Reading time: 11 minutes
Fan-out queries are the extra searches Google runs when it splits one AI Mode search into several sub-queries. In SEO, the map of fan-out queries shows what a page must answer. Google does not publish the list, so you have to estimate it. This guide shows the sources, the limits of the tools and how to use the map.
In short
- Google confirms that AI Mode splits the question into subtopics and runs many searches at once, but it does not show which ones.
- The map comes from four clues: autocomplete, “People also ask”, Search Console and a reading of the AI Mode answers.
- Tools that simulate the fan-out use an AI model to guess likely queries. They help with ideas, but they are not a record of what Google did.
- Each sub-question becomes an H2 or, when it has its own intent, a new page. Then it becomes a fixed tracking prompt.
What is query fan-out?
Query fan-out is the process in which AI turns one question into several related searches and runs them at the same time. Then it merges the results into one answer. So the cited page may have answered one of the sub-questions, not the search the person typed. For this reason, query fan-out changed content planning in SEO.
Google itself uses the term. Its generative AI guide for Search defines query fan-out as a set of related queries that the model generates and runs concurrently.
The effect on how Google picks sources, with the official lawn example, is in the guide to Google AI Mode and AI Overviews. To see where this fits in an AI visibility strategy, start with the guide on what GEO is. Here, the focus is the method: how to find the sub-questions and what to do with them.

How does Google’s query fan-out split a question?
According to Google’s announcements, AI Mode breaks the question into subtopics and runs many searches at once, across several data sources. Then it merges the results into one answer with links. Google describes the process, but it does not publish the list of sub-queries for each search, not even in Search Console.
The AI Mode announcement, from March 2025, says the technique runs related searches across subtopics and sources such as the Knowledge Graph and shopping data. In Google’s example, one question compares the sleep tracking of a ring, a watch and a mat. Then the model makes a plan, searches and adjusts the plan based on what it finds.
At Google I/O in May 2025, the company added Deep Search, which uses the same technique “taken to the next level.” According to Google, it can run hundreds of searches to build a report with sources. In addition, AI Overviews can also use query fan-out, according to the documentation.
| Point | What Google documents |
|---|---|
| Split | The question becomes subtopics and simultaneous searches |
| Sources | Web, Knowledge Graph and shopping data |
| Scale | Deep Search can run hundreds of searches |
| Reach | AI Mode and AI Overviews may use the technique |
| List of sub-queries | Not published |
In practice, nobody outside Google sees the list. Semrush, which sells SEO tools, wrote that there is no way to know which queries Google will use. According to the company, they are “generated dynamically and can vary with each search.” So in SEO, any map of Google fan-out queries is always an estimate.
How do you map fan-out queries by hand?
To map fan-out queries by hand, you gather clues about what people ask around the main search and group those questions by intent. Because Google does not show the sub-queries, the map is an estimate. The four clues below are free. For SEO, they cover fan-out queries from both sides: what the audience types and what the AI answers.
Autocomplete and “People also ask”
According to Google’s help page on autocomplete, predictions reflect real searches done on Google. However, the system also considers the language, the location, trending interest and the past searches of the person typing. So use an incognito window and set the location and language of your audience, such as the US and English. In SEO, this is the simplest clue for estimating Google fan-out queries, because it shows what people actually type.
Type the main search and then add modifiers: price, “worth it”, “how it works”, “for” and competitor names. Then open the “People also ask” box and note the questions. For example, clicking one of them usually shows others, which widens the map.
Search Console: the questions that already bring visits
In the Performance report, filter the queries that start with “how”, “which” or “why” to find the questions that already bring people to the site. The regex filters in Search Console do this in one step. This way, each long question becomes a candidate sub-question.
According to the Search Console help page, a follow-up question in AI Mode counts as a new query. In other words, conversation questions also enter the report. Still, the report does not mark which queries came from AI. What you can and can’t see is in the guide on how to read AI Mode in Search Console. In addition, the help page does not say it shows the sub-queries the model generates. Even so, the real questions in the report are a good base for the map of Google fan-out queries.
The answers in AI Mode itself
Run the main search in AI Mode and read the answer as a table of contents. Usually, each block covers one subtopic, and the sources cited in each block show who already answers that part. This is a clue, not a record of the searches. Therefore, repeat the question on different days, since the answer changes. This way, reading the answer completes the map of Google fan-out queries without replacing the other clues.
However, do not ask the model for the list of searches it ran. In a simulated test, with search turned off, Dan Petrovic of DEJAN asked Gemini 3, through the API, for the queries used in an answer. The model filled the list with invented searches and later said it had rebuilt them from its own answer. Petrovic maintains a fan-out tool, and it was a single test. For AI Overviews, see also how to tell whether a keyword has an AI Overview.
Bing grounding queries
Outside Google, Bing shows something similar. The AI Performance report in Bing Webmaster Tools lists grounding queries, that is, the phrases the AI used to retrieve the site content it ended up citing. But they apply to Copilot and partners, not to AI Mode. See how to read the AI Performance report in Bing Webmaster Tools.
Example: health insurance for the self-employed
Consider the search “health insurance for self employed”, which matters to a US broker who wants to appear in AI Mode. Suppose autocomplete and “People also ask” point to themes such as cost, family coverage, small business owners and state names. These themes are hypothetical, chosen to show the method; they are not Google’s actual suggestions, which change. The table groups them as if they were the clues.
| Likely sub-question | Clue | Becomes |
|---|---|---|
| How much does it cost? | Cost | H2 |
| Can one plan cover a family? | Family | H2 |
| Is it different for small business owners? | Small business owners | H2 |
| Which plans are sold in my state? | State names | State pages |
So for the broker’s SEO, the map of Google fan-out queries becomes three H2s and a set of state pages.
Which tools simulate query fan-out, and how reliable are they?
Tools that simulate query fan-out use an AI model to generate the likely sub-queries of a search. None of them sees the list AI Mode used. In fact, Google itself warns that no third-party tool has access to its internal ranking or AI systems. Therefore, in SEO, use the simulator for ideas about fan-out queries and check each one against the real clues.
| Tool | What the source says | Limit |
|---|---|---|
| Otterly (paid) | Queries that ChatGPT, AI Mode and AI Overviews may generate | The tool’s estimate |
| Qforia, by iPullRank (free) | Generates queries for AI Mode or AI Overview | Needs a Gemini API key |
| Gemini API with Search | Returns the queries the model ran | Applies to the API, not to AI Mode |
Otterly, which sells brand monitoring in AI, says its tool shows the queries the engines “might decompose” the prompt into. It also classifies each query by type, such as comparison or implicit question, and lets you track the chosen ones as prompts. NotHuman also sells AI visibility monitoring, so it has a stake in this market. We describe each tool from its own public page.
According to Semrush, Qforia, a free tool from the agency iPullRank, asks for the keyword, a Gemini API key and the mode. Then it generates related queries.
The Gemini API documentation explains that the model decides whether search helps, generates one or more queries and runs them. The API response, in turn, returns those queries. In other words, the data is real, but it shows Gemini through the API, not the query fan-out of AI Mode.
What one test showed
In the Semrush experiment, the team updated 4 articles from its own blog. For each one, it listed 10 to 20 synthetic queries, generated by Qforia and by a Screaming Frog script. Then it edited the text to answer as many of them as possible. It measured AI Mode and ChatGPT search for one month, with Semrush’s own paid tool.
Citations for the four articles went from 2 to 5. However, brand mentions fell from 18 to 10, and share of voice from 23.4% to 20.0%. Semrush blamed the drop on fewer brand mentions overall. The author concludes that optimizing for query fan-out helps, but with “a huge asterisk.” After all, the sample is small.
How do you turn the fan-out map into a content plan and prompts?
Group the sub-questions by intent and decide, for each group, whether it becomes an H2 on an existing page or a new page. Next, write each H2 as a question, with the direct answer right below. Finally, turn the main sub-questions into fixed prompts, to measure whether the brand appears in the answers.
- Group by intent: questions that need the same answer stay together, even with different words.
- Decide H2 or page: Google’s guide says you do not need to split content into small chunks, because its systems understand several topics on one page. Also, creating separate content for each fan-out query to manipulate search violates the scaled content abuse policy. So a new page makes sense only when the sub-question has its own intent. To connect those pages without one competing against another, see how to organize pages into a topic cluster.
- Write the answer first: the H2 repeats the question, and the first sentence already answers it, in 40 to 60 words.
- Link the pages: each new page links to the main page of the topic, and the main page links back.
- Track: put 10 to 20 sub-questions on the fixed prompt list and run it every month, following the protocol in how to measure brand mentions and citations in AI.
- Review every quarter: redo the clues for Google fan-out queries, since autocomplete and the answers change.
In short, for SEO, the map of fan-out queries only has value when it becomes a content plan and measurement.
How does NotHuman use query fan-out on its blog?
At NotHuman, the blog’s headings follow this logic. In each article, the H2s are questions that answer the sub-questions the AI generates from the main search. In this guide, for example, the five H2s came from the sub-questions of the search about query fan-out.
Main sources
- Google Search Central, optimizing your website for generative AI features on Google Search (2026).
- Google, AI Mode announcement in Search Labs (2025).
- Google, AI Mode updates at Google I/O (2025).
- Google Search Help, how autocomplete predictions work (2026).
- Search Console Help, how Search Console counts impressions, position and clicks (2026).
- Google AI for Developers, grounding with Google Search in the Gemini API (2026).
- Semrush, query fan-out optimization experiment (2025).
- Otterly, query fan-out tool (2026).



