Does Gemini use live search or memory

Both. Gemini can ground an answer in a live Google Search, and it can answer from what the model learned in training. Which one happens depends on the question, the surface and the model version, and Google does not publish the rule.

What Gemini does differently

Because Gemini is Google's assistant, its live half runs through Google's own search systems rather than a third party search provider. That makes it the engine where classic Search work carries over most directly.

It also means the live half is limited by Google's index. Grounding is not the live web, it is pages Google has already crawled. A page you published this morning is not usable until it has been crawled and indexed, which takes as long as it takes.

Google publishes a robots.txt token, Google-Extended, that controls use of your content for Gemini training and grounding without affecting Google Search. That is a real lever, and it is one people block by accident when they copy a robots.txt from somewhere else.

The six signs that tell you which one you got

You cannot be certain, but the following six signs are reliable enough to sort your readings into two piles.

• Links or source chips under the answer. The clearest sign that pages were used.
• An offer to check the answer against Search. A grounded answer usually comes with a way to verify it.
• Specific recent detail: a date, a version number, a price that matches your current page.
• No links at all, with confident general statements. This is the usual look of a memory answer.
• A price or a product description that is exactly one version out of date. A very strong sign of memory, because a stale fact that was once true came from training, not from today's web.
• A direct test: ask what a named page says today. Memory cannot answer that, and it will either hedge or be wrong.

What that means for whether you are named

These are two different jobs with different timescales.

The memory half reflects how much was written about you, on how many independent sites, before the model was built. You cannot edit it and you cannot rush it. You can only keep adding to the record so the next model knows more.

The live half is pages, and pages you can change. If your buyer's question triggers grounding, then being on the pages Google ranks for that question is the whole game. That includes third party pages, which are often easier to get onto than they are to outrank.

A worked example

Two readings from our own work show why this one column matters more than the mention count.

On 27 July 2026 we put 78 buyer questions to ChatGPT for clawlaw.in, our sister company, with no brand named in any of them. ChatGPT then confirmed that it had run no live web search for any of the 78. So the whole reading, all 78 answers, described what the model remembered rather than what it could find that week. Any conclusion drawn from it about recently published pages would have been wrong, because no recent page was involved in it.

The second reading is about trusting an assistant's account of itself. In September 2026, on our own site aiknowsus.com, we asked Perplexity afterwards how many of the questions it had actually searched for. It withdrew its earlier statement, saying it could not honestly substantiate the claim that it had run a live search for each question, and in another batch that its claim to have searched all five was not adequately supported. That happened in three separate batches.

Put those two together and the method follows. Do not ask Gemini whether it searched and then write its reply down as a fact. Record whether links appeared, because that is the part you can see. Take your two or three most valuable questions three times each, in fresh chats, and keep two columns only: were you named, and were links shown. Questions that show links are being decided today by pages, so your job there is to be on those pages and a change can show up in weeks. Questions that never show links are being answered from memory, so nothing you publish this month will move them, and the realistic route is to publish the answer anyway so that it is available when grounding does fire.

The pattern can also flip from one month to the next. That is normal and it does not mean your work failed. It means the grounding decision changed, which is why you keep the links column rather than only counting mentions.

What to change

Make sure you are indexed and rendered as plain text. Answer one buyer question per page, with the answer at the top. State facts that can be checked, with dates, because a grounded answer needs something concrete to lift. Keep your description of yourself identical across your site and every listing, so the model is not choosing between versions of you.

Then do the slow thing as well, because the memory half is real. Being written about on sites you do not own is the only input to it. That means listings, directories, trade publications, podcasts, conference pages and anywhere else your name can legitimately appear. None of it pays off this quarter and all of it compounds.

What nobody can tell you

Google does not publish when grounding fires, how many pages it reads, how it chooses between them, or how often the grounding behaviour changes. Any article that gives you a rule is reporting a pattern it saw, not a documented behaviour, and patterns here expire.

There is a second limit worth stating plainly. Even a perfectly grounded answer is not a promise. Two people in different countries, on different surfaces, in different accounts, can get different answers to the same words on the same day. So the honest unit of measurement is a direction across many readings, not a position.

How to tell whether it worked

Track two columns, not one: whether you were named, and whether the answer showed links. Those two numbers move for different reasons. If links appear more often over time and you are still not named, the problem is what your page says. If links never appear for your questions, grounding is not firing for that kind of question and the work belongs on the memory side, which means being written about in more places.

Whatever the pattern, no one can guarantee a position in an AI answer, and a measurement built on a handful of questions needs a range around it before it means anything.

Common questions

Can I force Gemini to search?

You can ask it to, and asking often helps, but you cannot rely on it. Write the instruction in the prompt, then check whether links actually appeared rather than trusting the reply. If you want a forced live reading, an engine built entirely around retrieval is the more honest instrument for that job.

Why do two runs of the same question disagree?

The most common reason on this engine is that one run was grounded and the other was not. After that, the retrieved page set can differ between two grounded runs, and your account, location and language all add variation. None of this is a fault. It is why a single reading is not evidence.

How often should I check?

Monthly for a full set, weekly for your two or three most valuable questions. Any faster and the variation swamps the signal. Keep every reading, because six months of monthly readings is a real picture and two readings a week apart is noise.

An engine cited a competitor's page for a fact about us. What now?

Treat it as a ranking of clarity rather than of truth. Their page was easier to find or easier to quote for that fact. Put the fact on a page of your own built for that exact question, state it in the first paragraph, date it, and make sure it is indexed. If their page is wrong about you, write and ask for a correction, and keep a dated screenshot.

Is being cited the same as being recommended?

No. On this engine the two are especially easy to confuse, because a grounded answer can link your page while naming somebody else in the text. Record them as separate columns. A link means you are reachable and relevant. A mention means your page was worth repeating.

Does it matter that Gemini appears in Docs and Gmail too?

For measurement, yes. Those surfaces work on different material, including the user's own documents, so an answer there is not comparable with the Gemini app. If your buyers use Gemini inside Workspace, test there separately and label the readings, rather than assuming one surface stands for all of them.

What to do first

Add one column to whatever sheet you already keep: "links shown, yes or no". Then re-run your three most valuable questions today and fill it in. You will learn more from that single column than from another month of counting mentions, because it tells you whether the work in front of you is a pages problem or a reputation problem.