How do AI Overviews work?

Google's AI Overviews now sit above the normal results for a large share of searches, summarising the answer and citing a few sources. For a marketer, the useful question is not what they look like but how they are built, because the mechanism decides what you can actually do about them.

An AI Overview is generated, not looked up. When a query triggers one, Google breaks it into sub-questions, retrieves passages from pages already ranking for those components, and uses a Gemini model to write a single answer with a handful of cited sources.

The short version: it is a layer built on top of the search index you already know, not a separate system you submit to. This is how it works, step by step, and what each step means for whether you get cited.

How AI Overviews work, in one pass

At a high level an Overview runs through four stages: a query triggers it, the query is fanned out into sub-questions, pages are retrieved for each, and a model synthesises the cited answer.

None of those stages is a new ranking engine. Each one sits on top of Google's ordinary index and relevance signals. That single fact is why most advice framing Overviews as a brand-new discipline points people at the wrong work.

What triggers an AI Overview

Not every search gets one. Google decides, per query, whether a synthesised answer is worth showing.

Overviews appear most on informational and how-to queries, where summarising genuinely helps. They appear less, or not at all, on navigational and transactional searches, and Google is more cautious on health, finance, and other high-stakes topics where a wrong synthesis carries risk.

You do not control the trigger. There is no setting that forces an Overview to appear for your query, and the only way to keep a page out costs you your normal snippet too. So the trigger is context, not a lever.

Query fan-out: how the query is broken up

The step that matters most for content is fan-out. Google does not answer your query as one lump. It decomposes it into a set of related sub-questions.

A search like "best crm for hospitality" quietly becomes several questions: what features hospitality needs, which tools serve it, how they compare on price, and so on. Google retrieves pages relevant to each of those components separately.

This is why breadth of coverage helps. The Overview is assembled from answers to the sub-questions, so a page that answers the whole cluster gives Google more surfaces to cite than one that answers only the headline query.

Where the answer comes from

The retrieved passages are grounded in Google's live index. The Gemini model does not write the Overview from training data and hope it is current. It is anchored to real pages it can cite and link.

That grounding is the reason Overviews can reference recent content and send clicks at all. It also means your page has to be in the index, cleanly, to be a candidate. A page that cannot be crawled or rendered is invisible to this step, no matter how good the writing is.

Because the candidate pages come from the same index that produces the blue links, the pages cited in an Overview are usually pages already ranking for the query or its parts. The Overview is built on conventional relevance, not in place of it.

Why AI Overviews cite the sources they do

The final step is synthesis, and it selects passages, not whole pages. The model reaches for the source that already stated the answer to a sub-question cleanly.

So two things decide whether you are cited. First, are you a strong enough conventional result to be in the candidate set for that sub-question. Second, did you answer it in a self-contained passage that lifts cleanly, rather than one the reader has to assemble from several paragraphs.

This is the same logic assistants use everywhere, which I set out in how AI assistants choose sources. A page that makes the reader do the summarising loses to one that did the summarising itself.

What this means for your visibility

Once you see the mechanism, the work stops being mysterious. You cannot optimise an Overview directly, so you optimise the things it is built from.

Earn conventional visibility for the query and its sub-questions. Cover the cluster of related questions, not just the headline. Write answer-first passages that extract cleanly. That is the whole route in, and I have set it out in full in how to rank in Google AI Overviews.

The one trap to avoid is reading an Overview's arrival as pure loss. Sometimes it satisfies the query and costs you the click, which shows up as rankings holding while traffic falls. Knowing the mechanism is what lets you tell that pattern apart from a genuine ranking drop.

The takeaway

AI Overviews work by triggering on certain queries, fanning the query into sub-questions, retrieving passages from pages already ranking in Google's index, and synthesising a cited answer with a Gemini model. It is a layer on top of ordinary search, not a separate one. So the way to be cited is to be a strong, extractable conventional result for the query and every sub-question around it.

If you want a measured read of where you currently appear across AI Overviews and the assistants your buyers use, that is what an AI visibility audit provides.

This article is part of the SEO in the AI Era: The Complete Guide guide.