Q
QueryFanout.wiki The technical reference and research archive for query fan-out in AI search

Query fan-out

From QueryFanout.wiki, the technical reference and research archive for query fan-out
This article is about the retrieval technique. For Google's own statements, quoted in full, see What Google has said. For the patent register, see Patents. For the raw test data, see Data.

Query fan-out is a retrieval technique in which a search system decomposes a user query into multiple related subqueries, searches for information relevant to those subqueries, and combines the retrieved information into a response. Google publicly uses the term "query fan-out" to describe part of how AI Mode in Google Search handles complex questions.[1][3]

The earliest public use of the phrase by Google that we located is 5 March 2025, the day it announced AI Mode.[1] It gave the technique a formal one-sentence definition in its Search Central documentation in May 2026.[9] Google has never published how many subqueries a fan-out produces, has not said that every query fans out, and has not confirmed that any of its patents describes the production system. Those three gaps are the reason this reference exists. Every claim below carries a label saying what kind of evidence stands behind it.

DOCUMENTED stated publicly by Google   PATENT described in a Google patent   TESTED observed in a logged test on this site   INFERRED a reasoned or third-party claim, attributed   UNKNOWN nobody outside the platform knows
Contents
  1. What is query fan-out?
  2. How does query fan-out work?
  3. How does Google describe query fan-out?
  4. What do Google's patents describe?
  5. Does every query fan out?
  6. How many subqueries does a fan-out generate?
  7. What does a query fan-out look like?
  8. Is query fan-out the same as query expansion?
  9. What does query fan-out change for SEO?
  10. History
  11. What we do not know
  12. See also
  13. References

What is query fan-out?

DOCUMENTED Query fan-out is a retrieval technique in which one user query is turned into several related subqueries, each subquery is searched, and the results are combined into a single response. Google's own definition, published in May 2026, calls it "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."[9]

Three things are true of every description Google has published. The subqueries are generated by a model rather than typed by the user. They run concurrently rather than one after another. And they can target different data sources, not only the web index: Google names its Knowledge Graph, its Shopping Graph and local data from Maps.[4] The purpose Google states is breadth and depth: "This approach helps you access more breadth and depth of information than a traditional search on Google."[1]

The phrase is Google's. The underlying idea is much older. Query expansion, query reformulation and multi-query retrieval have decades of literature behind them, and Google's own patent on generating query variants with a generative model has a 2017 priority date.[12] What Google named in 2025 was a specific product behaviour: model-generated concurrent subqueries feeding a generated answer.

How does query fan-out work?

DOCUMENTED PATENT In the simplest defensible model there are five stages: the user prompt is received, a model decomposes it into candidate subqueries, the candidates are filtered, each surviving subquery is searched, and the results are synthesised into one response. Google has described the first, fourth and fifth stages publicly. The filtering stage appears only in patents.
Five-stage pipeline: user prompt, model generates candidate subqueries (documented), filter for relatedness and diversity (patent), four subqueries searched concurrently across web index, Knowledge Graph and Shopping Graph (documented), one response with links (documented), with a loop back from synthesis to retrieval for gaps
The five-stage model. Stage 3 is documented only in patents; the loop from synthesis back to retrieval is from Google's I/O 2025 keynote.

Do not assume this pipeline is identical across platforms. Google has described it for AI Mode. Perplexity exposes something structurally similar in its interface, and our tests record what that looks like.[15] ChatGPT and other systems disclose their searches in their own ways. The five stages are a model for reading the evidence, not a claim about any one system's code.

DOCUMENTED One detail from Google's keynote is easy to miss and matters for the SEO argument. After synthesis, the system "checks its work" and, "if it detects any gaps, it issues even more searches to fill them in."[4] Fan-out is not necessarily a single pass.

How does Google describe query fan-out?

DOCUMENTED Google says AI Mode divides a question into subtopics, issues multiple related searches at the same time across multiple data sources, and brings those results together into one response. It has repeated that description, in nearly identical words, in at least a dozen posts and documents since March 2025.

The three descriptions below are the ones most worth quoting exactly. The full set, with every date and URL, is on the Google page.

"It 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."Robby Stein, VP of Product, Google Search, 5 March 2025[1]
"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."Elizabeth Reid, VP and Head of Search, 20 May 2025[3]
"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."Google Search Central, "AI features and your website", last updated 10 December 2025[7]

Note the word "may" in the third quotation. It is the only Google document that attaches fan-out to AI Overviews, and it hedges. Note also what is absent from all of them: a number, a model version, and any statement that the technique runs on every query.

Liz Reid on stage at Google I/O 2025 with a screen showing one query fanning out into subqueries such as art galleries, breweries with live music, parks and green spaces for a picnic, off the beaten path restaurants, and best food tours in Nashville
The technique was named on stage at Google I/O on 20 May 2025, two months after its first appearance in a blog post. The slide shows a Nashville weekend query fanning out into subqueries. Frame from Google's keynote video; credit Google.

What do Google's patents describe?

PATENT Google's 2025 application on responding to "multifaceted queries" claims the full pipeline: a large language model generates candidate subqueries, the candidates are filtered by relatedness and diversity metrics, a search is run for each surviving subquery, and a response is generated from the combined results. Its dependent claims describe conditions under which the system decides whether to fan out at all. This is evidence of claimed engineering, not proof that production AI Mode works this way.

The closest patent to the public description is US 2025/0117381 A1, "Utilizing large language model (LLM) in responding to multifaceted queries", filed October 2024 with a priority date of 10 October 2023 and still pending.[10] Its abstract begins:

"Implementations leverage a generative model (e.g., a large language model (LLM)) to generate a plurality of candidate subqueries for multifaceted natural language (NL) based input, where each of the candidate subqueries is potentially directed to a facet or problem of the multifaceted NL based input. Those implementations further select, from the plurality of candidate subqueries and using one or more evaluation metrics, a subset of the candidate queries."US 2025/0117381 A1, abstract[10]

The filtering stage is described in unusual detail. A candidate subquery is kept "only if (a) its relatedness metric satisfies a threshold (indicating relatedness to the NL based input) and (b) (i) the selected set is empty or (ii) each diversity metric, between the given candidate subquery and any already selected candidate subqueries, satisfies a threshold".[10] In plain terms: keep subqueries that are about the question, and drop any that duplicate one already kept. The application even suggests asking the LLM itself to score relatedness "on a scale of 1-10".

Four other Google families describe adjacent mechanisms: "Search with stateful chat" (synthetic queries generated with session context), "Thematic search" (results clustered into themes that fire a second query), "Generating query variants using a trained generative model" (the 2017-priority ancestor, with a control model that decides whether to generate variants at all), and the "Generative summaries for search results" family behind AI Overviews (documents responsive to related, recent and implied queries pulled into the summary). All six entries, with claim language quoted and status badges, are on the Patents page.

None of these documents uses the phrase "query fan-out". The in-patent terms are "candidate subqueries", "synthetic queries", "themes", "query variants" and "related queries". Several patent numbers that circulate in SEO articles as "Google's fan-out patent" belong to Microsoft, Citibank or other assignees; the Patents page lists the ones we checked and excluded.

Does every query fan out?

UNKNOWN Google has not said. The evidence points to fan-out being conditional rather than universal, but the conditions Google uses in production are not public.

How many subqueries does a fan-out generate?

UNKNOWN Google has not published a fixed or typical number for AI Mode or AI Overviews. Every description it has given is non-numeric: "multiple", "a multitude", "several". The only numbers Google has attached to fan-out are "hundreds" for Deep Search, "a dozen" in one engineer's informal interview, and "dozens of sub-queries" for visual image search.
FigureWhoContextLabel
"a multitude of queries"Google, Elizabeth ReidAI Mode, May 2025[3]DOCUMENTED non-numeric
"hundreds of searches"Google, Elizabeth ReidDeep Search only, May 2025[3]DOCUMENTED
"a dozen searches"Google, Dounia BerradaVisual search interview, March 2026[8]DOCUMENTED informal, illustrative
"dozens of sub-queries"Google, Brad KellettVisual image fan-out, July 2026[11]DOCUMENTED images only
"estimated eight to 20 or more"Semrush, Zach ParuchExplicitly labelled an estimate; "Google hasn't published exact counts"[13]INFERRED
average 10.7, range 3 to 28Seer InteractiveGemini API with forced grounding, 501 prompts. Not AI Mode.[14]INFERRED API, not product
"5 to 11" typical for AI ModeAhrefsDistribution matches the Seer API data; capture method not stated[16]INFERRED
"two per prompt"Ahrefs help centreFan-outs returned with ChatGPT and Perplexity responses[17]INFERRED other platforms
3 per multi-part prompt, 1 for simple promptsThis sitePerplexity, 11 prompts plus 2 repeats, 8 September 2026[15]TESTED Perplexity, not Google

Two points follow. First, no published number for Google AI Mode comes from observing Google AI Mode. The measured figures all come from the Gemini API, from ChatGPT, or from Perplexity, each of which is a different system. Second, the numbers that are observed are smaller than the numbers that are estimated. In the launch Perplexity sample published here, multi-part prompts generally exposed three search subqueries. That result should not be generalised to Google AI Mode.

What does a query fan-out look like?

A fan-out is a short list of searches that together cover the parts of a question. Below are three examples of three different kinds. The label on each one is the point.

DOCUMENTED Google's own illustration

The only worked example we located in Google's public documentation, as of September 2026, is in its May 2026 optimisation guide:

"For example, if the original user's query is "how to fix a lawn that's full of weeds", fanout queries might include "best herbicides for lawns", "remove weeds without chemicals", and "how to prevent weeds in lawn"."Google Search Central, "Optimizing your website for generative AI features on Google Search"[9]

Google says "might include". It is an illustration written by Google, not a capture of the system.

TESTED An observed fan-out

How do I stop my sourdough starter from developing a grey liquid layer on top
  • sourdough starter grey liquid layer on top what is it
  • how to prevent hooch sourdough starter grey liquid
  • sourdough starter gray liquid hooch fix
Perplexity, 8 September 2026, record qfo-0003. Three searches shown by the platform under one step labelled "Checking how to avoid drying layers on sourdough starter". 16 sources retrieved, 4 domains cited. The word "hooch" does not appear in the prompt: the system named the phenomenon before searching for it.[15]

SYNTHETIC An illustrative decomposition

What is the best CRM for a 50-person law firm concerned about security and migration cost?
  • CRM for law firms
  • CRM security compliance
  • CRM pricing for 50 users
  • CRM migration costs
  • Salesforce vs HubSpot for legal firms
  • legal CRM integrations
  • CRM onboarding time
Illustrative decomposition written by hand. Not observed on any platform. The same prompt was run three times on Perplexity for comparison, where the platform showed three subqueries per run, not seven; see Examples.
No. Query expansion broadens one query by adding terms. Query fan-out takes one user query, generates several related subqueries and searches across them. The table separates the terms that are most often confused.
ConceptSame thing?Difference
Query decompositionClosely relatedBreaking a complex input into component information needs. The first stage of a fan-out, not the whole of it.
Query expansionNoAdds synonyms or related terms to broaden a single retrieval. One query, wider net.
Query rewritingNoReplaces the query with a reformulated single query. One in, one out.
Query variantsRelatedMachine-generated alternative forms of a query; Google's 2017-priority patent term. May be issued alongside the original.[12]
Synthetic queriesRelatedAny query generated by a model rather than a user. Fan-out subqueries are synthetic queries; not all synthetic queries are part of a fan-out.
Retrieval-augmented generationNoAn architecture: retrieve documents, then generate from them. Fan-out is one way of doing the retrieval step. Google's glossary lists both, separately.[9]
Deep researchNoAn agentic, multi-step process that plans, searches, reads and searches again. Google says Deep Search "uses the same query fan-out technique but taken to the next level".[3]
Query augmentationNoAn older SEO idea about one page ranking for a cluster of related terms. A ranking-side observation, not a retrieval-time mechanism. Some practitioners treat the two as "the same thing"; Google's descriptions do not support that.[18]

What does query fan-out change for SEO?

INFERRED The retrieval surface becomes the space of questions around a query rather than the literal query. A page is retrieved for a subquery it answers well, whether or not that subquery is ever typed by a person. The defensible response is to identify the distinct information needs a prompt decomposes into and make sure a strong, retrievable answer exists for each of them. It is not to put twenty keywords on one page.

Three documented facts anchor this. Google says the technique surfaces "a wider and more diverse set of helpful links" than classic search.[7] Google's patent describes diversity filtering whose purpose is to stop the subqueries collapsing into near-duplicates, which means the subqueries are meant to be genuinely different questions.[10] And Google has warned, in the same document that defines the term, that creating "separate content for every possible variation of how people might search (for example, by focusing on other queries that people have asked, or fan-out queries)" primarily to manipulate results "violates Google's scaled content abuse spam policy".[9]

The practitioner claims, what is measured and what is not, and the one published experiment that tried to optimise for fan-out and got mixed results, are on the SEO page.

History

Timeline from 2017 to 2027: five Google patent priority dates below the line (2017, late 2022, three in 2023) and seven Google public statements above the line from March 2025 to July 2026
Patents below the line, public statements above. The phrase arrived eight years after the first filing.

What we do not know

This list is maintained deliberately. An item leaves it only when a public source answers it.

See also

References

  1. Stein, Robby. "Expanding AI Overviews and introducing AI Mode." The Keyword, Google, 5 March 2025. blog.google. Earliest public use of the phrase that we located.
  2. Google. "Get AI-powered responses with AI Mode in Google Search." Google Search Help. support.google.com.
  3. Reid, Elizabeth. "AI in Search: Going beyond information to intelligence." The Keyword, Google, 20 May 2025. blog.google.
  4. Reid, Elizabeth; Patel, Rajan. Google I/O 2025 keynote, Search segment, 20 May 2025. Google's YouTube channel, English captions. youtube.com.
  5. Stein, Robby. "Bringing multimodal search to AI Mode." The Keyword, Google, 7 April 2025. blog.google.
  6. Reid, Elizabeth. "Google Search with Gemini 3: Our most intelligent search yet." The Keyword, Google, 18 November 2025. blog.google.
  7. Google Search Central. "AI features and your website." Last updated 10 December 2025. developers.google.com.
  8. McHugh-Johnson, Molly; Berrada, Dounia. "Ask a Techspert: How does AI understand my visual searches?" The Keyword, Google, 5 March 2026. blog.google.
  9. Google Search Central. "Optimizing your website for generative AI features on Google Search." Published 15 May 2026, last updated 10 July 2026. developers.google.com. Google's only formal definition and worked example that we located, as of September 2026.
  10. Revach, Asaf; Lee, Hongrae; Liang, Zhengzhong. "Utilizing large language model (LLM) in responding to multifaceted queries." US 2025/0117381 A1, Google LLC. Priority 10 October 2023, published 10 April 2025. patents.google.com.
  11. Kellett, Brad. "Celebrating 25 years of visual search innovation." The Keyword, Google, 14 July 2026. blog.google.
  12. Alakuijala, Jyrki, et al. "Generating query variants using a trained generative model." US 11,663,201 B2, Google LLC. Priority 29 April 2017, granted 30 May 2023. patents.google.com.
  13. Paruch, Zach. "Query Fan-Out: What It Is & How to Optimize for It." Semrush Blog, updated 17 August 2026. semrush.com.
  14. Haigler, Nick. "Gemini 3 query fan-outs research." Seer Interactive, 21 November 2025. seerinteractive.com.
  15. QueryFanout.wiki. "Query Fan-Out Observatory, batch 1." 8 September 2026. Dataset version 0.1.0. queryfanout.wiki/data/.
  16. Gavoyannis, Despina. "Query Fan-Out: What It Is and How to Optimize for It." Ahrefs Blog, 2 March 2026. ahrefs.com.
  17. Ahrefs Help Center. "How to view fanout queries generated by AI." help.ahrefs.com.
  18. Dooley, James; Salazar Jurado, Luis. "STOP Using Query Augmentation and Query Fan Out Until You Watch This." FatRank podcast, episode 290, 6 March 2026. fatrank.transistor.fm.
  19. Gray, Matthew K., et al. "Generative summaries for search results." US 11,769,017 B1, Google LLC. Priority 30 December 2022, granted 26 September 2023. patents.google.com.
  20. Rofouei, Mahsan, et al. "Search with stateful chat." US 2024/0289407 A1, Google LLC. Priority 28 February 2023, published 29 August 2024. patents.google.com.
  21. King, Michael. "How AI Mode Works and How SEO Can Prepare for the Future of Search." iPullRank, 27 May 2025. ipullrank.com.
  22. Stein, Robby. "AI Mode can now help you search and explore visually." The Keyword, Google, 30 September 2025. blog.google.
  23. Paruch, Zach. "We Ran a Query Fan-Out Experiment. Here's What Happened." Semrush Blog, 26 September 2025. semrush.com.
Categories:
  • Information retrieval
  • AI search
  • Google AI Mode