Where Gemini gets its sources

Gemini answers from two places: what the model learned during training, and pages it pulls from Google Search while it is answering. When it uses Search it usually shows a set of links or a way to check the answer against Search. When it does not, the answer comes from memory and there is nothing to click.

What Gemini does differently

Gemini is Google's own assistant, so the live half of its answer runs through Google's search infrastructure rather than a third party search provider. That has three practical effects.

Grounding is not switched on for every prompt. A general question can be answered from training alone. A question about a company, a price, a launch or anything recent is far more likely to trigger a search. Google does not publish the rule that decides this, and it is not stable, so nobody can tell you exactly when it fires. Anyone who tells you they know the trigger is guessing.

The pages it can reach are mostly pages Google has indexed. If Googlebot cannot fetch and render your page, Gemini has very little to work with. This is the one engine where your ordinary Search health and your AI visibility are close to the same problem, which is good news if you have been doing search work for years.

Google publishes a separate control for the model side. The robots.txt token Google-Extended controls whether your content is used for Gemini model training and grounding. Blocking it does not remove you from Google Search. Blocking Googlebot removes you from Search, and takes the grounded Gemini answers with it. Behaviour also differs between the Gemini app, Gemini inside Workspace and Gemini on a phone, so a result in one place does not prove anything about the others.

The six kinds of source that turn up most

Across the buying questions we have watched, the link lists under a grounded Gemini answer are made of the following six kinds of page. The mix changes with the question, but the list itself is stable.

• Official and government pages, whenever the question touches a rule, a registration, a tax matter or a public record. These are the right answer and you will not displace them.
• Publisher roundups and "best tool" articles, for any question phrased as a choice between options. This is where the company names in the answer usually come from.
• Vendor pages, including yours, most often the pricing page, the documentation and any page that states scope plainly.
• Review and directory listings, which carry your category, your price band and sometimes a description you have never read.
• Community threads, for questions about what a thing is actually like to use and what goes wrong with it.
• Google's own company data, meaning your Google Business Profile and the details attached to it, for anything with a place or a phone number in it.

Notice that four of those six are pages you do not own. That is the single most useful fact on this page.

What that means for whether you are named

You are working on two different things at once, and they move at different speeds. The training half is slow, you cannot edit it, and it mostly reflects how widely your name was written about before the model was built. The live half can change this month, because it is just pages.

So the useful work is the live half. If a page that ranks for your buyer's question names you, Gemini can reach it. That page is often not yours: a category directory, a review listing, a comparison article. Being on those is usually faster than trying to outrank them, and it is the step most businesses skip because it feels like somebody else's job.

A worked example

Here is a reading from our own work rather than a scenario. 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. Across all six answers, not one third party review site and not one directory reached the sources at all. A well known review site did come up in the raw results and was dropped, because the list it offered was of American products and so was not an answer to an India question.

Read what that does to a plan. If your plan for getting found was to collect profiles on the big review platforms, that run says those platforms were not in the answers for an India question, and one of them was set aside for being about the wrong market. The pages that did the work were official pages and the companies' own pages.

Now take the same reading for yourself. Ask your single most valuable buyer question in a fresh Gemini chat, and write down the answer, the companies named in order, and every link shown. Then put each link in one of three buckets: a page you can never be on, a page you could be added to, and a page you could outwrite. A government page is bucket one, so ignore it. A roundup on a review site is bucket two, and getting listed is an email and a form rather than a content project. A competitor's pricing page is bucket three, and it tells you exactly what a page of yours would have to state to compete: the same facts, at least as clearly, with a date on them.

That one hour of work replaces a quarter of guessing. You now know which of your problems is an outreach problem and which is a writing problem, per question. And you know it for your own market, which is the part the review site in our run got wrong.

What to change

• One page per buyer question, with the answer in the first paragraph rather than the conclusion.
• Server rendered text. If the content only appears after JavaScript runs, assume some crawlers see an empty page.
• Leave Google-Extended allowed if you want to be in Gemini's answers, and decide that on purpose rather than by accident in a copied robots.txt.
• A plain facts page: legal name, what you sell, who it is for, how pricing works, where you operate, and the date it was last updated.
• Get listed on the third party pages that already rank for your category question.
• For anything local, keep your Google Business Profile correct, because the same company data feeds a lot of Google surfaces.

What this page cannot tell you

Google does not publish how Gemini chooses which pages to ground an answer in, how many it reads, or how it weighs one source against another. Everything above is drawn from what is visible in the answers and from the controls Google documents. Treat any confident account of Gemini's internal ranking, including one that sounds technical, as a guess.

Two more honest limits. The link list shows what was shown, not everything read, so being absent from it is weaker evidence than being present in it. And the same question can produce different links on different days, in different countries and in different Google accounts, so no single reading is your position.

How to tell whether it worked

Ask Gemini your buyer questions without naming your company. Naming yourself tells it what to talk about and ruins the reading. Run each question in a fresh chat, run it more than once, and write down two things: whether you were named, and which links it showed.

The link list is the more useful half. It tells you which sources Gemini reaches for in your category, and there is almost always one you had not thought about and are not on. Watch the direction across several readings rather than the jump between two, and remember that no one can promise you a place in an AI answer.

Common questions

How often should I check this?

Once a month for the full question set is enough, and once a week for the two or three questions that matter most to your revenue. More often than that and you are mostly recording the engine's own variation rather than your progress. Keep the dates, because the pattern over six readings is worth far more than any single one.

Why do two runs of the same question disagree?

Because two different things can vary. The model may ground one run in Search and answer the other from memory, and those are two different sources of truth. Even within grounded runs, the set of pages retrieved can differ. On top of that, Gemini sits inside a Google account, so your history and personalisation can colour what you see. Use a fresh chat, and where you can use a signed out or separate browser profile.

Gemini cited a competitor's page for a fact about us. What do I do?

This happens often and it is a signal, not an insult. It means your own page either was not reachable or did not state that fact clearly enough to be the better source. Do two things. First check that you state the fact plainly, on its own page, in text, with a date. Second read the competitor's page and be honest about whether it is simply clearer than yours. If their claim about you is wrong, write to them and ask for a correction, and publish your own corrected version so there is something to point at.

Is being cited the same as being recommended?

No, and keeping the two apart is one of the most useful things you can do. Your page can be in the link list while the answer recommends somebody else. That is progress worth recording in its own column, because it tells you the mechanics work and the wording does not. Being named in the answer text is the outcome that changes a buyer's mind.

If I block Google-Extended, do I disappear from Google Search?

No. Google documents Google-Extended as a control over use of your content for its generative products, separate from Googlebot, which is what feeds Search. Blocking Googlebot is the one that removes you from Search. The mistake we see is a robots.txt copied from a blog post that blocks more than the owner intended, so read your own file rather than trusting your memory of it.

Does a Google Business Profile matter if I am not a local business?

Less, but it is not nothing. It is a short, structured, Google owned statement of your legal name, your category and your contact details, and it is cheap to keep correct. If your profile says something you stopped doing two years ago, that is a wrong fact sitting in a place Google trusts.

What to do first

Pick your single most valuable buyer question. Ask it in Gemini in a fresh chat today, and write the answer and every link into a sheet with the date. Sort the links into the three buckets above. Then do the cheapest thing on the list: if you are missing from a roundup you could be on, send that email this week, because it is faster than anything you can write.