How Google AI Overviews Pick Their Sources

AI Overview placement is already being sold as a package. It is worth knowing exactly how AI Overviews pick sources before you buy anything, because the mechanism rules out most of what gets sold. AI Overviews are generated by a Google model that is grounded in Google Search itself — the answer is written from pages Search retrieved, and the links shown are the pages the model drew on.

How do AI Overviews pick their sources?

They pick from what Google Search already retrieved. The system issues the user's question along with a set of related queries, gathers results for all of them, and generates a summary grounded in those results, linking the pages that supported it. There is no separate index, no separate submission process and no separate eligibility list.

That single fact settles several arguments. If a page is not indexed, it cannot be a source. If a page is indexed but never surfaces for a query or any of its near neighbours, it will not be a source either. Everything that makes a page retrievable in ordinary Search is a prerequisite here, which is why the boring technical work still comes first.

What is query fan-out, and why does it change your keyword thinking?

Query fan-out is the practice of answering one typed question by running several related searches behind the scenes and synthesising across all of them. It matters because it breaks the one-page-per-keyword habit that most SEO plans are built on.

Take "should I use Shopify or WooCommerce for a store in India". A fan-out on that question can reasonably be expected to reach for transaction fees, payment gateway support, hosting responsibility, developer availability and migration difficulty. A page that argues the headline comparison well but never mentions payment gateways simply is not in the candidate set for that strand — and a smaller, sharper page that covers gateways properly might be. Depth across the sub-questions beats one long page that gestures at all of them.

Practically, that means building topic coverage rather than keyword coverage: a cluster where each page owns a real sub-question, all of them linked, none of them duplicating the others. That is a different content plan from the one most agencies produce, and it is the plan behind our SEO work.

Is there special markup that gets you into an AI Overview?

No. Google has stated plainly that there is no additional structured data, meta tag or markup that makes content eligible for its AI experiences — normal Search eligibility is the requirement. Anyone selling "AI Overview schema" is selling something that does not exist.

Structured data still earns its place for other reasons: it disambiguates your organisation, products and articles, and it makes machine parsing of your page cheaper and less error-prone. We go through what is genuinely worth implementing in schema markup for AI search visibility. Just do not implement it expecting a switch to flip.

Can you keep your listing but stay out of AI Overviews?

Partly, and it costs you something. Google's snippet controls apply to AI experiences as well as to ordinary results, so the same directives that suppress a snippet also keep your content out of a generated summary.

• nosnippet — no text snippet at all is shown for the page, and it becomes unusable for a summary. You keep the blue link and lose the preview.
• max-snippet:[n] — caps the characters Google may show. Setting it to zero behaves like nosnippet; setting it to a small number limits what can be lifted.
• data-nosnippet — an HTML attribute that excludes a specific element rather than the whole page. This is the surgical option, useful for pricing tables or proprietary methodology you do not want summarised.

The trade-off is real. Those directives do not distinguish between a generated summary and a featured snippet, so opting out of one opts you out of the other. For most Indian service businesses that is the wrong trade — the preview is often what earns the click. For a publisher whose entire product is the text, the calculation is different. Note also that the Google-Extended control governs a different set of Google AI products and is not the lever for AI Overviews.

Which kinds of pages get pulled in most often?

Pages that answer a discrete sub-question cleanly, early, and in language close to how the question is asked. Beyond that, a few patterns show up repeatedly when you look at what gets cited:

1. Definition and mechanism pages. Content that explains how something works, rather than why you should buy it, is easier to ground an answer in.
2. Genuine comparisons. A page weighing two real options with stated trade-offs supplies exactly the shape of content a summary needs.
3. Pages with clean passage boundaries. One idea per section, answered in the opening sentence, with the elaboration below it.
4. Sites with corroborating identity signals. A recognisable organisation with consistent details across the web is a safer thing to cite than an anonymous site, which is the argument for entity SEO.
5. Locally specific content. For queries with an Indian context, pages that address Indian regulation, pricing structures or logistics are more useful to ground an answer in than generic international content — one reason our Gurgaon SEO work leans hard on local specificity.

How do you measure AI Overview performance in Search Console?

You largely cannot separate it. Google reports clicks and impressions from AI experiences inside the overall Web search totals rather than as a distinct type you can filter, so there is no clean AI Overview column to report on.

What you can do is read the pattern. Impressions holding steady while click-through rate falls on informational queries is the signature of an answer being shown above you. Segment your Search Console queries into informational, comparison and transactional buckets and watch the click-through rate of each separately — an aggregate number hides the whole story. We break down how to build that report in what to do about zero-click search, and the broader strategic shift in generative engine optimisation versus SEO.

What should you actually change on the page?

Five things, in this order. Make sure the page is indexed and renders server-side. Rewrite each heading as the question a person would type. Make the sentence directly under each heading a complete answer. Add the specific detail — the number that comes from a real source, the exact setting, the actual constraint — that separates your page from the rewritten consensus. Then interlink the cluster so the sub-question pages support each other.

None of that is a trick, and that is the point. The mechanism rewards content that is genuinely easier to verify and quote, which is also what a reader wants. If you would rather have the whole cluster planned and executed as one campaign, that is what our digital marketing team builds.