Where Google AI Overviews gets its information

AI Overviews draw on pages from Google's own search index and link the ones they used underneath or beside the summary. That link list is the most useful thing on the screen for anybody trying to work out why they were left out, because it names the pages you are actually competing with.

What the linked sources usually are

The mix depends heavily on the question, and the pattern is fairly predictable once you look at a few. The following five kinds of page do nearly all of the work.

• Official and institutional pages for anything involving rules, registration, public records or definitions with a legal meaning.
• Publisher roundups and comparison articles for "best" and "top" questions, which is why those questions so often name companies that did nothing on their own site.
• Vendor pages and product documentation for specific, technical or feature level questions, where a vendor genuinely is the best source.
• Community threads for questions about real experience, problems and workarounds.
• Local listings and maps content for anything with a place in it.

The useful exercise is to work out which of those five dominates each of your own questions, because your next action is different for each one.

How to turn that list into a plan

Collect the linked sources for each of your top buyer questions. Count the domains. You will usually find three groups: sources you will never displace and should not try to, such as government pages; publisher pages you could realistically be included in; and vendor or documentation pages where a better page of your own could win.

Spend nothing on the first group. Spend outreach on the second. Spend writing on the third. Most businesses do the opposite, writing more pages aimed at questions that only ever resolve to official sources.

A worked example

Here is a real tally rather than a description of one. On 17 September 2026 we put six buyer questions to Claude for clawlaw.in, our sister company, with no brand named in any of them, and recorded every source that reached an answer.

Two findings came out of it. On the questions about how to look a case up, official court portals took every position above any commercial product, and the assistant said plainly that no commercial product should rank above the official portal for a question about using that portal. And across all six questions, no third party review site and no directory reached any answer at all. One well known review site did appear in the raw results and was discarded, because the list it offered was of American products and so was not an answer to an India question.

Read that as three decisions. Official pages are group one, and a page of yours competing with them on the rules is a page nobody will use, so stop writing those. Review platforms are not automatically group two: check whether they reach an answer in your own market before you spend a quarter on them, because in that run they did not, and one of them was set aside for being about the wrong country. Competitors' own pages are group three, and they tell you what a firm's own page has to contain to be used at all, so open them and note the specific facts they state, which is usually a fee, a timeline, a document checklist or an eligibility rule.

A community thread is the interesting case. Where one turns up in your tally, it is there because somebody described a real problem plainly and nobody else did. You cannot write that thread, and you can write the page that answers the same problem honestly, including what it costs to fix and how long it takes. That page has almost no competition, because most firms will not put a number or an unflattering detail on a public page.

How to get onto a source that is already being linked

This is the step people treat as impossible and it is mostly administrative. For a listing site or directory that keeps appearing in your tally, work through the following five steps in order.

• Find out whether you already have a profile. Many of these sites create entries from public company data, so you may have a listing nobody in your company has ever logged into. Search your own name plus the site's name before you fill in any form.
• Claim it rather than creating a second one. Two entries for the same company split whatever authority you have and produce contradictory facts, which is the exact problem you are trying to avoid.
• Complete every field that carries a fact. Category, locations, what you do, pricing model, and anything with a number in it. Empty fields are why a listing describes you in one line.
• Match your own wording. The description there should say the same thing as your website, in the same terms. A listing that disagrees with your site is a source of confusion rather than support.
• Put a reminder in the calendar to check it twice a year. These entries go stale silently, and a stale listing is where most wrong facts about a company come from.

For a publisher roundup the process is different and slower: read the article, work out whether you genuinely fit the criteria it used, and write to the author with the specific facts they would need, including a price and a date. Asking to be added because you exist does not work. Asking to be added with the four facts the article states about every other entry sometimes does.

Why your own page can still be the source

Vendor and firm pages appear regularly for the specific questions. What gets them chosen is being the clearest available answer with real detail: exact scope, real prices with dates, honest limits, documented steps. A page that reads as a brochure will not be used even when the vendor is the obvious authority.

Documentation deserves a separate mention. Help pages, technical docs and process guides are specific by nature and rarely exaggerate, which is exactly what a summariser wants. If yours are behind a login, your most citable asset is invisible.

A caution about reading this

The linked sources show what was used for that overview at that moment. Overviews do not appear for every query, they change, and two people can see different things for the same words depending on country, language, device and account. Take several readings across a few days before you treat any of it as your position.

Google also does not publish how the sources are chosen from the ranking set or how many are read, and it includes overview performance inside overall Search reporting rather than separately, so there is no report to check this against. Your own sheet is the only record you will have, which makes keeping it properly more important here than on engines that give you numbers.

Common questions

How often should I redo the source tally?

Monthly. The domain list is reasonably stable, and what moves is your own position within it. Weekly readings mostly record whether an overview was generated that day.

Why do two readings of the same question show different links?

Because the overview may be generated differently or not at all, and because location, language, device and account all affect it. Take several readings before you conclude that the source mix for a question has changed.

A competitor's page is linked for a fact about my company. What should I do?

Open it and check the fact. If it is wrong, request a correction with the right value and a link to your own page, and keep a dated copy. If it is right, publish it yourself in a liftable block on the page a buyer would land on, with a date, so that a better source exists.

Is being linked the same as being recommended?

No. You can be linked without being named in the summary text, and named in the text without being linked. Record both, plus whether you ranked in the ordinary results underneath, because that third one still brings visits.

Should I worry that the overview reduces clicks to my page?

It is a real effect and it is worth planning around rather than resenting. Informational questions are where a summary can satisfy the reader completely. Deciding questions, which involve comparing, pricing or contacting somebody, still need your page. Put your effort there.

Do the same sources get used across different questions?

Often yes, and that repetition is the most valuable thing in your tally. A domain linked across many of your questions is a source the summariser leans on in your category, and being present on it once can affect several questions at the same time.

What to do first

Pick five buyer questions, search them from a signed out browser, and paste every linked URL into a sheet. Sort by domain and count. Then draw a line: government and official pages get no effort, the most linked joinable source gets an email this week, and the vendor pages get read so you know exactly which facts your own page is missing.