How Google AI Overviews chooses what to cite

Google's AI Overview does not read your page and decide whether it is good. It runs a retrieval pipeline, and the page is only ever in contention for the passages it holds.

1. 01

   The query is decomposed

   One typed question becomes a set of related sub-queries, commonly called query fan-out. They include narrower specifications, canonical rephrasings, translations and clarifications of the original.

2. 02

   Each sub-query retrieves separately

   Google runs retrieval for each branch and collects candidate documents. A page can enter the candidate pool through a sub-query it was never written for.

3. 03

   Passages are ranked, not pages

   The system chunks candidates and scores the chunks. It never quotes a page. It quotes a paragraph, a table row or a list item.

4. 04

   The answer is synthesised and attributed

   Gemini writes a single answer from the selected passages and attaches links to the sources it drew on. Attribution is per-passage, which is why cited pages often look unrelated to the ranked results below.

Ranking gets you into one pool. Fan-out decides how many pools you are in.

If citations are won at the sub-query level, then a page's job changes. It is no longer to be the best answer to one keyword. It is to hold defensible passages on several adjacent questions at once.

• Cover the question's neighbourhood on one page. Definition, mechanism, exceptions, comparison, cost, limits. Each under its own heading, each answerable without the others.
• Write headings as the sub-questions themselves. A fan-out query is a question a person would type. A heading phrased that way matches it directly; a heading like 'Our approach' matches nothing.
• Front-load. Over half of sampled AI Overview citations came from the first 30% of the cited page.
• Keep passages self-contained. A paragraph opening with 'it' or 'this approach' cannot be lifted without its neighbour, so it is not lifted at all.
• Put structure where the facts are. Tables and lists are cleanly extractable units. Prose with the numbers buried mid-sentence is not.

Before optimising for AI Overviews it is worth knowing how much of your query set actually triggers one. The published figures disagree enormously.

The spread is not a scandal, it is a sampling artefact. Trigger rate depends on query mix, country, device and vertical. Informational and question-shaped queries trigger far more often than navigational ones. The only number that matters for a business case is the rate across the questions your own buyers ask, measured on your own list.

• No one controls the citation. There is no submission, no index request, no paid placement. Anyone selling a guaranteed AI Overview citation is selling something they do not own.
• The fan-out set is itself unstable. Only about 27% of extracted fan-out queries stayed consistent across repeated runs. Optimising for a specific sub-query list is building on sand; optimising for topical breadth is not.
• Roughly 85% of AI references point at third-party sites, not the brand's own domain. Your own page is one lever of several, and not the largest.
• Coverage moves under you. Google recalibrated AI Overview triggering downward between July and November 2025, halving prevalence on one measure. Programme design should survive that happening again.