How Does Google AI Overview Choose Sources? – Retrieval and Citation
Learn how an AI Overview selects the pages it cites: retrieval over ranking, passage quotability, corroboration, and what makes a page uncitable.
Source selection for a generated answer is a retrieval problem before it is a ranking problem. The system needs passages it can compose from, which is a different requirement from a page being a good destination.
1. Retrieval Comes First
Generation is grounded in passages the system retrieved. A page that ranks well but contains no passage that answers the question directly can be passed over in favour of one that does.
This is the single most useful reframing: the unit competing for citation is a passage, not a page.
2. What Makes a Passage Quotable
| Property | Why it helps |
|---|---|
| Answers one question completely | Can be lifted without surrounding context |
| Sits near its heading | Chunking keeps question and answer together |
| States specifics | Numbers, mechanisms, and conditions survive summarisation |
| Uses plain structure | Lists and tables extract cleanly |
| Avoids ambiguity | Pronouns and vague antecedents make a passage unusable alone |
A paragraph beginning "This is why it matters" cannot be quoted, because nothing in it identifies the subject.
3. Corroboration Reduces Risk
A generated answer asserts things, so the system prefers claims it can find in more than one independent place. A unique, uncorroborated claim is riskier to state and less likely to be repeated.
The practical consequence is counter-intuitive: being the only source making a claim is a disadvantage for citation, even when the claim is correct.
4. Accessibility Is a Precondition
None of the above matters if the fetcher cannot read the page.
- Content assembled only in the browser may return an empty shell to a fetcher.
- Blocking the crawler removes the page from the retrieval pool entirely.
- Login walls and interstitials do the same.
- Slow or unstable responses reduce the chance of successful retrieval.
5. Ranking Still Correlates
Retrieval draws on the same index that produces rankings, so authority, relevance, and quality signals still matter. Rank and citation correlate — imperfectly.
The gap between them is where the work is: a well-ranked page with no clean answer passage is exactly the page that ranks and is not cited.
6. What This Suggests in Practice
- Put the direct answer in the first sentence under each heading.
- Keep one question per section so chunk boundaries fall cleanly.
- Name the subject explicitly instead of relying on "it" or "this".
- Use tables for comparisons, which extract and quote well.
- Make sure the page renders server-side for fetchers.
- State claims specifically enough to be checkable and dated where relevant.
7. Measuring Rather Than Guessing
Selection varies between runs, so any claim about "what gets cited" needs repeated sampling: a fixed query set, several runs each, with the cited sources recorded every time.
Track your citation rate, not a single observation, and record which passage was quoted when you can — that reveals more about what the system found usable than any ranking report.
8. Summary
An AI Overview cites sources it could retrieve, quote, and corroborate. Passage-level answerability, explicit subjects, clean structure, and fetcher accessibility drive selection more directly than page-level ranking does.
Write so that any single section can be lifted out and still make sense on its own, then measure citation rates across repeated runs rather than trusting one check.