How do Google AI Overviews choose which sources to cite?

Last updated October 14, 2026. Written by the Third Coast AI team.

Google AI Overviews choose their sources from the ordinary Google index, by running several related searches behind the one question a person typed, pulling candidate pages for each of those searches, and writing a short answer that links to the pages whose passages supported it, which means the first condition for being cited is the same as for ranking at all (indexed, eligible for a snippet, and readable in plain HTML) and the second condition is that one of your pages answers one of the hidden sub-queries better than the alternatives. Google's own documentation says there are no extra requirements, and the independent studies agree that being in the top ten helps but is no longer enough on its own.

This page sets out what Google has published about the mechanism, what the measurement studies found when they compared organic rankings to AI Overview citations, and the handful of page controls that decide whether a page is even eligible. It is written for an owner or marketing director who wants to know what is in their control before paying anyone, including us, to change it.

Does Google say how AI Overviews pick sources?

Partly, and the part it says is worth reading closely. Google's AI features documentation describes AI Overviews as a feature that helps people "get to the gist of a complicated topic or question more quickly, and provide a jumping off point to explore links to learn more," and it says both AI Overviews and AI Mode may develop a response by "issuing multiple related searches across subtopics and data sources." That second sentence is the whole mechanism in plain words: the system does not search for the phrase the person typed, it searches for a set of narrower questions underneath it and assembles an answer from the pages that came back for each one.

On eligibility, the same page is unusually direct. Pages need to be indexed and eligible to be shown in Google Search with a snippet, fulfilling the standard technical requirements, and Google adds that "there are no additional requirements to appear in AI Overviews or AI Mode, nor other special optimizations necessary." It also says you "don't need to create new machine readable files, AI text files, or markup to appear in these features," which closes the door on a good deal of what gets sold as AI optimization. Traffic from AI Overviews is reported in Search Console inside the ordinary Web search type, so there is no separate report to buy or build.

Which pages get cited, according to the studies?

The studies agree on the direction and disagree on the size. Ahrefs measured the overlap between Google's top ten and the pages cited in AI Overviews and watched it fall from 76% in July 2025 to 38% in March 2026, so a majority of citations now come from pages that do not rank in the top ten for the typed query. Kirsten and colleagues (Ruhr University and MPI-SWS, 2025) found 53% of the domains cited in AI Overviews were not in Google's organic top ten for the same query, and 40% were outside the top 100. Rank still helps, and Cyrus Shepard's May 2026 synthesis of 54 studies scores search rank at 9.4 out of 10 as a citation factor, second only to URL accessibility at 9.5, but the relationship is loosening every quarter and a page that ranks first for the typed phrase can rank nowhere for the hidden sub-queries that decide the answer.

The pages that do get cited share a shape, and the Princeton researchers (Aggarwal and colleagues, KDD 2024) measured the parts of that shape: adding quotations, statistics and cited sources raised a page's visibility in generated answers by roughly 41%, 33% and 28%. Sistrix's May 2026 citation drift study adds a reassuring note for Google specifically, because across 82,619 prompts tracked over 17 weeks, AI Overviews replaced only about 5% of their cited sources week to week, the most stable of the three platforms it measured, so a citation earned in AI Overviews tends to hold longer than one earned in ChatGPT.

What page controls decide whether you are eligible?

Google's robots meta tag documentation is the authoritative list, and three of the controls matter more than the rest because they are easy to set by accident in a CMS or a security plugin. The table below is our reading of that documentation.

The one that catches companies most often is max-snippet. Several SEO plugins ship with a default of max-snippet:-1 (unlimited), which is fine, but a well-meaning developer who sets a small number to tidy up search listings has also capped how much of the page AI Overviews can draw on. We check the rendered head of every page in an audit for exactly this reason.

Why can a competitor on page three be cited over you?

Because the competitor's page matched one of the hidden sub-queries and yours matched the typed phrase. When a buyer asks "best commercial roofing contractor for a flat roof in Grand Rapids," the engine goes looking for membrane types, warranty terms, local licensing, winter installation and reviews, and a page that states and proves three of those in self-contained passages will be retrieved for three searches while a page optimized for the headline phrase is retrieved for one. BrightEdge's tracking puts AI Overviews on nearly half of Google queries and Semrush measured a 71% rise on commercial searches between November 2025 and April 2026, so the number of buyer questions where this mechanism decides the answer is growing, not shrinking.

That is why our method starts with capturing those hidden searches rather than with a keyword list. We build a prompt bank the way buyers ask, run it with retrieval on, capture the sub-queries and the decision factors the engine uses, and then audit the client's pages for proof of each factor, which turns "get cited" into a list of specific sentences that are missing from specific pages. The decision factors page explains what those criteria look like in practice.

Should you change anything on your site today?

• Confirm every important page is indexed and carries no nosnippet or small max-snippet directive, by viewing the rendered source rather than the CMS settings screen.
• Put the direct answer to the page's question in the first 150 to 200 words, in plain HTML, with a number or a named source behind it.
• Break long pages into passages with question-form headings, one per sub-question a buyer would ask, so each passage can be retrieved on its own. The structure guide covers the pattern.
• Keep classic SEO health, because the index is the pool the engine draws from; a page Google cannot crawl cannot be cited.
• Measure before and after by running the buyer questions several times and recording who is named, since a single screenshot proves nothing (see why answers change between runs).

None of this requires a new file format or a new markup vocabulary, and Google says as much. What it requires is reading what the engine already cites for your topics and writing pages that are at least as useful, which is slower than installing a plugin and is the only version of the work we have seen move citations.

Frequently asked questions