AI search
AI search: what changed and what did not
AI features sit on top of a search index, they do not replace it. This guide explains retrieval-augmented generation, how AI Overviews and AI Mode assemble an answer, what query fan-out means for content, and which widely repeated AI-optimisation tactics have no documented basis.
Almost every AI search feature in production is a retrieval-augmented generation system. The pattern is consistent: take the user’s question, retrieve relevant documents from an index, give those documents to a language model as context, and have the model write an answer grounded in them.
The important consequence is that the retrieval step is still a search engine doing its ordinary job. Crawling, indexing and ranking still decide what the model is allowed to see. A page that is not indexed cannot be retrieved and therefore cannot be cited.
AI Overviews and AI Mode
An AI Overview is a generated summary shown above search results, with links to the pages it drew on. AI Mode is a conversational interface where a question can be followed by further questions, with the context of the earlier ones retained.
Google has described AI Mode as using a query fan-out technique: the question is broken into multiple related searches, run across subtopics and data sources, and the results are synthesised into one answer. Practically, this means a single question can pull from many pages, each answering a fragment.
What this changes for content
The shift is from ranking a page to being the clearest available source for a specific claim. Content that gets extracted tends to share properties:
- The answer comes first, in a self-contained block, before the context and caveats.
- Claims are specific and attributable rather than hedged and general.
- Structure is genuine — real headings, real lists, real tables — so a passage can be lifted without losing meaning.
- The source of a factual claim is named on the page.
None of that is new advice. It is the same thing that made content extractable for featured snippets, applied to a system that extracts far more often.
What has not changed, despite the claims
Two things worth stating plainly because they circulate constantly:
There is no AI-specific structured data. No schema type exists that marks content for AI systems, and adding invented properties does nothing.
Google has said it does not use llms.txt. The file costs nothing to
publish and does no harm, but it is not a mechanism for influencing Google’s AI
features. Treat any tactic that promises AI visibility through a file or a tag
with the scepticism you would apply to a meta keywords tag.
The levers that do exist are the ordinary ones: be indexable, answer the question directly, and be the source worth citing.
Articles in this guide
The articles for this guide are being written. The guide above covers the whole topic in outline, and the article index lists everything published so far.