Google details how AI Overviews and AI Mode choose and cite pages

What Google's guide says, and since when

The AI features and your website page, from Google Search Central, was last updated on December 10, 2025, according to the document's own footer. Ten months later, it remains the official reference for how AI Overviews and AI Mode decide which pages to cite as support in AI-generated responses within Search. The text doesn't announce a new feature: it consolidates rules that SEO teams had already been trying to infer by trial and error.

For those working in growth and technical SEO, the document confirms something uncomfortable: appearing as a source for an AI response isn't the same as appearing in a traditional search result, and Google doesn't promise that the two will behave the same way in terms of metrics.

How AI chooses who to cite: query fan-out

According to Google, both AI Overviews and AI Mode can rely on a technique called query fan-out: the user's original question is broken down into several related queries, covering different subtopics and data sources. The models use these sub-queries to build the response and, in the process, identify more supporting pages than a traditional search would bring up.

In practice, this means a page can be cited in an AI response without ever having ranked well for the term the user actually typed. It came in through the side door, via one of the sub-queries generated by the fan-out. It's a broader discovery mechanism, but also a more opaque one: Search Console doesn't show which sub-query brought in that citation.

AI Overviews, according to the document, only appears when Google's system judges that the generated response adds something beyond what a traditional search would already deliver, which is why the feature is often not triggered at all. AI Mode, on the other hand, is described as better suited for searches that call for exploration, reasoning, or comparison, the kind of question that used to require several back-and-forth trips to the search engine.

Eligibility with no extra technical requirement, but a strict filter

The documentation is direct on a point worth underlining: there's no specific technical optimization for appearing in AI Overviews or AI Mode. The requirement is that the page already be indexed and eligible to appear in traditional Search with a snippet, meeting Google Search's standard technical requirements. Nothing beyond that.

This has a practical consequence: a page that Google can't properly index, or that's blocked by robots.txt, doesn't enter the race for citation in any AI response. The eligibility funnel is the same as always for Search, it just feeds a destination that behaves differently on the metrics side.

Best practices remain the same, they just cost more when they fail

Google reinforces a list of practices that technical SEO should already be applying regardless of generative AI:

• Making sure robots.txt and any CDN or hosting infrastructure allow crawling;
• Making content discoverable through internal links;
• Delivering a high-quality page experience;
• Making sure important content exists as text, not just in images or video;
• Supporting the text with quality images and videos, when it makes sense;
• Keeping structured data consistent with the visible text on the page;
• Keeping information up to date in Merchant Center and Business Profile.

None of these items is unknown to anyone who already takes technical SEO seriously. The difference is that, with fan-out expanding the universe of candidate pages, basic failures, like structured data that doesn't match the visible content, cost more: the chance of losing a citation grows in proportion to the number of entry points the AI now tests before building the response.

The blind spot for those measuring performance

Here's the detail that matters most to anyone handling organic attribution: Search Console doesn't separate AI Overviews and AI Mode traffic from the rest. All of that traffic falls under the "Web" search type in the Performance report, mixed in with clicks from traditional organic results.

This means an aggregate increase or decrease in the "Web" report can hide a change in composition: more citations in AI responses, fewer clicks on conventional blue links, or the opposite. Without native segmentation, the only way to investigate is to cross-reference Search Console performance data with tools like Google Analytics, looking at time on page and conversion by landing page.

Google itself states, in the document, that clicks coming from results pages with AI Overviews tend to be of higher quality, in the sense that the user spends more time on the site after clicking. It's a statement favorable to the format, but one that doesn't solve the underlying problem: it speaks to the click that happened, not to the responses that satisfied the user without generating any click at all. That second category simply doesn't appear in any Google report.

How to get closer to an honest attribution measurement

Given that Google doesn't provide a dedicated report, a practical approach to justifying SEO investment is to treat the problem in layers, without expecting a single number that solves everything.

1. Segment by query type in Search Console: long, comparative queries, or ones that look like complex questions, are the most likely candidates to trigger AI Overviews or AI Mode. Tracking the evolution of impressions in this group, even without an associated click, gives a proxy for visibility.
2. Look at the click-through rate (CTR) per page, not just the site's aggregate. A CTR drop in a group of pages that maintains or increases impressions is a reasonable signal that the response is being satisfied before the click.
3. Cross-reference with Google Analytics to check whether the remaining "Web" traffic arrives with more intent (time on page, browsing depth), as Google itself suggests when commenting on the quality of AI Overviews clicks.

None of these layers replaces an "AI response citations" number that Google simply doesn't publish. They only reduce the size of the blind spot.

The controls the site still has in hand

The documentation reinforces that, since AI is integrated into Search, robots.txt directives for Googlebot remain the central control over crawl access. To specifically restrict what appears as a content snippet, Google points to nosnippet, data-nosnippet, max-snippet, or noindex.

For those who want to prevent content from feeding model training or being used as grounding in other Google systems, the documentation points to the Google-Extended token, kept separate from the standard Googlebot crawling directives used for Search.

For those who have set up any of these controls and still see content appearing where it shouldn't, Google recommends three steps: confirm with the URL Inspection tool that Googlebot is receiving the expected HTML, wait for the recrawl (which can take anywhere from days to months, depending on how often the site updates), and, if the problem persists, open a thread in the Search Central help community.

What remains open

The documentation doesn't promise, and couldn't, a native indicator of how many times a site appeared cited in an AI response without generating a click. Until that number exists, any AI Overviews or AI Mode impact a growth team presents internally will be an estimate built from cross-referenced data, not a ready-made Google report. It's worth treating it as such when reporting results to whoever decides the budget.