How to check whether Claude knows about your company

Run three checks: your category questions without naming yourself, your company by name in a separate conversation, and a direct question about how many live searches it actually ran. The third one is what makes the first two mean anything, and it is the step almost everybody misses.

The blind check

Take ten to twenty questions your buyers ask, in their own words. Ask them without naming your company anywhere. Record who was named, in what order, and which sources were cited.

Naming yourself in the question is the single mistake that ruins this, and we have the two readings side by side. On 27 July 2026 the same 78 questions were run twice in one day for clawlaw.in, our sister company. The first run used a wrapper that named the brand, and ChatGPT came back ranking it first on almost every question. The second named nothing, and it 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. Three weeks later, on 18 August 2026, eighteen blind commercial questions made that company the top source on exactly one of the eighteen.

Use a fresh conversation for each question. An answer earlier in the same conversation shapes the answers after it, so ten questions in one thread is one long answer rather than ten readings.

The search check

In the same conversation, ask for how many of the questions it ran a live search, and which pages it retrieved. Write the number down next to the result.

If it searched for none of them, you have measured its memory, which cannot reflect anything you published recently. That is still useful, but only as a picture of the past, and it must be recorded separately from a live reading rather than mixed with it.

If you want a live reading, say so in the prompt: ask it to search for each question and to tell you next to each answer if it did not. Then check the reply against the instruction rather than assuming it complied.

The brand check

Separately, ask about your company directly. What does it do, what does it cost, who is it for, and are there any concerns. Read the concerns carefully rather than defensively. They are often accurate and they are what your buyers are being told.

Keep these answers out of your visibility number. A question naming you will almost always mention you, so including it produces a figure that can only look good.

What to write down, and what to ignore

Record the following six things and nothing else. The list is short on purpose, because a sheet with twenty columns does not get filled in.

• Date and search count. These two together are what make a reading comparable.
• The exact question, copied.
• Named or not, and in what position.
• Who else was named, in order.
• Which pages it says it retrieved, by URL.
• Any concern it raised about you, quoted word for word.

Ignore the tone. It is tempting to read a warm paragraph as a good result and a blunt one as hostility, but the wording changes between runs while the names and sources are the part that actually moves. If the concerns raised about you are accurate, they are the most valuable output of the whole exercise, because that is what your buyers are being told when you are not in the room.

A worked example

Two readings from our own work show why the search count is the column that decides everything else.

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. It confirmed afterwards that it had run no live web search for any of the 78. So the whole set was a reading of memory, and no page published in the weeks before could have appeared anywhere in it.

Then, in September 2026, on our own site aiknowsus.com, we asked Perplexity 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.

Two rules fall out of those readings, and they are the whole method. Never compare one reading with another unless the search counts sit beside both. And never take an engine's word for the count without checking it against the pages it says it retrieved.

Applied to a blind set you run twice, a month apart, that changes the report you write. If more questions were searched in the second month, then a rise in mentions is partly a change of instrument rather than a change in your position. So compare searched questions with searched questions, and unsearched with unsearched. Expect the unsearched ones to show no movement at all, because memory does not move in a month. The honest report is duller and it survives scrutiny: more questions were answered with a live search, the rate within searched questions moved this much or not at all, and there is no evidence of movement in the memory based answers. That points at the right work, which is making pages more quotable rather than making more pages.

Reading it honestly

Take the whole set more than once. Answers vary between runs, so one reading of ten questions is a rough indication and no more. Put a range around any rate you report, and if two readings overlap, say nothing moved even when the middle number rose.

What this check cannot do

It cannot produce a percentage you should put in a board pack from ten questions. It can produce a direction over several months, which is more useful and harder to sell.

It cannot tell you what other users see. Different accounts, surfaces and product versions behave differently, and an API reading is not the same as a consumer app reading.

It cannot tell you why you were left out, only which of three situations you are in: not reached at all, reached and not quoted, or quoted with something wrong. Each of those has a different fix, and narrowing to one of them is the point.

We build a tool that runs this kind of check at scale, so treat that as our bias. The method above costs nothing and works in a spreadsheet, and you should do it by hand at least once before you pay anybody for it, including us.

Common questions

How often should I run the full set?

Monthly. This engine answers many questions from memory, and memory does not change week to week, so weekly readings mostly produce identical results and false alarms.

Why did two runs of the same question disagree?

Check the search count first, because that explains most of it. After that, two searching runs can retrieve different pages, and generated text varies anyway. Three readings before you believe a result is a good habit.

It named a competitor's page as the source for something about us. What should I do?

Ask which page, open it, and check the fact. Wrong facts get a correction request with the right value and a link to your page. Correct facts mean you never published that fact plainly yourself, so publish it, with a date, on a page built for that question.

Is being cited the same as being recommended?

No, and on this engine the gap is wide. It often uses a vendor page for facts while recommending somebody else, because a page about yourself is read as advocacy. Record the two separately or you will misread your own progress.

Should I test through the API instead, to automate it?

You can, and label it clearly as a different surface. API behaviour and consumer app behaviour are not identical, particularly around when a search happens. Keep a few manual readings alongside any automation as a sanity check.

What do I do with the concerns it raises about my company?

Treat them as free research. If a concern is accurate, answer it on your own site plainly, including what you changed and when. If it is out of date, find the page it came from and get it corrected. Either way, that paragraph is the most honest competitive feedback you will read this quarter.

What to do first

Ask it about your own company by name, once, today, and read the concerns section carefully. Most people discover something there that is either wrong and fixable or true and unanswered, and both are worth more than another week of counting mentions.