How Google Decides Which Sources Appear in AI Overviews

Google’s own documentation answers this question directly. As of October 2026, its guidance on AI features for site owners states that “there are no additional requirements to appear in AI Overviews or AI Mode, nor other special optimizations necessary.” Its AI optimization guide describes how the supporting links are found: the system relies on Google’s “core Search ranking systems” to retrieve pages from the Search index. Source selection for an AI Overview therefore begins inside the same ranking machinery that orders the classic results. What Google adds on top is a different way of asking the index questions, and a separate decision about whether an AI Overview should appear at all.

The practical picture has three layers: gates a page must pass, a retrieval process that widens the pool of candidates, and a triggering decision that sits above both. Google describes each layer in public. It does not publish how the layers are weighted.

The three gates a page must pass first

Before any judgment about quality or relevance, a page has to clear conditions that are binary and checkable.

It must be indexed. A page Google has not indexed cannot be retrieved, so it cannot be cited.

It must be eligible to show a snippet. The AI features page says a supporting link must be “indexed and eligible to be shown in Google Search with a snippet.” This is where the preview controls matter. Google’s robots meta tag documentation says nosnippet “will also prevent the content from being used as a direct input for AI Overviews and AI Mode,” and that max-snippet limits how much of the content can be used as input. A template-level nosnippet added years ago for an unrelated reason quietly removes a page from consideration.

It must not be excluded in Search Console. Google says it rolled out a Search generative AI control in Search Console to all websites worldwide on August 31, 2026. “Include” is the default. “Exclude” removes the site’s links from AI Overviews, AI Mode and generative AI features in Discover, and stops its content being used as grounding input. Google states that the setting “isn’t used as a ranking or inclusion signal affecting other parts of Search,” so choosing Exclude does not cost classic rankings.

One control is easy to confuse with these gates. Google-Extended, a token you can address in robots.txt, governs training of future Gemini models and grounding in Gemini Apps and Vertex AI. Google’s crawler documentation says it “does not impact a site’s inclusion in Google Search nor is it used as a ranking signal in Google Search.” The Search Console help page adds that Google-Extended is also the way to limit training of the models behind Search’s generative features. Blocking it is a training decision. Whether a page can be cited in an AI Overview is decided by the three gates above.

A page that fails any gate is out, however good it is. A page that passes all three has only earned the right to be considered.

How fan-out widens the pool of candidates

The AI features page explains that both AI Overviews and AI Mode may use a “query fan-out” technique, which Google defines as “issuing multiple related searches across subtopics and data sources.” The AI optimization guide gives an example. A search for “how to fix a lawn that’s full of weeds” can trigger related searches such as “best herbicides for lawns,” “remove weeds without chemicals” and “how to prevent weeds in lawn.”

This widens what “ranking” means for citation purposes. If the candidate pool for one AI Overview is drawn from several related searches, a page can be cited for a query it does not rank well for, because it ranks for one of the subqueries. Three levels of evidence need to be kept apart here. That Google can use fan-out is documented. That most cited URLs no longer come from the query’s own top 10 is a measurement: in Ahrefs’ 2026 analysis of 4 million AI Overview URLs, 37.1% of cited URLs ranked in the organic top 10 for the same query, against 76.10% of the top three citations in its 2025 study, which used a different method. That fan-out caused the change is Ahrefs’ interpretation, not something Google has stated, and Google does not say how many subqueries are issued or how their results are combined. Treat fan-out as a documented capability and a plausible explanation, not as a proven account of any particular citation.

Fan-out also has a policy edge. The AI optimization guide warns that creating separate pages for each fan-out variation, when the primary purpose is to manipulate rankings or generative AI responses, violates Google’s scaled content abuse spam policy. Covering a subtopic properly inside a page that serves the reader is one thing. Spinning out thin pages to match guessed subqueries is the behavior the policy names.

The decision about whether an AI Overview appears at all

A page can pass every gate and rank for every subquery and still never be cited, because for many queries no AI Overview is shown. The AI features page says AI Overviews “are only shown when our systems determine that it is additive to classic Search, and as such, often don’t trigger.”

Google has named some categories where it holds back. In a May 2024 post on AI Overviews, Liz Reid of Google wrote that “we aim to not show AI Overviews for hard news topics, where freshness and factuality are important,” and described added “triggering refinements” for health queries and restrictions for queries where AI Overviews “were not proving to be as helpful.”

The two AI surfaces also differ. Google says AI Mode and AI Overviews “may use different models and techniques, so the set of responses and links they show will vary.” A citation in one tells you little about the other.

What Google says it looks for in the content

Google’s description of what makes a page worth citing is the same description it gives for Search generally. The AI optimization guide puts it plainly: “From Google Search’s perspective, optimizing for generative AI search is optimizing for the search experience, and thus still SEO.”

The guide does add one useful distinction. It contrasts commodity content, with the example “7 Tips for First-Time Homebuyers,” against non-commodity content, with the example “Why We Waived the Inspection & Saved Money: A Look Inside the Sewer Line.” The second kind contains information the reader cannot get from a hundred near-identical pages. When fan-out pulls candidates for a narrow subquery, a page holding a specific, first-hand answer has something the generic pages do not.

The same guide rules out several popular tactics for Google’s surfaces. It says Google Search does not use llms.txt files, that there is no ideal page length and no need to “chunk” content, that rewriting pages specifically for AI is unnecessary, and that inauthentic “mentions” do not help. Structured data is not required for AI features, though it remains useful for rich results. The AI features page adds that no special schema.org markup and no new machine-readable files are needed.

Where Google’s explanation stops

Google describes the process. It does not publish weights, thresholds or a scoring formula. The AI optimization guide states that “no third-party tool has access to our internal ranking or AI systems.” Any claim about a specific percentage weight for freshness, a correlation coefficient for “semantic completeness,” or a named multi-stage pipeline that runs separately from ranking comes from outside observers, not from Google’s documentation. Some of those observations are carefully measured and some are not, and none of them is Google describing its own system.

Google has also described how the system fails, which is informative about how it chooses. In the May 2024 post, Reid wrote that when AI Overviews get something wrong, “it’s usually for other reasons: misinterpreting queries, misinterpreting a nuance of language on the web, or not having a lot of great information available.” That last reason, which Google calls a data void, means that when few good pages exist on a narrow topic, the system has little to select from. For a site owner, a data void is also an opening: a careful page on an under-covered question faces less competition for retrieval.

What to check on your own site

Each gate and layer above maps to something you can verify.

1. Indexing. Use the URL Inspection tool described in Google’s recrawl documentation to confirm that the pages you care about are indexed.
2. Snippet eligibility. Search your templates and HTTP headers for nosnippet, max-snippet and data-nosnippet. Confirm each one was placed on purpose.
3. The Search Console setting. Open the Search generative AI control and confirm it is set the way you intend. Remember that Include is the default.
4. Impressions. The Generative AI performance report, which Google says reached all websites on August 31, 2026 (its help page still notes that some properties may not have access yet), shows impressions in AI Overviews and AI Mode by page, country, date and device. It is the first place to see whether any of your pages are being surfaced.
5. The content itself. For each priority page, ask what it contains that a dozen competing pages do not. If the answer is nothing, the page is commodity content in Google’s own terms, and no technical adjustment will change that.

The useful conclusion from Google’s documentation is a narrow one. Google decides which sources appear in AI Overviews by retrieving from its normal index with its normal ranking systems, sometimes across several related queries, and only when it judges an overview worth showing. The levers you control are the gates, which are checkable today, and the substance of the page, which is the same work good SEO has always required. Effort spent on tactics Google has said it ignores is effort taken from the parts it has said matter.