In plain terms
Ask a good librarian for “the best CRM for a mid-sized company” and they will not look for that exact sentence. They will check prices, read user reviews, look at integration options, and then give you a summary. AI search does the same thing in a few seconds, and the user sees only the final answer.
Why it matters
It changes what it means to be found. The pages used in an AI answer are the ones that come up for the sub-searches, which the user never typed and you cannot see. A site that does not rank for the main question can still be cited because one of its pages is the best result for a sub-question. Content that covers a topic's obvious follow-up questions, such as price, comparison, integration and limits, has more chances to be picked up.
Example
A user asks an AI search engine which CRM suits a 200-person company. Behind the scenes the system runs eight searches, among them “CRM pricing per user”, “CRM reviews mid-size companies” and “CRM integration with accounting software”. A vendor that sits fourteenth for “best CRM” is cited in the answer, because its integration guide is the top result for the third search.
Most often confused with
Query Fan-out vs. Query expansion
Query expansion is an older technique that adds terms to a single search so that it matches more documents. Fan-out breaks the question into separate questions, each with its own results, and has a model write one answer from all of them. Expansion improves a list of links; fan-out produces a synthesis.
Origin: Google used the term publicly in 2025 to describe how AI Mode and AI Overviews work.
Under the hood
A language model generates the sub-queries: facets of the topic, rephrasings, comparisons, specific entities and likely follow-up questions, sometimes using the user's location or history. The sub-queries run in parallel against the web index and other sources such as product, map and knowledge-graph data. Relevant passages, and not whole pages, are selected from the results, and a model writes the answer with citations. The number of sub-queries grows with the complexity of the question, from a handful to hundreds in deep-research modes. The sub-queries vary from run to run and are not reported to site owners. Practical consequences: relevance is judged at passage level, coverage of sub-topics matters, and passages should make sense on their own. The same pattern drives agentic RAG and deep research.