 2025 | Interview with Robby Stein, reported by Search Engine Journal | Stein described the AI using Google Search as a tool. Using the example of things to do in Nashville with a group, he said it may think of questions like restaurants, bars and things to do with kids, "and it'll start Googling basically."3 |
| November 18, 2025 | Gemini 3 in Search post (Elizabeth Reid) | "Google Search's query fan-out technique is getting a major upgrade." It can now "perform even more searches to uncover relevant web content" and "find new content that it may have previously missed."4 |
| December 10, 2025 (last update) | Google Search Central, "AI features and your website" | Both AI Overviews and AI Mode "may use a 'query fan-out' technique," issuing "multiple related searches across subtopics and data sources" to "develop a response."5 |
Two things stand out from Google's own words.
- Fan-out is not only an AI Mode feature. Google's documentation says AI Overviews "may" use it too.5 AI Overviews are the AI summaries that appear at the top of normal Google results.
- It is meant to find a wider range of pages. Google's help page says that while answers are being made, its models find more supporting pages, so it can "display a wider and more diverse set of helpful links" than a classic web search.5 So fan-out pulls in pages that might not show up for the original question.
Google's help page also says no special trick is needed: "There are no additional requirements to appear in AI Overviews or AI Mode, nor other special optimizations necessary."5 To be shown as a supporting link, a page "must be indexed and eligible to be shown in Google Search with a snippet."5 Indexed means Google has stored the page. A snippet is the short text preview under a search result. So the basics of normal SEO (search engine optimization) still decide whether your page can even be picked up by a fan-out search.
Why AI systems split one question into many
The idea behind fan-out is older than AI Mode. One search often misses useful pages. A person might use different words from the page, or ask a question with several parts that no single page covers.
Google has filed patent applications on this kind of system. A patent does not prove a company uses the idea in a live product. Google's application called "Search with stateful chat," published in August 2024, describes a chat-style search system where "Synthetic queries may be generated" from the AI model's output.6 A synthetic query is a search the system writes itself rather than one the user typed. The document lists extra queries such as "alternative query suggestions, supplemental queries, rewritten versions of the user's query," and "drill down" queries.6 It also describes "implied queries" that can be "automatically generated" from context.6
Researchers have tested the same pattern. A 2024 paper on a method called RAG-Fusion looked at a chatbot for Infineon, a chip maker, that needed to give engineers and customers product information fast.7 RAG stands for retrieval-augmented generation. It means an AI first looks up documents, then writes an answer based on them. RAG-Fusion works "by generating multiple queries," then ranking and combining the results.7 The author found it could "provide accurate and comprehensive answers," because the extra queries looked at the original question "from various perspectives."7
The same paper also found a weakness. It said "some answers strayed off topic when the generated queries' relevance to the original query is insufficient."7 So fan-out is not perfect. If a sub-query drifts from what the person meant, the answer can drift too.
What 15 million ChatGPT fan-out queries show
Google does not publish its actual sub-queries. With ChatGPT, researchers can sometimes capture the searches it runs. In October 2026, the SEO software company Ahrefs published a large study of these searches.8
Ahrefs analyzed 66.6 million free, logged-out responses from ChatGPT's GPT-5.6 Luna model, and the 15 million fan-out queries behind them. The data came from August and September 2026.8 Ahrefs used logged-out sessions because "Logging out also strips out chat history and memory."8
Here are the main numbers:
| Finding | Number | What it means |
|---|---|---|
| Responses that triggered a web search | 14.4% | Most answers in this sample were written without searching the web at all.8 |
| Fan-out queries per response that searched | About 1.6 on average | 15.0 million queries came from 9.6 million searches. On average, that is far from dozens of searches per answer.8 |
| Fan-out queries using a "site:" operator | 6.8% | "site:" tells a search engine to look inside one website only. ChatGPT sometimes went straight to a specific site.8 |
| Most common "site:" searches aimed at a brand's own site | 75.9% | Among the most-searched sites, ChatGPT's "site:" searches usually pointed at a company's own website.8 |
| English fan-out queries containing "official" | 10.1% (August) to 13.1% (September) | ChatGPT asked for official sources more often in September.8 |
| Fan-out queries containing "2026" | 17.6% (August) to 22.9% (September) | The current year was the most common of the terms Ahrefs tracked, which suggests ChatGPT is looking for fresh pages.8 |
| Growth in "price" and "review" fan-outs | Up to 47% in a month | "Price" rose from 6.04% to 8.88% of fan-outs (a 47% relative rise), and "reviews" grew by about a third.8 |
Ahrefs also ran a separate test of 20,000 US prompts. Factual questions about "pricing, product features or company details" cited at least one official source 84.4% of the time. For comparison questions, like "A vs B" or "which is cheaper," that share "dropped to 68.3%."8 So for simple facts, ChatGPT tends to go to the company. For comparisons, it more often goes elsewhere.
The study also looked at the words ChatGPT added to its searches. Ahrefs wrote: "The terms ChatGPT adds to its fan-out queries hint at the page types it wants to retrieve."8 In one example, it filtered US ChatGPT answers that mentioned "running shoes." Among the 1,474 of those answers that included fan-out queries, "best" appeared 206 times, "2026" appeared 160 times, "review" 117 times, "vs" 105 times and "official" only 65 times.8 That is one topic, not the whole dataset. But for this shopping topic, ChatGPT mostly searched for recent reviews and comparisons.
A note on the data: Ahrefs says these numbers "are a snapshot of one model between August and September 2026," and that "OpenAI changes how ChatGPT retrieves information constantly."8 It also says free, logged-out sessions "search less than paid or reasoning models," and that "The prompts in our study are broad."8 So treat the percentages as one view of one model. Google, by comparison, has said Deep Search "can issue hundreds of searches,"2 which is a very different scale from about 1.6.
What fan-out means for your content
You compete on the sub-questions, not only the main one
A long buyer question rarely matches any page word for word. So the AI searches for shorter, more specific pieces. Your page gets its chance when it is a strong result for one of those pieces. That means one "ultimate guide" may matter less than a set of clear pages that each answer one sub-question well.
The type of page matters
The words ChatGPT adds to its searches, like "best," "vs," "review," "price" and "official," each point to a kind of page.8 "Best" points to roundup lists, "vs" to comparison pages, "price" to pricing pages and "official" to the company's own site. If a buyer's question fans out into a "price" search and you have no clear pricing page, the AI may take the price from someone else's page instead.
Your own site should handle the facts
The Ahrefs data suggests AI often goes to a brand's own site for factual details like pricing and features.8 So the boring pages matter: pricing, features, integrations, help docs and company details. Keep them easy to find, current and in plain text, not hidden in images or PDFs.
Other sites often handle the comparisons
Comparison questions cited official sources less often (68.3% versus 84.4% for factual questions).8 So when a fan-out search asks "A vs B" or "best X for Y," the AI may rely on review sites, independent blogs and roundups. Publish your own comparison pages, but also make sure you are listed accurately on the third-party pages that rank for those searches.
Freshness shows up in the search itself
The current year was the most common of the terms Ahrefs tracked in ChatGPT's fan-out queries.8 When the AI adds "2026" to a search, pages that clearly show they are current have a better chance of matching. That does not mean stuffing the year into titles. It means truly updating key pages and showing a real "last updated" date.
An illustrative worked example: one buyer question, ten hidden searches
ILLUSTRATIVE ONLY: The sub-queries below are not real logged output from Google, ChatGPT or any other AI tool. They are a plausible illustration, written by Ripplix Research, based on the patterns Google has described and the query terms Ahrefs reported ("best," "vs," "review," "price," "official," the current year and "site:" searches). Real fan-out queries vary by tool, by model and from one run to the next.
Let's take one realistic question a buyer might type:
"What's the best project management software for a 20-person creative agency that bills clients by the hour?"
No single page is likely to match all of that. So an AI system might break it into smaller searches like these.
| # | Illustrative sub-query | What the AI is checking | Page type that answers it | Usually owned by |
|---|---|---|---|---|
| 1 | best project management software for agencies 2026 | Which tools are usually recommended right now | "Best tools" roundup or list article | Review sites, publishers, sometimes vendors |
| 2 | project management software with time tracking and client billing | Which tools have the key feature the buyer needs | Feature page for time tracking and invoicing | Vendor |
| 3 | project management software pricing per user small team | What a 20-seat team would pay | Pricing page with clear per-user prices | Vendor |
| 4 | Tool A vs Tool B for agencies | How two shortlisted tools compare | Comparison page | Vendors and review sites |
| 5 | Tool A reviews agency | What real users say | Review platform profile, case studies | Review sites, vendor (case studies) |
| 6 | site:tool-a.com pricing | The official price, straight from the source | Official pricing page | Vendor |
| 7 | project management client portal guest access | Whether clients can view work without paying for a seat | Feature page or help center article | Vendor |
| 8 | Tool A integrations Slack Google Drive | Whether it fits the tools the agency already uses | Integrations directory | Vendor |
| 9 | agency resource planning capacity scheduling software | Whether it can plan staff time across projects | Use-case or industry page ("for agencies") | Vendor |
| 10 | how to choose project management software for a creative agency | What criteria matter for this kind of business | Buying guide or how-to article | Publishers, vendors, consultants |
Only sub-queries 1 and 10 sound like the original question. The other eight are narrow, practical checks. Six of the ten point mainly at pages a software company would own: feature pages, pricing, integrations, use-case pages and help docs. The rest point at roundups and review sites, where the company needs to be present and described accurately.
Now picture two imaginary vendors. Vendor X has a strong homepage and one long "ultimate guide" blog post. Vendor Y has the same post, plus a pricing page, a time tracking and invoicing page, a "for agencies" page, an integrations list and a help article on client access. Across these ten searches, Vendor Y has a chance in six or more. Vendor X has a chance in one or two. The gap is about coverage, not writing quality.
A reusable content-gap checklist
To reuse the example, take one key buyer question and write down eight to twelve likely sub-queries. Then check each type below against your own pages.
| Sub-query type | Typical words the AI might add | Page you need | Check yourself |
|---|---|---|---|
| Category "best" search | best, top, 2026 | A presence on the main roundups, plus your own honest "best for" page | Are you listed on the top-ranking roundups for your category, with correct details? |
| Feature check | with, supports, feature names | One clear page per key feature | Can a stranger confirm each key feature in one click, in plain text? |
| Price check | price, pricing, cost, per user | A public pricing page | Are real numbers on the page, or only "contact sales"? |
| Official facts | official, site:yourbrand.com | Pricing, features, company facts on your own domain | Would a "site:" search of your domain return the right page for each basic fact? |
| Comparison | vs, alternative, compared to | Fair comparison pages against your main rivals | Do you have a page for each "you vs rival" search buyers run? |
| Reviews and proof | review, reviews, case study | Review platform profiles, named case studies | Are your review profiles active and recent? Do case studies name the customer type? |
| Fit for a segment | for agencies, for small teams, for [industry] | Use-case or industry pages | Do you have a page for each segment you want to win? |
| Compatibility | integrations, works with, tool names | Integrations directory with one entry per tool | Is each integration on its own crawlable page or section? |
| How-to and setup | how to, set up, can I | Help center articles that are public and indexable | Is your help center open to search engines, or behind a login? |
| Freshness | 2026, latest, new | Pages with real, visible update dates | When did each key page last change in a meaningful way? |
Each "no" is a likely content gap. Start with the gaps closest to a buying decision: pricing, features and comparisons.
How to find your likely fan-out queries
You usually cannot see the exact sub-queries an AI ran. But you can get close with free or common tools.
- Start with real buyer questions. Collect the questions people actually ask your sales and support teams. Ripplix's guide to finding the buyer questions people ask AI covers this step in more detail.
- Check Google's "People also ask" box. Search your main question on Google. The "People also ask" box shows related questions. These are not fan-out queries, but they make good sub-query guesses.
- Ask the question in AI Mode and read the sources. Each page it links to hints at a sub-question the AI needed answered. Then try the follow-up questions AI Mode suggests. Google's launch post says people can "ask follow-ups to dig deeper."1
- Look at long queries in Google Search Console. Google says traffic from AI Overviews and AI Mode is "included in the overall search traffic in Search Console" and reported in the Performance report under the "Web" search type.5 It is not split out separately, and Google does not label which searches were fan-out searches. Still, filtering for long, question-style queries can show the detailed questions people bring to your pages.
- Use Bing Webmaster Tools' AI Performance report. This is the closest thing to real fan-out data that a site owner can see today. Microsoft launched it in public preview in February 2026. It covers "Microsoft Copilot, AI-generated summaries in Bing, and select partner integrations."9 Its "Grounding queries" view "Shows the key phrases the AI used when retrieving content that was referenced in AI-generated answers."9 Microsoft notes the data "represents a sample of overall citation activity."9 In June 2026, Bing added Topics, which "group related grounding queries into broader thematic clusters," and Citation Share, which shows "how much of the citation space your site receives for a specific grounding query."10
- Write the sub-queries down and map them. Use the checklist above. For each likely sub-query, note which page on your site answers it, or write "none."
- Re-check over time. Ahrefs saw the mix of fan-out terms shift within a single month,8 and Google has upgraded its fan-out system at least once.4 Repeat this exercise every few months, not once.
What nobody outside the AI companies knows yet
A lot of advice talks about fan-out as if every step were known. It isn't.
- Site owners mostly cannot see the sub-queries. Google describes fan-out, but its help page for site owners offers no report of the actual sub-queries.5 Bing's grounding queries are the main exception, and even those are a sample.9
- The same question may not fan out the same way twice. There is no public promise that one question always produces the same sub-queries. The ChatGPT terms Ahrefs tracked shifted from August to September,8 and Google's fan-out changed when Gemini 3 arrived.4
- Different tools fan out very differently. Free, logged-out ChatGPT ran about 1.6 searches per searching answer in Ahrefs' sample.8 Google says Deep Search can run hundreds.2 Advice built on one tool may not carry over to another.
- Being found is not the same as being cited. A fan-out search can retrieve your page, and the AI can still leave it out of the final answer. The research on RAG-Fusion shows the step that combines results can also go wrong when sub-queries drift off topic.7
None of this makes fan-out too uncertain to act on. The safest strategy is coverage: clear pages for the sub-questions buyers are very likely to need answered.
See which buyer questions your pages actually show up for
Fan-out means your brand can win or lose an AI answer on a sub-question you never tracked. Ripplix runs the real buyer questions in your category across AI platforms and shows where you are mentioned, where you are cited and which pages AI picks instead of yours. That makes your content gaps visible, question by question.
Get your free AI Visibility Report →
Sources:
- Google: Expanding AI Overviews and introducing AI Mode (March 5, 2025)
- Google: AI in Search: Going beyond information to intelligence (May 20, 2025)
- Search Engine Journal: Query Fan-Out Technique in AI Mode: New Details From Google (July 30, 2025)
- Google: Google Search with Gemini 3: Our most intelligent search yet (November 18, 2025)
- Google Search Central: AI features and your website (last updated December 10, 2025)
- Google Patents: US20240289407A1, Search with stateful chat (published August 29, 2024)
- arXiv: RAG-Fusion: a New Take on Retrieval-Augmented Generation (January 31, 2024)
- Ahrefs: How ChatGPT Does Query Fan-Out (15 Million Queries Studied) (October 12, 2026)
- Microsoft Bing Blogs: Introducing AI Performance in Bing Webmaster Tools Public Preview (February 10, 2026)
- Microsoft Bing Blogs: New AI Visibility Insights in Bing Webmaster Tools: Intents, Topics, Citation Share, Compare (June 16, 2026)
