Blog · Engines

How Google AI Overviews chooses its sources

Published 4 September 2026 · Updated 8 October 2026 · 9 min read

How Google AI Overviews chooses its sources
Photograph: Tahamie Farooqui / unsplash

The short answer: Google AI Overviews do not pick a winner from the ten blue links. A model splits your question into several related searches, Google's ordinary ranking systems retrieve pages from the Search index for each one, the model reads those pages, writes an answer, and then attaches links to the pages that support what it wrote. Three gates decide whether you are even a candidate: the page is indexed, it is allowed to show a snippet, and your site is still switched on for generative AI features.

That last gate is new. Nothing in the ordinary search results changes when it is closed, which is what makes it easy to miss.

The pipeline, in the order it runs

Google's own optimisation guide names the two mechanisms rather than hinting at them. One is retrieval augmented generation, which Google also calls grounding: core Search ranking systems fetch relevant, current pages from the index, and the model reads the specific information on them before writing. The other is query fan-out, described as a set of concurrent related queries the model generates to go and get more.

Google's own worked example is a weed-choked lawn. Behind that one question sit several others the user never typed: which herbicide to buy, how to clear weeds without chemicals, how to stop them returning next season. Each gets its own retrieval pass. Each pass can surface pages you would never find by reading the rank report for the phrase the user actually entered.

So the citation list you see is a union of several result sets, filtered by whatever the model found useful while drafting. That is a different object from a ranking. It explains most of the confusion people have when a page sitting at position two is nowhere in the Overview above it.

Eligibility is a gate, not a ranking factor

Before any of the clever work happens, a page has to clear three conditions. Google states them plainly. It must be indexed. It must be eligible to appear with a snippet, which means no noindex, no nosnippet, no max-snippet:0, and no data-nosnippet wrapped around the part that answers the question. And the site must be included in Search generative AI features in Search Console.

Read that third one twice. Since Google shipped a Search Console control for appearing in AI features, there is now a switch inside your own property that can remove you from AI Overviews and AI Mode entirely, while leaving ordinary rankings untouched. Two paths to it being flipped are plausible, and we have not measured how often either one happens: a legal or brand-protection instruction given during the AI-scraping panic, and an agency changing the setting without telling anyone. Either way the symptom resembles losing on quality, and the fix is a setting rather than a content project.

Check it this week. Open Search Console, open the property settings, find the generative AI features control, and confirm the site is included. Then run the live URL inspection on your three most important pages and confirm they are indexed. Then view source, not the rendered DOM, and search for nosnippet. If your template inherited a data-nosnippet attribute on the main content block, the passage that answers the question is not available for a snippet, and Google's guide makes snippet eligibility a condition of appearing in generative AI features. Ordinary rankings carry on regardless.

Access problems sit upstream of everything else here. A chain of hops that strips the destination, for instance, fails quietly in exactly the same way, which is why redirect chains and lost sources is a dull topic that keeps paying for itself.

The page that ranks and the page that gets cited are drifting apart

Measurement work published in 2026 puts numbers on this. A longitudinal study issued 55,393 trending queries across 19 topical categories over a 40-day window, 13 March to 21 April 2026, and was accepted to the 2026 ACM Internet Measurement Conference. It found an AI Overview on 13.7 percent of those queries overall and on 64.7 percent of queries phrased as a question. Close to 30 percent of the sources it cited were absent from the ordinary first page shown underneath. Those are rates for trending queries in that window, not for your category.

Two practical readings follow. Phrasing is a real input rather than a detail, because in that sample a question-form query drew an Overview roughly five times as often as the overall rate. And a rank report is a weak proxy for citation, because close to 30 percent of the cited sources in that sample sat outside the results the rank report looks at.

Where do they come from? Partly from the fan-out queries, which have their own result sets. Partly from the kinds of domain Google reaches for when a question is factual rather than commercial. Pew Research Center analysed the browsing activity of 900 US adults who agreed to share it, covering Google searches made in March 2025 and published on 22 July

  1. Wikipedia, YouTube and Reddit together supplied about 15 percent of the sources listed

    2. in AI summaries against 17 percent of standard result sources, while .gov sites supplied 6 percent of summary sources against 2 percent of standard results. That is one US panel in one month, and it is the only first-party-grade number we can open on source mix. The government share, not the first three sites, is what is at least consistent with reference-grade sources being reached for ahead of vendors, which is the argument behind the third-party sources AI engines keep citing.

Citations are attached after the answer, not before it

Here is the mechanism most people get backwards. The model does not select a source and then write what the source says. It writes, and links are attached to support the statements it has already produced. Grounding is a quality control step, and the study below puts an error rate on it.

The same measurement study broke 98,020 atomic claims out of AI Overview responses and found 11.0 percent of them unsupported by the pages cited next to them, with omission rather than contradiction as the dominant failure mode. In plain terms: the Overview says something your page does not say, and points at your page anyway.

That changes what a win looks like. Being cited is not the goal if the sentence attached to your name is wrong, and the correction is not a link-building exercise. It is a writing exercise. The claim you want quoted has to exist on your page as a complete, self-contained sentence, near the question it answers, with the qualifier inside the sentence rather than two paragraphs below it. Omission was the dominant failure mode in that study, so a passage that needs its neighbours is the kind that tends to lose them.

This is also where AI Overviews and AI Mode start to part company, because the second one fans out harder and shows fewer links per answer. We cover that split in Google AI Mode: what changes for a publisher, and the cross-engine version in how each AI engine picks the sources it cites.

What you can actually measure now

Search Console added Search Generative AI performance reports on 3 June 2026 and finished rolling them out to sites worldwide on 31 August 2026. They show impressions from generative AI features in Search and Discover, broken down by page, country, device and date, with hourly granularity.

Note what is absent. No queries, no clicks, no click-through rate, no position. Google's announcement says additional metrics may come over time. So the report answers whether your URL showed up in an AI feature, and nothing about what happened next, which is honest and narrow and still the only first-party number on this surface.

Pair it with a prompt log you keep yourself. Write down the question as a buyer would type it, run it several times over several days, and record whether you were named, whether the claim was accurate, and which domains were cited beside you. Expect the answer to vary between runs and treat a single run as one observation rather than a reading. Ten questions, five runs each, is a weekend of work and gives you something a screenshot cannot.

If you want the structured version of that exercise, the self-check lists ten tests you can run this afternoon without buying anything.

Honest limits

Google publishes the mechanisms, not the weights. We know fan-out happens; we do not know how many queries it generates for a given question, how candidates are scored against each other, or how diversity constraints keep two pages from one domain out of the same answer. Anyone presenting those as known numbers is modelling, not reporting.

The research above is also a snapshot. Activation rates move with Google's own tuning, the 13.7 percent figure comes from 40 days of trending queries rather than commercial ones, and your category may behave nothing like the average. The source-mix figures come from a single US browsing panel in March 2025 and may not hold elsewhere or now.

Two things here are not measurements and should not be read as any. We have no count of how often the Search Console generative AI setting is switched off, only reasons it plausibly could be, so treat it as a check worth running rather than a common fault. And the prompt log described above is a method we recommend, not a study we have published: ten questions at five runs each is a sample of one site's surface on the days you ran it, never a census of what the engine does.

FAQ

Does ranking in the top ten get me into the AI Overview? It helps and it is not sufficient. Retrieval runs off core Search ranking, so rank is a real input, but close to 30 percent of cited sources in the 2026 measurement study were not in the organic results shown next to the Overview.

Do I need special structured data to be cited? No. Google's guide says structured data is not required for generative AI search and there is no special schema.org markup to add. Schema still earns its keep for rich results and for making your entity unambiguous.

Why does my page get cited for a claim it does not make? Because links are attached to support an answer that was already drafted. Omission is the common failure: the Overview compresses, loses a qualifier, and still points at you. Writing the qualifier into the same sentence as the claim is the lever you hold.

Can I see my AI Overview traffic in Search Console? You can see impressions by page, country, device and date from the generative AI performance reports. Queries, clicks and position are not in those reports today.

Should I block AI crawlers to protect my content? For Google, robots.txt and the snippet controls are what its documentation points to, and using them to shut the door also removes you from the AI answers Google writes. For a paywalled archive it may be right. For a product site it usually costs more than it protects, and it is a commercial decision rather than a technical default.

Sources

Related reading

How each AI engine picks the sources it citesGoogle AI Mode: what changes for a publisherThe third-party sources AI engines keep citingRedirect chains and lost sourcesGrok and social sourcing: when posts become citationsWhat a model update does to your visibility