A reproducible Gemini visibility check for a business: prompts, scoring, and results

To check whether Gemini knows about your company, do not ask it whether it knows you. Write down your own facts first, then ask questions that would force those facts to come out, and score each fact as correct, wrong, missing or unverifiable. The result is a count: correctly identified facts out of the facts tested, on a stated date. We publish the fact sheet, the rubric and the whole procedure below, and we state plainly that as of 29 September 2026 we have not run this check on Gemini, so this page carries no Gemini accuracy count.

Everything else on this page is usable today, including the rubric, which is the part people get wrong and the part that makes two people's results comparable.

What "Gemini knows about your company" can mean

Four different things, and a business that has not separated them cannot say what is broken.

• Recognition. The assistant can say what your company is when the name is given to it. This is the weakest form and the one businesses test first, because it feels good.
• Fact accuracy. The specific things it says about you are true: what you sell, where you operate, what it costs, who it is for.
• Retrieval. It can find and cite your own pages when the question calls for current detail.
• Consideration. It names you unprompted when somebody asks for options in your category. This is the one with commercial value and the hardest to get.

A company can be recognised, described wrongly, never cited and never considered, all at once. Those are four separate pieces of work.

Before you test

Write the fact sheet first

This is the step that decides whether the test means anything. Before you ask a single question, list the facts you will score, each one with the page on your own site that states it. Twelve to twenty facts is a workable set, drawn from these eight kinds.

• What you sell, in one sentence a stranger would recognise.
• Who it is for, stated as a buyer type rather than a market.
• Where you operate, named places, not "pan India".
• Price, a figure or a band, with the date it applies from.
• Scope or coverage, counted and named: which courts, which states, which integrations, which certifications.
• Founding facts, the year, the founders, the registered entity name.
• Proof points, anything you claim publicly that a reader could check.
• Limits, what you do not do and who should buy something else.

Every fact needs a URL on your own site next to it. A fact you cannot point to on your own site is a fact you are asking the assistant to get right by luck.

Check that a machine can read those pages at all

Do this before testing, because otherwise you will be measuring a problem you already have. Fetch your own pages the way a crawler would, with scripts not running, and see what text is actually there. We have a dated case of what goes wrong: a pricing page rendered its prices only after scripts ran, so what a crawler received contained no prices at all, and the prices the assistants did quote had come from an app store listing instead of from the company's own site. Claude, 6 August 2026, on clawlaw.in.

Then look for contradictions in public

A fact that exists in two different versions in public is worse than a missing fact. On the same programme, the website and the app store listing carried different plan names and different prices for the same product, and ChatGPT noticed the contradiction and said so in its answer. ChatGPT, 6 August 2026, on clawlaw.in. Fix the contradiction before you test, or your test will just rediscover it.

The test prompts

Three groups, fifteen prompts, run in clean sessions with no uploads and nothing pasted in.

Group one: five named prompts, for recognition and fact accuracy

• "What is [company name] and what does it do?"
• "Who is [company name] for, and who should not use it?"
• "What does [company name] cost, and what is included?"
• "Where does [company name] operate, and what does it cover?"
• "What are the limits of what [company name] does?"

Group two: five blind prompts, for consideration

Your name appears in none of these. They are your buyers' selection questions in your buyers' words, and they are the only prompts that can tell you whether you are in the running. A prompt that names you cannot answer this question, which is why our own first blind run on another assistant replaced an earlier run that did.

Group three: five verification prompts, for retrieval

Questions that can only be answered correctly from a page on your site, and which should therefore produce a link if retrieval is working. Your current price, your current coverage list, your published policy on something, a specific named feature, and something you published in the last month.

Score answers and verify claims

The rubric, in four values

• Correct. The statement matches your fact sheet, with no material detail wrong.
• Wrong. The statement contradicts your fact sheet. Record the exact wording.
• Missing. The fact was called for by the question and did not appear.
• Unverifiable. The statement is too vague to be either, for example "affordable pricing" where you asked for price. Count these separately and never quietly as correct.

The headline figure is correct facts over the total facts tested, on a date, with the unverifiable count stated next to it. Then run one more column, which most tests skip: was the correct fact supported by a link. A fact that is right with no source behind it is right this week and unreliable next week.

Rules that stop the score drifting

• Score against the fact sheet, not against your impression of the answer.
• One fact, one row, even when one sentence carries three facts.
• Paste the quoted wording into the row. Scores without wording cannot be rechecked.
• Where the answer is partly right, score it wrong and write what was wrong. Half marks make the total meaningless.
• Have a second person score ten rows blind. If they disagree with you on more than a couple, tighten the fact sheet.

Repeat the test and compare results

Run the same fifteen prompts monthly, on the same weekday, and keep every log. Compare four things run to run: the correct count, the wrong count, which specific facts moved, and whether links appeared where they did not before. The movement of one named fact is more useful than the movement of the total, because it points at one page you can go and fix.

What we have measured on Gemini: nothing, as of 29 September 2026. We hold no Gemini fact log and publish no Gemini accuracy count. What we hold on other assistants is in the two dated cases above, on clawlaw.in on 6 August 2026, plus the September 2026 aiknowsus.com audit in which, across 24 batches and 72 conversations, the phrase recording that we were not cited appears 161 times in Perplexity's own self audits.

What the result does not prove

• A high correct count is not consideration. An assistant can describe you accurately when asked and never name you when a buyer asks for options. Group two is the commercial test, not group one.
• A wrong fact is not a fixed state. Answers vary between runs, so a single wrong answer needs to be seen twice before you treat it as the model's position.
• It cannot tell you where the error came from. That takes finding the public page that says the wrong thing, which is separate work and usually finds a listing, an old press item or your own outdated page.
• It does not transfer to other assistants, or to another person's account, or to a different country.
• A correction is not guaranteed to hold. We do not promise that fixing a page changes an answer, and we hold no dated measurement of correction durability to offer instead.

Common questions

Should I be worried if Gemini has never heard of my company?

Not in itself. Many small and mid-sized companies are absent from model memory and still get cited once they publish pages that answer specific buyer questions, because that is a retrieval problem rather than a memory problem. The first citation we can point to in our own programme came fourteen days after the page was published, when ChatGPT cited clawlaw.in/blog/how-to-check-a-companys-court-cases-in-india on 6 August 2026 for a vendor due diligence question. One case, one date, and not a promise.

Why does asking "is my company good" produce a nice answer?

Because a question that names you and invites approval gets approval. Our own first run of 78 questions used a wrapper naming the brand and came back ranking it first on almost every question. It was discarded, because the only thing it had measured was our own prompt. ChatGPT, 27 July 2026, on clawlaw.in.

How many facts should I test?

Twelve to twenty. Fewer than twelve and one wrong fact dominates the score. More than twenty and nobody finishes the scoring, which is worse than a short test finished properly.

What do I do about an "unverifiable" answer?

Treat it as your own page's failure first. If you asked for price and got "flexible pricing", check whether your own site states a figure a machine can read. Vague answers usually come from vague sources.

Is it cheating to test with my name in the prompt?

No, as long as the two groups stay separate and the named results are never reported as visibility. Named prompts are the right tool for fact accuracy. They are the wrong tool for whether buyers will meet you.

Worksheet, example dataset, and sources

Everything needed to run this is four sheets you can build in an hour: a fact sheet with a URL per fact, a prompt sheet with the three groups, a log with one row per fact per run carrying the quoted wording, and a folder of saved answers with their source lists. No tool is required.

Our dated observations come from the clawlaw.in programme from July 2026 and the aiknowsus.com audit of September 2026. The crawler-readable pricing failure and the two contradicting price lists are both from 6 August 2026 and both are checkable from outside, because they concern pages that were public.

Change log. First published 29 September 2026, with the Gemini count absent and marked absent. Disclosure: AI Knows Us sells AI visibility measurement, so treat the recommendation to measure as coming from an interested party.

What to do first

Write the fact sheet. Twelve facts, each with the URL on your own site that states it, in one sitting. Most companies discover in that exercise that three or four of their own facts are nowhere on their own site in a form a machine could read, and fixing that is worth more than the first test result.