Definition
Query fan-out is a technique used in AI search: the system breaks a question into subtopics, runs several searches of its own and combines the results into an answer with sources. Google describes it for AI Mode and AI Overviews, where these searches run in parallel.
Key takeaways
- Query fan-out means an AI system runs several searches of its own for your question. Google describes it for AI Mode and AI Overviews.
- The number of fan-out queries varies widely. Many ChatGPT answers contain none, while Google’s Deep Search can run hundreds of searches.
- AI systems find pages through fan-out queries. So answer your customers’ sub-questions, each one clearly and with verifiable facts.
- Google says a page for every phrasing doesn’t work in the long run. If it’s mainly meant to manipulate AI answers, it counts as spam.
- Site owners rarely see fan-out queries. Bing Webmaster Tools shows a sample of the phrases the AI used to retrieve the pages it cited.
How does query fan-out work?
With query fan-out, an AI system breaks your question into subtopics and comes up with its own sub-searches for them, called fan-out queries. It reviews the results and writes an answer with links to its sources. Google runs these searches in parallel. Other systems can also search one after another and build on the first results.
Google explains the technique with an example query: “how to fix a lawn that’s full of weeds.” Its fan-out queries might include “best herbicides for lawns,” “remove weeds without chemicals” and “how to prevent weeds in lawn” (Google Search Central, as of Jul 10, 2026). Each fan-out query asks for one part of the answer, and none of them repeats the question.
Based on the documentation from Google and Anthropic, the process has four steps (Google AI Studio, Anthropic):
- Decide: The system checks whether a search would improve the answer. For established knowledge, it often answers without searching.
- Formulate searches: It derives one or more search queries from the question.
- Retrieve: Each fan-out query returns results from a search index, the catalog of pages a search engine knows about. At Google, the core Search ranking systems select the results (Google Search Central, as of Jul 10, 2026).
- Answer: The model evaluates the results and links to the pages the answer is based on.
According to Google, this lets AI answers show a wider and more diverse set of links than a classic web search (Google Search Central, as of Dec 10, 2025). One study is consistent with this. On average, 53% of the domains that AI Overviews drew on were not among the top 10 organic results for the original query (Kirsten et al., May 31, 2026). The tests ran in September 2025, with queries in English. The study didn’t look at whether fan-out queries are the reason.
Which AI systems use query fan-out?
Google uses the term query fan-out for AI Mode, AI Overviews and Deep Search. OpenAI describes the same principle for ChatGPT, Microsoft for Copilot, Anthropic for Claude and Perplexity for Pro Search: the system derives its own search queries from your question and searches more than once if needed.
- Google’s AI Mode breaks a question into subtopics and runs a large number of searches at the same time (Google, May 20, 2025). With the Gemini 3 model, Google says, it can run even more searches (Google, Nov 18, 2025).
- AI Overviews can use the technique as well. According to Google, the two features may use different models and therefore show different links (Google Search Central, as of Dec 10, 2025). Our article on AI Overviews and AI Mode explains how Google picks its sources there.
- Deep Search in AI Mode uses the same technique on a larger scale and can run hundreds of searches (Google, May 20, 2025).
- ChatGPT typically rewrites your question into one or more targeted search queries when it uses search partners. After reviewing the first results, it can send more specific follow-up queries (OpenAI, October 2026).
- Microsoft Copilot, in its business version, derives short search queries for Bing from the question. In Microsoft’s example, a request to summarize a company’s finances and business strategy turns into the searches “Fabrikam strategy” and “Fabrikam financials” (Microsoft, Oct 29, 2026).
- Claude searches when a request depends on current or changing information, according to Anthropic’s developer documentation. It can search several times within one request (Anthropic, accessed Nov 7, 2026), and later searches can build on earlier results (Anthropic, May 7, 2025).
- Perplexity describes follow-up searches for Pro Search that build on previous results (Perplexity, Jul 2, 2024).
Our guide to generative engine optimization (GEO) compares where each system gets its results and how you can control access.
How many fan-out queries does an AI system run?
There’s no fixed number. It depends on the system, the type of access and the kind of question. Many answers contain no fan-out query at all. Google loosely describes AI Mode as doing a dozen searches and says Deep Search can run hundreds.
The numbers in the “Source” column refer to the list of sources at the end of this article.
| System | Figure | Type and scope | Source |
|---|---|---|---|
| Google AI Mode | about a dozen searches in the time one search takes | figure of speech by a Google Search engineering director, not a measurement | 13 |
| Google Deep Search | hundreds of searches possible | Google’s own statement | 8 |
| Claude | single facts usually 1–3 searches; comparisons and research on several companies or products 10 or more | Anthropic’s figure for its developer API | 3 |
| ChatGPT, free and logged out | 14.4% of answers contained fan-out queries, about 1.6 each on average | Ahrefs measurement: 66.6 million answers from August and September 2026 | 14 |
| Perplexity, ChatGPT, Grok | 1.4, 2.1 and 6.8 fan-out queries per prompt | Peec AI measurement: 5 million fan-out queries from Apr 1–21, 2026 | 15 |
The figures depend on how each measurement was set up. Ahrefs analyzed free, logged-out sessions, which it says search less than paid plans and reasoning models. Other studies found far more fan-out queries per prompt, often using commercial prompts such as “best X” (Ahrefs, Oct 12, 2026). Ahrefs also points out that OpenAI constantly changes how ChatGPT searches. Ahrefs and Peec AI sell tools for measuring AI visibility, and the figures come from their own data.
What this means for your pages: in ChatGPT, there are often only one or two fan-out queries. If your page matches neither, ChatGPT never sees it, Ahrefs concludes. Comparisons and research involve more fan-out queries, so there are more ways to reach your page.
What does query fan-out mean for your content?
When AI systems search the web, they find your pages through fan-out queries. So answer every sub-question customers have about your services, clearly and with verifiable facts such as price range, process and service area. You don’t need a page for every phrasing. Google treats such pages as spam if they’re mainly meant to manipulate rankings or AI answers.
For small businesses, we think query fan-out matters most for questions where customers compare providers or want to clarify several things at once. Google says AI Mode is particularly helpful for complex comparisons like these (Google Search Central, as of Dec 10, 2025).
Which sub-questions come up about what you offer?
Fan-out queries often ask for current and official information. In ChatGPT’s English-language fan-out queries in September 2026, 22.9% contained the year 2026, 13.1% the word “official” and 8.9% the word “price” (Ahrefs, Oct 12, 2026). Start by collecting the questions customers ask before they hire you, for example about price, process, timeline and service area. Step 4 of our how-to, How to get your business recommended by ChatGPT, shows how to answer them.
Which facts belong on your website?
Everything customers ask about before they hire you belongs on your own website, in plain text. ChatGPT looks there on purpose: 6.8% of its fan-out queries used the site: operator, which searches a single website. Among the roughly 160 websites searched this way most often, three in four of these searches went to a brand’s own website (Ahrefs, Oct 12, 2026).
ChatGPT can also add a location to the search on its own, based on the user’s IP address. If you ask “What are some good restaurants near me?” from the San Francisco area, ChatGPT may search for “top restaurants San Francisco” (OpenAI, October 2026). We think a page that doesn’t mention a location is a weaker match for fan-out queries like these. Mention your location and service area in the text.
Does every sub-question need its own page?
No. Google says AI systems understand synonyms and general meaning, so you don’t have to cover every variation of a search (Google Search Central, as of Jul 10, 2026). A study shows how systems create variations themselves. Asked about the goals of a US commission, Gemini ran two searches: one with “goals” and one with “objectives” (Kirsten et al., May 31, 2026). Separate pages for variations or fan-out queries break Google’s spam policy on scaled content abuse if their main purpose is to manipulate rankings or AI answers.
One page can answer several sub-questions. Google says its systems can understand multiple topics on one page and show users the relevant part. You don’t need to chop your content into small pieces for that. Structure it for your readers, with headings and with sections that start with the answer.
On the technical side, a page only appears in AI Mode and AI Overviews if it’s indexed and eligible to be shown in Search with a snippet (Google Search Central, as of Dec 10, 2025). That means crawlers have to be allowed to read it. Our article on AI crawlers and robots.txt explains which ones.
Can you see which fan-out queries an AI system runs?
Only partly. Google doesn’t describe any report on fan-out queries in Google Search Console. Bing Webmaster Tools shows a sample of the phrases the AI used to retrieve pages it cited. Developers can see the search queries in the responses of Google’s and Anthropic’s APIs.
- Google Search Console: AI Mode data counts toward the question the user asked, and a follow-up question counts as a new query (Google Search Console Help, accessed Nov 7, 2026). The generative AI performance report breaks data down by page, country, date and device (Google Search Console Help, Aug 31, 2026).
- Bing Webmaster Tools: Since February 10, 2026, the AI Performance report has shown what Microsoft calls grounding queries. These are the phrases the AI used to retrieve content that was then cited in Microsoft Copilot, in AI-generated summaries in Bing or in select partner integrations. The data is a sample (Microsoft Bing, Feb 10, 2026). Since June 2026, Microsoft has also grouped them by intent and topic, as a preview (Microsoft Bing, Jun 16, 2026). Our how-to on measuring AI visibility shows how to analyze this data.
- Microsoft Copilot, business version: In Copilot Chat, the sources of an answer list the exact search queries sent to Bing. They stay visible for 24 hours (Microsoft, Oct 29, 2026).
- Developer APIs: Google and Anthropic return the executed search queries together with the answer (Google AI Studio; Anthropic, accessed Nov 7, 2026). That isn’t AI Mode, but it shows how the models phrase search queries.
- Third-party tools: Some record fan-out queries from ChatGPT and Perplexity, for example (Ahrefs, Oct 12, 2026). Google, however, states: “No third-party tool has access to our internal ranking or AI systems” (Google Search Central, as of Jul 10, 2026). If a tool shows fan-out queries from AI Mode, ask where they come from.
Some explanations online derive details of AI Mode from patents. One Google patent application, for example, describes search queries that a large language model (LLM) generates from the user’s query, which it calls “synthetic queries” (Google LLC, Aug 29, 2024). A patent application doesn’t prove whether or how Google uses such a method in AI Mode.
What is query fan-out called in German?
There’s no fixed German term. In German, Google uses “Query-Fan-Out-Verfahren,” “Abfrage-Fan-out-Verfahren” and “parallele Anfrageverarbeitung,” which means parallel query processing. If you work with German teams or read German sources, you may come across all three.
The terms come from Google’s German blog post on the launch of AI Mode (Google, Oct 8, 2025). They also appear in the German versions of two Google Search Central pages (Google Search Central, as of Dec 31, 2025; as of Jul 15, 2026). There, the individual searches are called “parallele Anfragen” (parallel queries) or “ähnliche Suchanfragen” (related searches). Our German pages use the English term and call the individual searches “Teilsuchen.”
The name comes from the verb fan out, which means to move in different directions from a single point (Cambridge Dictionary). Here, one question spreads out into several searches.
How does query fan-out differ from query expansion and grounding?
Query fan-out creates several additional search queries for subtopics of a question. Query expansion, by contrast, adds more terms to a single query. Grounding means retrieving the pages that an AI answer is based on. Deep Search applies query fan-out on a much larger scale.
The numbers in the “Source” column refer to the list of sources at the end of this article.
| Term | What happens | Search queries | Source |
|---|---|---|---|
| Query fan-out | The model breaks the question into subtopics and runs additional related searches for them. At Google, these run in parallel. | several | 1 |
| Query expansion | A query gets additional terms so the search system finds more relevant documents. | one | 22 |
| Query rewriting | The system turns the question into a suitable search query, for example by adding a location or a year. | one or more | 9 |
| Grounding (at Google, also called retrieval-augmented generation or RAG) | The system retrieves current pages from a search index and bases its answer on them. | depends on the system | 1 |
| Deep Search | In-depth research in AI Mode using the same technique. The result is a report with sources. | up to hundreds | 8 |
Our article What is AI grounding? explains what grounding is and what you can control. What is RAG? shows how the technique works.
The idea isn’t new. Back in 2022, researchers showed that a language model answers compositional questions more accurately when it breaks them into follow-up questions and lets a search engine answer those (Press et al., 2023). Google introduced the technique with AI Mode in March 2025 (Google, Mar 5, 2025).
What’s different about query fan-out in Germany?
Google began rolling out AI Mode in Germany on October 8, 2025, and describes it there with the same query fan-out technique. For German-language prompts, ChatGPT often searches in English as well. That means German companies also compete with English-language sources in fan-out queries.
Peec AI analyzed more than 20 million fan-out queries, including German prompts from users in Germany. In nearly 78% of non-English cases, ChatGPT added research in English. As of February 2026, 43% of all fan-out queries for such prompts ran in English (Peec AI, February 2026). According to Peec AI, Google’s products stay closer to the language of the question.
We think a good German-language page is still the foundation. If you sell into Germany with English content, English fan-out queries give that content a chance to be retrieved for German prompts as well. And if your field has many English sources, also check which of them show up in AI answers on your topics. Our page SEO and GEO for Germany explains how we help international companies reach German buyers.
From our work: how we split one customer question across our pages
When we planned datafalke.de, we gave every keyword we want to cover exactly one target page in each language. Variants of the same question are secondary keywords of that page, and no variant gets a page of its own. This page, for example, has the main keyword “what is query fan out” and secondary keywords such as “query fan out” and “query fan out seo.” Its German version has the main keyword “was ist query fan out.”
The table takes a sample customer question and shows how its sub-questions map to our pages and their German main keywords: “Who in the Ruhr area builds a website that gets my business into AI answers too, and what does it cost?” The sub-questions are our assumption, not a record of real fan-out queries.
| Sub-question | German main keyword | Page on datafalke.de |
|---|---|---|
| Are there providers nearby? | webdesign bottrop | Bottrop location page, in German only |
| What does a website cost? | was kostet eine website | Website cost in Germany |
| How does AI search work? | generative engine optimization | What is GEO? |
| What can I do myself? | in chatgpt gefunden werden (get found in ChatGPT) | Get recommended by ChatGPT |
| What does the project include? | website erstellen lassen (have a website built) | Website development |
Here’s how to do the same for your business:
- Write down the question customers ask most often before they hire you.
- Break it down into sub-questions, for example about the provider, price, process and results.
- For each sub-question, check which of your pages answers it in the first sentence of a section. If an answer is missing, add a section. A new page is only worth it for a question with its own intent.
We can’t prove yet that AI systems find our pages through fan-out queries like these. We’ll only publish our own measurements together with the measurement log. We plan to review the grounding queries in Bing Webmaster Tools every month. If a question shows up there twice, it goes on our topic list. That way, real fan-out queries show which sub-question our website still doesn’t answer.
How we’ll measure AI visibility