How Do Google AI Overviews Choose Their Sources?

Google says AI Overviews and AI Mode pick their supporting links using its core Search ranking systems, and that a page only needs to be indexed and eligible to show with a snippet. There’s no extra markup or file to add, though nothing guarantees you’ll be chosen.

That’s a less exciting answer than the AI search industry would like. It’s also the one in Google’s own documentation, and independent data now adds a twist: cited pages are less tied to page-one rankings than they were a year ago.

We read Google’s pages on 3 October 2026 and flag where they differ in date. If you’d like a second opinion on how your site fares, see our AI search service.

What Does Google Say a Page Needs to Appear?

To be a supporting link in AI Overviews or AI Mode, Google says a page must be indexed and eligible to be shown in Google Search with a snippet, and it lists no additional technical requirements. Its newer guide adds one setting you control in Search Console.

What does the December 2025 page say?

The rule comes from Google’s AI features and your website page, last updated on 10 December 2025. It says there are no additional requirements to appear in AI Overviews or AI Mode, and no other special optimisations necessary.

Google also says that meeting every requirement doesn’t mean it will crawl, index or serve a page. Indexing and serving aren’t guaranteed.

What does the July 2026 guide add?

A longer guide to optimising for generative AI features, last updated on 10 July 2026, repeats the eligibility rule. It adds that a site must also be included in Search generative AI features in Search Console to be eligible.

The setting defaults to included for all properties. The December page predates it and doesn’t mention it.

What does Google recommend in practice?

The list is familiar SEO:

• Allow crawling in robots.txt and by any CDN or hosting infrastructure.
• Make content easy to find through internal links.
• Provide a great page experience.
• Keep important content in textual form.
• Support text with high-quality images and video.
• Make structured data match the visible text.
• Keep Merchant Center and Business Profile information up to date.

How Do Retrieval and Query Fan-Out Choose the Pages?

Google says its generative AI features use retrieval-augmented generation to pull pages from its Search index, and query fan-out to run several related searches behind one question. Both are described in its guide.

What is retrieval-augmented generation?

Google describes retrieval-augmented generation, also called grounding, as a technique that improves the quality, accuracy and freshness of AI responses by relying on its core Search ranking systems to retrieve relevant, up-to-date pages. The model then reviews the specific information on those pages and shows prominent, clickable links to the ones that support its response.

In plain terms, the ranking systems you already know supply the candidates, and the model writes from them.

What is query fan-out?

Google calls it a set of concurrent, related queries generated by the model to fetch additional relevant results. Both AI Overviews and AI Mode may use it, and Google says it lets them show a wider and more diverse set of links than a classic search.

Step	Example from Google’s guide
What the user types	“how to fix a lawn that’s full of weeds”
Fan-out queries the model might add	“best herbicides for lawns”, “remove weeds without chemicals”, “how to prevent weeds in lawn”
Pages that could be linked	Any page that answers one of those narrower questions well

What does fan-out mean for your content?

A page can be chosen for a sub-question the user never typed. That favours pages that cover a topic’s follow-up questions properly.

Google warns against overdoing it. Creating separate content for every variation or fan-out query, primarily to manipulate rankings or AI responses, violates its scaled content abuse spam policy, and the guide says it’s an ineffective long-term strategy anyway.

Do You Need to Rank in the Top 10 to Be Cited?

Not as closely as you used to. Ahrefs reported that about 76% of cited URLs ranked in the top 10 results in July 2025, and about 38% in March 2026.

That makes classic rankings a weaker predictor of citation than they were a year ago.

Study	Date	Sample	Cited URLs also in the top 10
Ahrefs, first study	21 July 2025	1.9M citations from 1M AI Overviews (top 3 most visible citations in each)	About 76%
Ahrefs, update	2 March 2026	863K keyword SERPs and 4M AI Overview URLs	37.9% across all SERP blocks, 37.1% for organic listings only

Why do the two studies differ?

Ahrefs says it improved its parsing methodology between the two, so it can see more of the citations in AI Overviews, and the sample more than doubled. It also notes that from January 2026 AI Overviews are powered by Gemini 3.

So the drop isn’t a clean before-and-after of Google’s behaviour. It’s still the best independent data we found, with one caveat: Ahrefs sells an AI visibility tool, and Google hasn’t commented on the figures.

What did Ahrefs think changed?

In the update, 31.2% of cited URLs ranked 11 to 100 and 31.0% sat beyond the top 100 blocks. Ahrefs reads that as AI Overviews relying less on the original query’s results and more on pages from fan-out queries’ results.

That’s a hypothesis, not a Google statement. Google confirms fan-out exists, but its documentation doesn’t say how fan-out results are weighed against the original ones.

Does video play a part?

Ahrefs found that among cited pages outside the top 100 for the same keyword, 18.2% were YouTube URLs, which was 5.6% of all AI Overview URLs in its data. Google’s own guide says generative AI features can bring in relevant images and video, giving your pages more ways to appear.

What should you do with this?

Rank for the sub-questions as well as the head term, and give the topic a proper treatment on a page that deserves to be found. If video suits your subject, treat it as another route in rather than a separate project.

Related: What Is Generative Engine Optimisation (GEO)?

Do You Need Special Markup or Files?

No. Google says you don’t need new machine-readable files, AI text files or markup, and there’s no special schema.org structured data to add.

Its July 2026 guide goes further and names specific myths.

Idea	What Google’s guide says
llms.txt and “special” files	Google Search doesn’t use them; having one neither harms nor helps
Chunking content into small pieces	No requirement; its systems understand multiple topics on a page
Writing in a special way for AI	Not needed; systems understand synonyms and meaning
Inauthentic mentions across the web	Not as helpful as it might seem
Structured data	Not required for generative AI search; still useful for rich results

What about llms.txt?

The guide says it’s fine to create llms.txt files for other services that use them, and that doing so won’t affect your visibility in Google Search. Our view of the file is in the guide below.

Related: What Is llms.txt and Does Your Site Need One?

What about schema markup?

Google says structured data isn’t required for generative AI search, but it’s a good idea to keep using it because it helps with eligibility for rich results. The one rule that applies to AI features is that your structured data should match the visible text, and schema markup still earns its place for other reasons.

What about chunking and page length?

The guide says there’s no ideal page length and that shorter or longer pages can both work depending on your audience. Make pages for readers, not for the format.

Do other engines say the same?

Not entirely. Microsoft’s October 2025 guidance for its own AI answers recommends schema, clear headings and Q&A formats, so cheap structure that helps readers is worth keeping even where Google doesn’t require it.

What Does Google Say It Rewards?

Google says creating unique, non-commodity content will likely influence your presence in generative AI search over the long run more than any other suggestion in its guide. Everything else is largely SEO hygiene.

What counts as non-commodity content?

Google’s example of commodity content is a “7 Tips for First-Time Homebuyers” listicle based on common knowledge. Its example of the opposite is a first-hand account of waiving an inspection and what it saved.

The common thread is a unique point of view from real experience, which is the territory of E-E-A-T and good content marketing. Google adds that content that only summarises what’s already online restates information available elsewhere.

Do the technical basics still count?

Yes. The guide lists meeting the Search technical requirements, following crawling best practice, using JavaScript carefully, reducing duplicate content and providing a good page experience, which includes the speed measures in Core Web Vitals.

It also says perfect semantic HTML isn’t required, though it helps people using screen readers.

What about local and product details?

Google says generative AI responses can include product listings and information about local businesses. Merchant Center feeds and a Business Profile can help them appear, and our guide to how Google ranks local businesses covers the profile side.

Want to know whether your pages are eligible and being cited in Google’s AI features? Talk to us about AI search for your business.

How Can You Control Your Content in AI Features?

Use the same preview controls as Search (nosnippet, data-nosnippet, max-snippet or noindex), plus the Search generative AI control in Search Console. Robots.txt for Googlebot manages crawling, because Google says AI is built into Search.

Control	What Google says it does
nosnippet, data-nosnippet, max-snippet	Limits the information shown from your pages in Search
noindex	Blocks your content from appearing in Google Search completely
Search generative AI control	Excludes a property from AI Overviews, AI Mode and generative AI features in Discover
Google-Extended	Limits AI training and grounding in some of Google’s other systems

How does the Search generative AI control work?

Google’s help page says it rolled the control out to all websites worldwide as of 31 August 2026. You’ll find it under Settings, then Search generative AI, in Search Console.

There are three choices. Include is the default, Exclude removes your links and content from those features, and Inherit follows a parent property and is the default for properties that have one.

What happens if you exclude your site?

Your links won’t appear, your content can’t be used as an input to generate an AI response or preview in those features, and you won’t receive impressions or traffic from them. Content from other sites will still appear and may look similar to yours.

Google says a change generally takes a few days, with content excluded within one to two days after the control goes live, although some may take longer. Think hard before you opt out, because you’d leave those features entirely.

Does the control affect AI training?

No. Google says the control doesn’t affect AI training, and points to Google-Extended for limiting the training of the models behind these features.

Whether to block training crawlers is a wider content decision, covered in should you block AI crawlers.

What if the controls don’t seem to work?

Google suggests using the URL Inspection tool to confirm Googlebot sees your control, then allowing time to recrawl. It says crawling can take anywhere from several days to several months.

How Do You Measure AI Overviews in Search Console?

Google’s position on reporting changed between December 2025 and July 2026: the older page says AI features sit inside the Performance report, and the newer guide points to a dedicated Generative AI performance report. Both are still live, and you need both to read your data.

What does each Google page say, and when?

Page	Date	What it says about reporting
AI features and your website	Updated 10 Dec 2025	AI features are included in overall search traffic, reported in the Performance report under the Web search type. No mention of a separate AI report
Optimizing for generative AI features	Updated 10 Jul 2026	Use the Generative AI performance report to measure performance in generative AI features on Search and Discover
Generative AI performance report (help)	Rollout note dated 31 Aug 2026	Rolled out to all websites worldwide; covers impressions for AI Overviews and AI Mode

As we read them on 3 October 2026, the two aren’t in conflict. AI feature traffic still counts in the Web search type, and the new report breaks out the impressions.

What does the Generative AI performance report show?

It shows impressions for AI Overviews and AI Mode, grouped by page, country, date or device, with a filter for text-based and multimodal searches. The usual Search limits apply, including the 1,000-row cap, and dates follow Pacific Time.

The chart aggregates by property, so two results from your site in one AI response count as one impression, while the page table aggregates by page. Experiments in Search Labs aren’t included.

The help page describes impressions only. It doesn’t list a clicks or position column, or a query dimension.

How are clicks, impressions and position counted?

Google’s Search Console help page on impressions, position and clicks sets out the rules for the Performance report:

Feature	Click	Impression	Position
AI Overviews	Clicking an external link counts	Standard impression rules apply	One position for the whole overview; every link in it gets that position
AI Mode	Clicking an external link counts	Standard impression rules apply	Follows the same methodology as a Search results page

A follow-up question in AI Mode counts as a new query, with its own impressions, position and clicks.

What does that mean for reading your numbers?

A page cited in an AI Overview takes that overview’s position, not a rank within it. On that reading, a shift in average position could come from an AI Overview appearing, not only from your page moving.

Google also warns against third-party tools that claim to use its internal metrics, saying none has access to its ranking or AI systems. For the wider routine, see our guide to measuring AI visibility.

Related: How Do You Measure AI Search Visibility?

What Is Still Unknown?

Google hasn’t published how it weighs sources within an AI Overview. It has said what makes a page eligible and what it values in general, not why one eligible page beats another.

Which details are unpublished?

The exact criteria for choosing supporting links, how often fan-out is used and how AI Mode and AI Overviews differ in practice. Google says only that the two may use different models and techniques, so the set of links they show will vary.

How stable are the answers?

Not very, on the evidence we have. SparkToro’s January 2026 study of ChatGPT, Claude and Google’s AI features found a less than 1 in 100 chance of the same list of brands in any two responses.

That study used brand-recommendation prompts rather than informational ones, so it doesn’t prove the same for every AI Overview. It does mean one screenshot proves little.

How do you keep up?

Check the date on any Google page before you rely on a guide, ours included. This area changed more than once between December 2025 and August 2026.

If you’d like an outside check of how your pages perform, an SEO audit covers crawling, indexing and snippets, which are the eligibility basics.

Need help turning Google’s documentation into a plan for your site? Explore Utterly Digital’s AI search service.