Google documents two things about how AI Overviews pick sources: an eligibility floor and a retrieval mechanism. A page has to be indexed and allowed to show a snippet, and the answer is built by running several related searches rather than one. That is the whole of the published record. Everything else you have read about AI Overview citations, including the ranked factor lists, is inference from samples that nobody outside Google can check.
This post separates the two piles. First what Google states in its own documentation, quoted. Then what people assert, with the reason each claim cannot currently be verified. Everything here is Google AI Overviews. ChatGPT, Perplexity and Claude are different products with different retrieval, and none of this transfers to them.
The short answer
- Observable: the eligibility floor, the fan-out mechanism, the snippet controls, and the fact that overviews often do not trigger.
- Not observable: any weighted list of citation factors. Google has not published one and nobody can verify one from outside.
- Google states plainly that no special markup, files or optimisations are needed.
- The one thing you can measure yourself is the citation list on your own queries.
What Google states, quoted
Every line in this section comes from Google Search Central's AI features documentation, last updated 10 December 2025, or from Google's Search Console performance report methodology page, Help Center article 7042828 on support.google.com.
There is an eligibility floor, and it is low. "To be eligible to be shown as a supporting link in AI Overviews or AI Mode, a page must be indexed and eligible to be shown in Google Search with a snippet, fulfilling the Search technical requirements. There are no additional technical requirements."
That sentence is the most actionable one in the entire documentation set. It also carries an immediate operational consequence: a nosnippet directive, a restrictive max-snippet, or a data-nosnippet wrapper on your main content removes the page from consideration. Sites have done this to themselves while trying to control snippet length.
The candidate pool comes from more than one query. "Both AI Overviews and AI Mode may use a 'query fan-out' technique, issuing multiple related searches across subtopics and data sources, to develop a response. While responses are being generated, our advanced models identify more supporting web pages, allowing us to display a wider and more diverse set of helpful links associated with the response than with a classic web search."
This is the mechanism behind the single most common surprise in AI visibility, which is that the cited set does not match page one.
Overviews often do not appear. "AI Overviews are only shown when our systems determine that it is additive to classic Search, and as such, often don't trigger."
AI Overviews and AI Mode are not the same system. "AI Mode and AI Overviews may use different models and techniques, so the set of responses and links they show will vary." A measurement of one is not a measurement of the other.
Nothing special is required, and Google says so twice. "There are no additional requirements to appear in AI Overviews or AI Mode, nor other special optimizations necessary." And: "You don't need to create new machine readable files, AI text files, or markup to appear in these features. There's also no special schema.org structured data that you need to add."
The controls are the ordinary Search controls. "AI is built into Search and integral to how Search functions, which is why robots.txt directives for Googlebot is the control for site owners to manage access to how their sites are crawled for Search." Google-Extended is described separately, as a control for "AI training and grounding in some of Google's other systems". The two are constantly confused, and we pulled that apart in Google-Extended vs Googlebot.
Every link in one overview shares one position. "An AI Overview occupies a single position in search results, and all links in the AI Overview are assigned that same position." There is no ordering signal to reverse engineer from Search Console.
What we can observe ourselves, and what it is worth
The second observable category is the citation list itself. It is public: ask Google a question, and the sources attached to the answer are right there. Read enough of them and you have a description of the cited set, which is not the same as knowing the rule that produced it.
Here is ours. On 15 August 2026 we asked Google nine buyer-intent questions about rank tracking tools through Bright Data's SERP API, US and English. Eight completed. All eight returned an AI Overview, and those eight answers carried 70 citation slots across 40 distinct domains, between 6 and 14 sources per answer.
Of those 70 citation slots, 31 pointed at a domain that also appeared in the organic results returned in the same API response for the same query. The other 39 did not. And 24 of the 40 cited domains never appeared in the organic block of any of the eight queries.
Eight answers on one topic in one country on one day. That is a probe, not a study, and we are not going to dress it up as one. What it supports is narrow and useful: the cited set and the ranked set are different sets, which is exactly what Google's fan-out description predicts. What it does not support is any statement about why a particular domain was chosen. We can see who was cited. We cannot see the reason, and neither can anybody else who is selling you one.
The claims that cannot currently be checked
Each of these is in wide circulation. None has a primary source behind it.
| Claim | Status |
|---|---|
| "The N ranking factors for AI Overviews" | Google publishes no factor list for AI features. Any list is somebody's inference from their own sample, and their sample is usually not published either. |
| "Schema markup increases your odds of being cited" | Contradicted by Google's own text: "There's also no special schema.org structured data that you need to add." Schema still earns rich results where a feature exists, which is a different argument. |
| "llms.txt gets you into AI answers" | No major AI search product has confirmed that it reads llms.txt, and Google's documentation explicitly says AI text files are not needed. Covered in what llms.txt actually is. |
| "Cited passages average N words" | Requires a corpus of cited passages, a comparison group of uncited ones, and a stated method. We have not seen a public dataset that has all three. |
| "AI Overviews appear on X% of queries" | Meaningless without the query set. Our own buyer-intent set returned 8 of 8, while provider documentation had us expecting 15 to 20 percent. The query type dominates the number. |
| "We changed the page and got cited" | A sample of one page, one query and one time period, against an answer that regenerates. Not evidence of a mechanism. |
| "FAQ schema helps you get cited" | The FAQ rich result was deprecated on 7 May 2026, and the schema was never claimed by Google to affect AI features. See what to do after FAQ schema. |
The pattern is the same in each row. There is a real observation somewhere underneath, and then a causal claim bolted onto it that the observation cannot carry.
Why "no special optimisation" is not the same as "nothing to do"
The honest reading of Google's position is not that effort is pointless. It is that the effort is the ordinary work, aimed at a surface that happens to be new. Google's own list of what still matters for AI features is short and dull. Crawling allowed in robots.txt and at the CDN. Content findable through internal links. Important content available as text. Structured data matching the visible page. Good page experience.
So the defensible programme has two halves and no third.
Half one, eligibility. Confirm the page is indexed, snippet-eligible, and that your primary content actually exists in the HTML rather than arriving later from JavaScript. This is checkable rather than arguable. Our free answer readiness checker and the AI content extractor run against a URL with no signup, and make your site readable to AI agents covers the same ground in more detail.
Half two, measurement. Ask your own questions on a schedule and record who gets cited. That gives you a number that belongs to your site rather than to somebody's aggregate sample. The method, the cost and the failure modes are in how to measure AI visibility.
There is no half three. The gap between those two halves is where the ranking-factor lists live, and it is empty.
What to do with a vendor's factor list
Ask three questions, in this order.
- What was the query set? If it is not published, the percentage means nothing. We have now watched the same method return 8 of 8 and 15 to 20 percent, depending only on the questions asked.
- What was the comparison group? A list of traits shared by cited pages is not a finding unless somebody checked whether uncited pages share them too. Most cited pages are in English. That is not a citation factor.
- Which product was measured? AI Overviews, AI Mode, ChatGPT and Perplexity are separate systems, and Google states that even its own two vary from each other. A single blended visibility score across all of them is a modelling choice, not a measurement.
If a vendor answers all three, read the work. Ours are answered in the open, including the parts that undercut us: our biggest finding rests on eight questions, on one topic, on one day.
What to do next
Check the eligibility floor first, because it is binary and it is the only thing on this page that is fully in your control. Confirm the page is indexed, confirm no nosnippet or max-snippet directive is limiting it, and confirm the main content is in the HTML. Then start measuring your own citation lists weekly, so that in three months you are arguing from your data instead of somebody's blog post. What that looks like for this engine specifically is on tracking AI visibility in Google AI Overviews. If you need the vocabulary for the conversation, AEO explained without the sales pitch and SEO vs AEO vs GEO cover the terms, and our free tools do the checks without an account.