Albrecht LippkyResearch

· 10 min read

What research and patents describe query fan-out?

Google describes AI Mode query fan-out as multiple related searches issued concurrently across subtopics and data sources, followed by combined results.

A tall stack of clipped research papers and patents with coloured tabs on a violet backdrop

Photo-illustration: Fan-out (AI-generated)

Summarize with ChatGPTClaudePerplexity

TL;DR

  • Google describes AI Mode query fan-out as multiple related searches issued concurrently across subtopics and data sources, followed by combined results.[12]
  • Query expansion, multi-query retrieval and rank fusion are distinct research layers rather than interchangeable names for one algorithm.[9]
  • RAG-Fusion is a close academic analogue because it generates multiple queries, retrieves for each and fuses the rankings, but it can add latency and off-topic noise.[9]
  • Google patents describe query variants and thematic search, but patent records and SEO interpretations do not prove the exact production architecture of AI Mode.[12]
  • Google Search Central says publishers do not need special AI Mode markup or technical optimization beyond ordinary Search eligibility.[14]

Query fan-out is best understood as an orchestration pattern: a system breaks a complex information need into related searches, retrieves evidence for those branches, and combines the results into an answer. Google publicly describes AI Mode as issuing multiple related searches concurrently across subtopics and data sources before combining the results, but it does not disclose the exact branch count, selection rules, or fusion algorithm.[1]

The academic lineage is broader than generative AI. Earlier information-retrieval research studied query expansion, in which additional terms or representations improve matching; later work examined question decomposition, multi-query retrieval, and rank fusion. These techniques illuminate parts of the fan-out pattern, but none of the papers below proves that Google uses its method in AI Mode.[2][3]

Papers behind the query fan-out pattern

Work Core technique Relationship to query fan-out Main limitation
Cormack, Clarke and Buettcher, “Reciprocal Rank Fusion” Combines document rankings from multiple retrieval systems.[4] Describes the result-fusion layer. It does not generate queries.
Bernhard, “Query Expansion based on Pseudo Relevance Feedback from Definition Clusters” Extracts terms from relevant definition clusters to expand a query.[5] An early query-expansion ancestor. Expansion does not necessarily create independent searches.
Wang, Yang and Wei, “Query2doc” Uses an LLM-generated pseudo-document to expand the original query.[6] Shows how one query can gain additional retrieval context. It is not, by itself, multi-branch retrieval.
Gao et al., “HyDE” Generates a hypothetical document, embeds it, and retrieves similar real documents.[7][8] Connects a user question to a richer retrieval representation. The generated document can contain hallucinated details.[8]
Rackauckas, “RAG-Fusion” Generates multiple queries, retrieves for each, and fuses the rankings with RRF.[9] A close research analogue to multi-query fan-out. Extra branches add latency and may introduce off-topic results.[9]
Zhuang et al., “EfficientRAG” Iteratively generates retrieval queries and filters irrelevant information.[10] Represents an iterative, multi-hop variant. Iterative retrieval is not identical to concurrent fan-out.[10]

How did query expansion lead toward fan-out?

Query expansion predates large language models. Its basic premise is that the words in a user’s query may not match the vocabulary used in relevant documents, so the system adds related terms or representations before retrieval. Bernhard’s 2010 work, for example, uses pseudo-relevant definition clusters to extract terms that address the lexical gap between a query and relevant material.[5]

This is related to fan-out but should not be treated as equivalent. A conventional expansion system may transform one query into a richer query and run one retrieval operation. Fan-out instead implies multiple related searches or branches whose evidence is later combined. The distinction matters because “more terms in one query” and “several independently retrieved subqueries” have different effects on coverage, ranking, latency and error propagation.

What does Query2doc contribute?

Query2doc uses few-shot prompting to generate a pseudo-document from the original query, then uses that generated text to expand retrieval. The paper reports a 3%–15% improvement in BM25 retrieval without fine-tuning.[6]

That result supports LLM-assisted query expansion, not a claim that the system ran multiple independent searches. Query2doc changes the representation of one information need; a fan-out system may generate several distinct branches such as “best products,” “pricing,” “technical limitations” and “independent reviews,” then retrieve evidence for each.

What does HyDE contribute?

HyDE, or Hypothetical Document Embeddings, asks an LLM to write a hypothetical document responding to the query. A dense encoder embeds that document, and the system retrieves similar real documents from the corpus.[7]

HyDE therefore sits between query reformulation and document retrieval. It demonstrates how a generated representation can bridge the vocabulary gap between a short question and relevant documents, but it does not require multiple query branches. The paper also notes a central risk: the hypothetical document may hallucinate, even though its embedding can still help locate relevant evidence.[11]

Why is RAG-Fusion the closest academic analogue?

RAG-Fusion explicitly combines multi-query generation with retrieval and rank fusion. The system generates several search queries from the original question, retrieves documents for each query, and applies Reciprocal Rank Fusion to produce a combined ranking before generation.[9]

Its architecture separates two operations that are often conflated:

  1. Query generation: create multiple formulations or perspectives on the original question.
  2. Result fusion: combine the ranked documents returned by those searches.

Reciprocal Rank Fusion itself is older than RAG-Fusion. The SIGIR paper presents it as a method for combining document rankings from multiple information-retrieval systems, reporting stronger results than individual systems and other tested approaches in its experiments.[4] RRF can therefore be used after query fan-out, but RRF is not synonymous with fan-out and does not explain how branches are generated.

The RAG-Fusion study also illustrates the tradeoff. Across ten runs, its average response time was 34.62 seconds, compared with 19.52 seconds for the RAG baseline; the study also reports that generated queries could take answers off topic.[9] Multi-query retrieval can improve coverage, but each additional branch creates more retrieval work and another opportunity for irrelevant evidence to enter the synthesis stage.

What does Google publicly say about AI Mode fan-out?

Google’s public description is specific about parallelism and breadth. AI Mode can issue multiple related searches concurrently across subtopics and multiple data sources, then combine the results into an answer.[12][13] Google’s announcement also describes a multistep approach that can plan, search and adjust its plan while answering complex comparison questions.[3]

That disclosure establishes a product-level pattern, not a complete implementation blueprint. Google does not publicly specify the number of branches, the rules for selecting subtopics, the retrieval systems used for each branch, or whether it uses Reciprocal Rank Fusion, another ranking method, or a proprietary process.[9]

Google Search Central says that pages must be indexed and eligible for a Search snippet to be considered as supporting links for AI features, while noting that indexing and serving are not guaranteed.[14] It also says there are no additional technical requirements or special optimizations for AI Overviews or AI Mode.[14] Consequently, there is no evidence-based “fan-out markup” that publishers can add to force inclusion.

What do the Google patents describe?

Two Google patent families are frequently connected with query fan-out in SEO coverage, but they should be treated as related patent designs rather than proof of deployed AI Mode internals.

Patent Disclosed subject Connection to fan-out What it does not prove
US11663201B2, “Generating query variants using a trained generative model” A trained generative model produces query variants, with control models determining whether additional variants should be generated in response to search-system results.[15] Search Engine Land characterizes the design as one query producing related variants whose results can be combined.[15][12] The patent does not prove that this exact system powers AI Mode.
US20240386046A1 / US12158907B1, “Thematic search” Generates themes from summaries of responsive passages and presents results arranged by theme.[16][17] Search Engine Journal describes it as resembling subtopic exploration associated with fan-out.[16] The patent does not establish that Google executes parallel subqueries in production.

The two families describe different mechanisms. The query-variant patent is directly concerned with generating alternative queries.[5][8] Thematic Search is concerned with themes and the organization of responsive results.[17][16] A thematic presentation may support exploration, but it is not automatically evidence of concurrent retrieval.

Search Engine Journal explicitly cautions that Google has not confirmed the Thematic Search invention is in use.[1] The appropriate conclusion is therefore limited: the patents show that query variants, thematic organization and related retrieval workflows are documented design possibilities. They do not reveal the exact architecture behind Google AI Mode.

How should researchers distinguish expansion, decomposition and fusion?

A useful taxonomy separates four functions:

Function What the system does Representative research
Query expansion Adds terms, generated text or a richer representation to one information need.[6] Pseudo-relevance feedback, Query2doc and HyDE
Query decomposition Breaks a complex question into subquestions or retrieval objectives. Multi-hop and structured-retrieval research.[10]
Multi-query retrieval Runs retrieval separately for several generated queries.[9] RAG-Fusion
Rank or evidence fusion Aggregates results from multiple rankings or branches.[4] RRF and RAG-Fusion

These layers can be combined, but they are not interchangeable. A system might expand a query without branching, decompose a question without using RRF, or retrieve multiple branches and combine them with a different fusion method.

EfficientRAG provides a related example of iterative retrieval. It generates retrieval queries over multiple steps and filters irrelevant information, with the authors reporting 60%–80% better time efficiency than other iterative methods while maintaining similar GPU utilization.[10] Its iterative design is relevant to multi-hop retrieval, but it should not be relabeled as concurrent fan-out.

Research on question decomposition also belongs in the lineage because it formalizes the idea that a complex question can be divided into smaller retrieval tasks.[10][18] Earlier interactive question-answering work explored predictive questioning and dialogue continuations, providing a precedent for anticipating related questions.[19] Those studies, however, do not establish concurrent backend retrieval and should not be presented as direct evidence of modern AI search fan-out.

What can patents and papers establish?

Papers provide evaluated methods and measured outcomes. Query2doc reports retrieval improvements, RAG-Fusion reports a concrete multi-query pipeline and latency tradeoff, and RRF evaluates a way to combine rankings.[4][9]

Patents provide disclosures of claimed or described systems. They can show that query-variant generation or thematic search was documented by a company, but publication or grant does not establish that a product uses every described mechanism. The fan-out connection for Thematic Search comes from SEO interpretation, while the patent itself describes thematic result organization.[1][12]

Google’s own product documentation is the strongest source for what Google publicly says AI Mode does: related searches across subtopics and sources are issued concurrently, and their results are combined.[1][12] It is not sufficient to infer the undocumented details of query selection, ranking or synthesis.

What can safely be claimed about Google AI Mode query fan-out?

A defensible description is that Google says AI Mode performs multiple related searches concurrently across subtopics and data sources and combines the results into an answer.[1][13] It is also defensible to say that query expansion, multi-query retrieval and rank fusion provide academic precedents for parts of this workflow.[9]

It is not defensible to claim a fixed number of branches, a specific fusion algorithm, a guaranteed patent-to-product connection or a special fan-out optimization for publishers. Google Search Central says there are no additional technical requirements or special optimizations for AI Overviews or AI Mode.[14]

A numbered timeline of query fan-out’s lineage

  1. 2005–2006: Predictive question systems. FERRET and the Predictive Dialogue Network explored suggesting likely related questions or dialogue continuations.[19]
  2. 2009: Reciprocal Rank Fusion. Cormack, Clarke and Buettcher described a method for combining rankings from multiple retrieval systems.[4]
  3. 2010: Definition-based query expansion. Bernhard used pseudo-relevant definition clusters to identify terms for query expansion.[5]
  4. 2023: Query2doc. Query2doc applied few-shot LLM prompting to generate pseudo-documents for query expansion and reported 3%–15% BM25 improvements.[6]
  5. 2023: HyDE. HyDE used hypothetical documents and dense embeddings to retrieve similar real documents without relevance labels.[7]
  6. 2023: Google query-variant patent grant. US11663201B2 described generating query variants with a trained generative model.[15]
  7. 2024: RAG-Fusion. RAG-Fusion combined generated multi-query retrieval with Reciprocal Rank Fusion and measured both coverage potential and latency costs.[9]
  8. 2024: Thematic Search patent publication and grant. Google’s Thematic Search records described generating themes from responsive passages and arranging results by theme.[17][16]
  9. 2024: EfficientRAG. EfficientRAG studied more efficient iterative retrieval for multi-hop question answering.[10]
  10. 2025: Google’s public AI Mode description. Google described query fan-out as concurrent related searches across subtopics and sources, followed by result combination.[1]

What does this mean for publishers and researchers?

The research lineage suggests that fan-out is not one algorithm. It is a pipeline in which a system may expand an input, decompose an information need, retrieve across branches, rank or fuse evidence, and synthesize an answer.

For researchers, the most precise terminology is therefore functional. Use “query expansion” when the system enriches one query, “multi-query retrieval” when it runs several generated queries, “question decomposition” when it divides a complex task, and “rank fusion” when it combines retrieval rankings. Reserve “query fan-out” for a broader orchestration pattern involving multiple related searches.

For publishers, Google’s documented guidance is comparatively simple. Pages need to be indexed and eligible for Search snippets, and Google says there are no special AI Mode files or markup requirements.[14] Useful content that answers distinct aspects of a user’s information need may be discoverable through ordinary search systems, but no paper or patent establishes a guaranteed content format for fan-out visibility. Finseo’s own replication of generative engine optimization techniques points the same way: the effect of an edit was rank-dependent, moving a rank-5 page +115% and a rank-1 page −30%.

The strongest overall conclusion is cautious: academic work explains the component techniques, patents document plausible system designs, and Google confirms the high-level behavior of concurrent related searches. Together they describe a credible lineage behind AI search fan-out, not a proven blueprint of Google’s production implementation.[12]

Frequently asked questions

Is query fan-out the same as query expansion?

No. Query expansion enriches one query with additional terms, generated text or a hypothetical document representation.[6][5] Query fan-out generally refers to issuing multiple related searches or branches and combining their evidence, although a fan-out system may use expansion inside each branch.

What is multi-query retrieval?

Multi-query retrieval generates several formulations of an information need, retrieves documents for each formulation, and aggregates the results.[9] It can improve coverage, but it may also increase latency and introduce irrelevant or off-topic material.[10]

What is RAG-Fusion?

RAG-Fusion is a retrieval-augmented generation approach that generates multiple search queries, retrieves documents for each, and uses Reciprocal Rank Fusion to combine and rerank the results.[9] It is a close academic analogue to multi-query fan-out, but it does not prove that Google uses the same pipeline.

What is Reciprocal Rank Fusion?

Reciprocal Rank Fusion is a method for combining ranked document lists from multiple retrieval systems or searches.[4] It is a fusion technique rather than a query-generation method, so it can be used after fan-out but is not itself fan-out.

Does a Google patent prove that AI Mode uses query fan-out?

No. US11663201B2 describes generating query variants, while Google’s Thematic Search patents describe generating themes and organizing results.[12] SEO publications connect these ideas with fan-out, but the coverage also notes that Google has not confirmed the patented Thematic Search invention is deployed in the product.[2][1]

Is there special markup for Google AI Mode fan-out?

Google Search Central says there are no additional technical requirements or special optimizations for AI Overviews or AI Mode.[14] Pages must still be indexed and eligible for a Search snippet to be considered as supporting links, although inclusion is not guaranteed.[14]

Sources

  1. 1Google's Thematic Search Patent — searchenginejournal.com
  2. 2Query Fan-Out Technique in AI Mode: New Details From Google — searchenginejournal.com
  3. 3Expanding AI Overviews and introducing AI Mode — blog.google
  4. 4Reciprocal rank fusion outperforms condorcet and individual rank learning methods | Proceedings of the 32nd international ACM SIGIR conference on Research and development in information retrieval — dl.acm.org
  5. 5Query Expansion based on Pseudo Relevance Feedback from Definition Clusters — aclanthology.org
  6. 6Query2doc: Query Expansion with Large Language Models - ACL Anthology — aclanthology.org
  7. 7Precise Zero-Shot Dense Retrieval without Relevance Labels - ACL Anthology — aclanthology.org
  8. 8ExpandR: Teaching Dense Retrievers Beyond Queries with LLM Guidance — aclanthology.org
  9. 9RAG-Fusion: a New Take on Retrieval-Augmented Generation — arxiv.org
  10. 10EfficientRAG: Efficient Retriever for Multi-Hop Question Answering — aclanthology.org
  11. 11Question Decomposition for Retrieval-Augmented Generation — aclanthology.org
  12. 12What is query fan-out & how does it work for AI searches? — searchengineland.com
  13. 13AI Overviews and AI Mode in Search — search.google
  14. 14AI Features and Your Website | Google Search Central  |  Documentation  |  Google for Developers — developers.google.com
  15. 15US11663201B2 - Generating query variants using a trained generative model - Google Patents — patents.google.com
  16. 16US12158907B1 - Thematic search - Google Patents — patents.google.com
  17. 17US20240386046A1 - Thematic search - Google Patents — patents.google.com
  18. 18Query, Decompose, Compress: Structured Query Expansion for Efficient Multi-Hop Retrieval | Proceedings of the 34th ACM International Conference on Information and Knowledge Management — dl.acm.org
  19. 19Proceedings of the... — aclanthology.org