Back to Blog
ai seo

How Google AI Overviews Choose Their Sources, and How to Become One

Nidhi Mevada

Nidhi Mevada

Founder, Brainito

October 8, 2026
7 min read
Article Insight

AI Overviews cite a few pages above every other result. What Google has said about how they are picked, what the cited pages share, and what to change.

What an AI Overview actually is

An AI Overview is the generated summary that appears above the regular results for some Google searches. It is written by a Gemini model, grounded in pages from Google's index, and it shows a set of links as its sources. Google launched it in the United States in May 2024 and has since expanded it to many more countries and languages. For the queries where it appears, it is the first thing the searcher reads, and the cited pages get a prominence that a position-one listing used to have.

Two points matter for anyone trying to be cited. First, AI Overviews are part of Google Search. They run on the same index and the same crawler, Googlebot, and the same quality systems apply. Second, they do not cite ten pages. They cite a few, and the cited set is often not identical to the top ten organic results. Something beyond classic ranking is being applied, and it is worth understanding what.

Where the sources come from

Google has not published a specification, but it has described the mechanism in enough places to piece together. For AI Mode, its more conversational search experience, Google explained a technique it calls query fan-out: the system breaks a question into several related sub-queries, runs them, and assembles the answer from the combined results. AI Overviews behave consistently with that description. The sources for an overview are frequently pages that rank well for one of the sub-questions rather than for the original query as typed.

That explains the pattern people notice. A page that ranks sixth for "best running shoes for flat feet" can be cited in the overview because it ranks first for "what is overpronation" or "do stability shoes help flat feet", which are the sub-questions the model asked. The overview is a composite, and each piece of it is retrieved separately.

The practical consequence is that you are not competing for one ranking. You are competing to be the best answer to any of the three to six questions a model would ask on the way to answering the user's question. Pages that answer one sub-question cleanly do better than pages that gesture at all of them.

What the cited pages have in common

Look at enough overviews in your field and the cited pages share a shape. They are not necessarily the longest or the highest authority. They are the ones where a specific passage answers a specific question without needing the rest of the page.

  • A direct answer under a heading that matches the question. The heading is "How long does a marketing audit take" and the first sentence answers it.
  • Self-contained passages. The paragraph does not depend on a definition three screens up. Pronouns refer to things named in the same paragraph.
  • Specifics. Numbers, ranges, named options, steps. Overviews quote facts, and vague pages have no facts to quote.
  • Recent and dated. For anything that changes, a visible updated date and current details beat an evergreen page from four years ago.
  • A clear entity behind the page. An author or an organization Google can identify, with consistent descriptions elsewhere on the web.
  • Nothing in the way. Content in the HTML, not injected after load, and no interstitial between the crawler and the text.

What to change on your pages

Start with the pages that already rank in the top twenty for the questions you care about, because those are the ones in the retrieval pool. For each, list the sub-questions a careful person would ask on the way to the main question. Then check whether the page answers each one in a passage that could stand alone. Usually it answers two of five and alludes to the rest.

Add the missing passages as their own sections, with the question as the heading and the answer as the first sentence. Convert comparisons into lists or tables with the dimensions named. Put the number in the sentence, not in a chart. Add an updated date and make sure it is honest. Add FAQ sections where the questions are the ones people type, and mark them up with FAQ schema so the structure is explicit.

Then check the plumbing. Brainito's SEO Suite runs the on-page and technical checks per page, including whether the main content is server-rendered and whether schema validates, and the AI SEO module adds the AI-specific checks and tracks whether Google's AI results, along with ChatGPT, Perplexity and Gemini, actually cite you for your prompts week to week.

Mistakes that keep pages out of overviews

The most common one is a robots.txt change made for a different reason. Google-Extended is the user agent token that controls whether your content is used for Gemini training and grounding in Gemini apps. Blocking it does not remove you from Search or from AI Overviews, which use Googlebot. But some sites, in an attempt to stay out of AI entirely, have added rules that also restrict Googlebot or that block the rendering resources it needs. Read the file and test the page in Search Console's URL inspection before assuming you are eligible.

The second mistake is writing for the overview instead of the reader. Pages stuffed with question headings and one-line answers to every conceivable variation read as thin, and Google's quality systems still apply. Answer real questions fully and let the structure make the answers findable.

The third is ignoring the rest of the web. Overviews lean on sources Google already trusts. If your brand has no mentions, no reviews and no coverage outside your own domain, a competitor with the same page quality and a visible reputation will be chosen. Citations are earned off-site as much as on.

How to know whether it is working

Search Console reports AI Overview impressions and clicks inside the regular performance data without separating them, so it cannot tell you directly. The reliable method is a prompt set: a fixed list of the questions your buyers ask, checked on a schedule, with the cited domains recorded each time. Over a few weeks the set shows which questions you appear for, which a competitor owns, and which have unstable sources, where a better passage can win quickly.

Run it monthly by hand if you must, or weekly with a tracker. Either way, pair every change you make to a page with a check the following week. AI Overviews reward the page that answers the sub-question best, and the only way to know you are that page is to look.

Share

Link copied

Nidhi Mevada

Nidhi Mevada

About the Author

The Brainito team consists of marketing experts and data analysts dedicated to helping businesses grow. We combine human expertise with AI-driven insights to create actionable marketing strategies that deliver measurable results.

Ready to Transform
Your Marketing?

Get your personalized AI-powered marketing strategy today and start growing your business with data-driven clarity.

Get Your Marketing Plan