How to check whether Gemini knows your business

Run two separate sets of questions. A blind set that never names your company tells you whether Gemini recommends you. A small named set tells you what Gemini believes about you. Keep the two apart and never mix them into one score.

What Gemini does differently when you test it

Gemini can ground an answer in Google Search and can also answer from training alone, and you often cannot tell which happened until you look for links under the answer. It also sits inside a Google account. Chat history and personalisation can colour what you get, so a test run in your own everyday account is a test of your account as much as of the world.

Two habits fix most of that. Use a fresh chat for every question, and where you can, use a signed out or separate browser profile. Also note that Gemini on the web, Gemini in Workspace and Gemini on a phone can answer differently, so record which surface you used. A screenshot from a colleague's phone is not comparable with your desktop reading, and treating the two as one number is how teams talk themselves into progress that did not happen.

The two sets of questions

The blind set. Ten to twenty questions in the words a buyer would actually type, with the buyer's context and none of your branding. "Best payroll software for a company with fifty staff in India." "Alternatives to the tool we use now." "How much should I expect to pay for this kind of software." If your name appears in the question, the answer is about your question, not about you.

That is not a theoretical risk. On 27 July 2026 the same 78 questions were run twice in one day for clawlaw.in, our sister company. The run whose wrapper named the brand had ChatGPT ranking it first on almost every question. The run that named nothing put the company second by breadth and absent altogether from the litigation due diligence questions it most wanted to win. The first run was discarded, because the only thing it had measured was our own prompt.

The named set. Five or six questions that do name you: what does this company do, who is it for, what does it cost, is it any good, who owns it. These are not a visibility measurement. They are a fact check, and they are how you find out that Gemini has your old pricing or has confused you with a company of a similar name.

The six columns your sheet needs

One row per question per run, and the following six columns. Fewer than these and you will not be able to explain your own result in a month.

• Date and surface. Gemini on the web, in Workspace, or on a phone. They differ.
• The exact question, copied, not paraphrased. A reworded question is a different question.
• Were you named, yes or no, and where in the answer: first sentence, middle of a list, or a passing mention at the end.
• Who else was named, in the order they appeared. This is your real competitive set, and it is usually not the one in your pitch deck.
• Every link shown, by full URL. Domains alone lose the detail you will want later.
• Whether links appeared at all. This is the grounded or not grounded column, and it is the one everybody forgets.

How to run the check

• Ask each question at least twice, on different days. Answers vary between runs on the same question, so one run tells you very little.
• Copy the answer text, not a summary of it. You will want the exact wording later when something changes.
• Do the same questions on at least one other assistant. They read different sources and disagree often, and knowing where you are weak matters more than an average.
• Where the Gemini app offers a way to check the answer against Search, use it, and note what it flagged. It is a free second opinion on which parts of the answer have a page behind them.
• Do not edit your questions between runs. If a question is wrong, add a new one and keep the old one going, or you lose your ability to compare.

A worked example

Take a reading we actually have. On 18 August 2026 we put eighteen blind commercial questions to ChatGPT for clawlaw.in, our sister company, with the company name kept out of every question. It was the top source on exactly one of the eighteen. On that one, the answer named its comparison page and said plainly that the page was still the vendor's own editorial page.

That is two findings in one line and they point at different work. Being the top source once in eighteen is a reach problem: on the other seventeen the answer was built out of somebody else's pages. Having your comparison page named as the vendor's own is a trust problem: you were read, and then discounted for being the interested party. A sheet that records only "named or not named" hides the second finding completely, which is why the columns above include who else was named and which links appeared.

There is a second habit worth taking from our own audit. In September 2026, on our own site aiknowsus.com, we asked Perplexity afterwards how many of the questions it had really searched for. It withdrew its earlier statement and said it could not honestly substantiate having run a live search for each one, and that happened in three separate batches. So treat the links shown under a Gemini answer as your evidence, and treat anything the assistant says about its own searching as a claim to check rather than a number to record.

Then run the same question twice, a week apart, and fill in the links column both times. A run with no links was answered from memory, so it says nothing about the pages deciding that question today. A run with links tells you which pages Google surfaces and which firms those pages name, and that half you can work on this week. Without that one column the two runs look like one flat result.

What the result tells you

There are only three useful outcomes and each points at different work.

You are not named and your pages are not in the links. Gemini has not reached you at all. Check indexing and rendering first, then get onto the third party pages it is citing.

Your page is in the links but you are not named in the text. It read you and had nothing worth repeating. That is a writing problem: the page is hedging where it should state something specific and checkable.

You are named with something wrong. Now the named set matters. Find where that fact lives, correct it there, and make the correct version easy to find in plain text with a date on it.

What this check cannot tell you

It cannot give you a percentage you should trust. Ten or twenty questions asked twice is a small sample, and a rate calculated from it moves on its own. Put a range around any figure and if two readings overlap, report no change.

It cannot tell you what other people see. Your account, your location, your language and your history are all part of the reading. A signed out profile gets you closer to a neutral view but not all the way.

It also cannot tell you why you were left out. It can only tell you which of the three outcomes above you are in, which narrows the question from "why are we invisible" to something you can actually work on. We build a tool that runs this kind of check at a larger scale, so treat that as our bias, but the method works the same way in a spreadsheet and costs nothing.

Common questions

How many questions do I really need?

Start with ten, because ten you will actually finish. Grow towards fifty as you learn which questions your buyers ask, and keep the original ten in the set so you can compare across time. A set that changes every month is not a measurement.

Why did two runs of the same question disagree?

Usually because one was grounded in Search and the other was not, and those two draw on different information. Even two grounded runs can retrieve different pages. Personalisation and location add more variation. This is normal and it is why you take several readings and watch the direction rather than the jump.

Should I use my own Google account or a clean one?

A clean or signed out profile for the blind set, because you want something close to what a stranger sees. Your own account is fine for the named set, since you are fact checking rather than measuring. Whichever you use, write it in the sheet.

Gemini cited a competitor's page for a fact about us. Is that bad?

It is common and it is fixable. It means that on the question asked, their page was the more reachable or clearer statement of that fact. Publish your own version plainly, on a page built for that question, dated, in text. If their page states something about you that is not true, ask them to correct it and keep a copy of what it said.

Is being in the link list the same as being recommended?

No. Keep two columns. Being linked means the mechanics are working: you are indexed, fetchable and relevant. Being named in the text means the page gave the model something worth repeating. Moving from the first to the second is a writing job, and it is usually a matter of replacing adjectives with facts.

Can I automate this with a script that calls an API?

You can, and you should know the catch: an API reading and a consumer app reading are not the same surface, and the grounding behaviour can differ. If you automate, keep a small set of manual readings alongside it as a sanity check, and label everything with the surface it came from.

What to do first

Write down ten questions in your buyers' words, in a sheet, with the six columns above. Ask the first three in a fresh Gemini chat today and fill in the rows, including the links column. Then read your own sheet and decide which of the three outcomes you are in for each question. That decision, per question, is the whole point of the exercise, and it takes one afternoon.