Query fan-out

Google Search Central is the primary statement: both AI Overviews and AI Mode may issue multiple related searches across subtopics and data sources to develop one response^[1]. Dounia Berrada, a senior engineering director on Search, put the everyday scale at about a dozen searches per question^[2]. The Deep Search mode inside AI Mode goes much further: Search Engine Journal, reporting Robby Stein, VP of product for Search, described dozens or hundreds of background queries for one question^[4].

Google holds a patent describing a related mechanism, Thematic search, US12158907B1, filed 16 May 2023 and granted 3 December 2024. It describes generating a set of themes, and then narrower sub-themes, from the documents an initial query returns^[5].

A 2024 survey of query optimisation for LLM retrieval pipelines names four atomic operations a query passes through before retrieval runs: query expansion, query decomposition, query disambiguation and query abstraction^[6]. Decomposition is the operation a fan-out exposes, since Google's own description is of multiple related searches issued across subtopics^[1]. The searches a page is retrieved for are therefore not the words the user typed.

Fan-out is not particular to Google, but the fan is narrower elsewhere. A study of more than 8,500 ChatGPT prompts across nine industries found 31% triggered at least one live web search^[7], averaging 2.17 searches on those prompts with a maximum of four^[3]. Perplexity documents a classifier that scores query complexity and routes harder questions to a multi-step Pro Search rather than a single pass^[8]. Pro Search separates planning from execution: the model writes a plan, then generates and runs the searches for each step in turn, so later sub-queries can depend on earlier results^[9].

Vendor guidance treats fan-out as a reason to cover a subject's sub-questions across a page or a cluster, on the argument that a page answering one narrow angle is less resilient as sub-queries multiply^[10]. Google's own position contradicts the stronger version of that advice: no additional requirements to appear in AI Overviews or AI Mode, and no special optimisation necessary^[11].