Does Claude know your company? A repeatable prompt and citation audit

Ask Claude what your company is and you will learn almost nothing. To find out whether it knows you, write your own facts down first, ask questions that force those facts out, then score every fact twice: was it accurate, and was it supported by a citation you can open. Those two numbers are different and must never be merged. As of 29 September 2026 we have not published an "X of Y fact checks were accurate and citation-supported" count for Claude, so this page gives you the fact sheet, the prompts, the rubric and the log format, plus the dated fact-level failures we have already found.

The reason for the two-column score is simple. A fact that is right with nothing behind it is right until something changes, and you will not be told when it stops being right.

The answer, first

Three findings that decide how you should read any result you get.

• Being described is not being cited. In our own programme the assistant's answer was shaped by our pages and our name arrived as an item in a list rather than as a linked recommendation. Claude, 17 September 2026, on clawlaw.in.
• Your own site may not be the source. 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.
• Contradictions get noticed and stated. 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.

How it was measured

Step one: the fact sheet, written before any prompt

Fifteen facts, each with the URL on your own site that states it. Draw them from these seven areas so the sheet is not all product description.

• What you sell, in one sentence.
• Who it is for, as a buyer type.
• Price, a figure or a band, with an effective date.
• Coverage or scope, named and counted.
• Entity facts: registered name, founding year, head office city.
• Proof points anybody could check.
• Limits: what you do not do and who should buy elsewhere.

Any fact with no URL next to it comes off the sheet or gets published first. You cannot fairly score an assistant on a fact you have not published.

Step two: the fifteen prompts

• Five named prompts, asking directly what your company is, who it is for, what it costs, what it covers, and what its limits are.
• Five blind prompts, your buyers' selection questions with your name absent, which is the only group that can tell you whether you are in the running.
• Five citation prompts, each answerable correctly only from a specific page on your site, including one page published within the last month.

Clean session per prompt. Nothing uploaded, nothing pasted, no follow-ups in the same thread. Three repeats per prompt on three different days, because a single answer is a sample of one.

Step three: score each fact twice

One row per fact per run, with these columns: the fact, the quoted wording from the answer, accurate or not, citation attached or not, citation opened and supports the fact or not, and the run date.

The two headline figures are accurate facts over facts tested and facts both accurate and citation-supported over facts tested. The second is always the smaller number and it is the one that predicts whether the description survives your next price change.

Step four: the four failure codes

• Stale: true at some point, not now. Look for an old page of yours.
• Foreign source: true of a listing or profile rather than of your own site. Look for a directory or app store entry.
• Category drift: true of your category, not of you, with no citation attached. This points at model memory rather than retrieval.
• Vague: too imprecise to score, for example "affordable pricing" when you asked for a price. Usually your own page's fault.

What the numbers were

We publish no accuracy count for Claude. We have not run the fifteen-prompt audit as a scored series, so there is no "X of Y" figure here, and inventing one would break the only rule on this site that has no exceptions.

What we hold from dated runs, at fact level, all on clawlaw.in:

• Prices absent from what a crawler received and taken from an app store listing instead. Claude, 6 August 2026. Failure code: foreign source, caused by our own page.
• Two public price lists disagreeing, and the contradiction stated in the answer. ChatGPT, 6 August 2026.
• A headline claim that could not be true. When an assistant checked it against public numbers it advised a buyer against the product, after which the company's accurate claims stopped counting for that answer. Claude, 6 August 2026. This is the most expensive fact-level failure in the whole programme.
• Correct competitor prices discounted for having no source. Two comparison pages stated competitors' prices with no link, no date and no source. The figures were correct when the assistant checked them independently, and it treated the pages as advocacy and used official sources instead. Claude, 17 September 2026.
• Our arguments used and the credit dropped, in two places, described by the assistant as a citation lapse. Claude, 17 September 2026.

And from our own domain, for scale: across 24 batches and 72 conversations on aiknowsus.com in September 2026, the phrase recording that we were not cited appears 161 times in Perplexity's own self audits. In one batch of six questions the assistant reported that it had not cited or recommended aiknowsus.com in any of the six answers, so there was no position for it to hold.

What this cannot tell you

• Accuracy is not consideration. An assistant can describe you correctly when asked and never name you when a buyer asks for options. The five blind prompts are the commercial test.
• One wrong answer is not the model's position. Answers vary. Two or three repeats before you act on it.
• A correct answer with no citation is fragile. It tells you the model remembers something right, not that your page is reachable.
• It cannot locate the error for you. The failure code narrows the search. Finding the page still means looking.
• It is not comparable with somebody else's audit unless their fact sheet and rubric are published too.
• It cannot be turned into a guarantee. We do not promise that a fixed page changes an answer, and we have no dated durability measurement to offer in place of that promise.

Common questions

Is it a bad sign if Claude has never heard of us?

Not on its own. Plenty of mid-sized companies are absent from model memory, and that is a retrieval problem, which is the more tractable of the two. The one dated publish-to-citation event we hold is ChatGPT citing a named clawlaw.in URL on 6 August 2026, fourteen days after that page went up. One page, one assistant, and not a schedule.

Why does the answer change every time we ask?

Because generation varies and retrieval varies. That is why the method asks for three runs per prompt and a report of the range. The facts that come back the same way every time are the settled ones.

Should we ask with our name or without it?

Both, in separate groups that never get averaged together. Named prompts test fact accuracy. Blind prompts test whether buyers will meet you. Reporting a named result as visibility is the single most common way in-house tests mislead their own management.

What if the answer is vague rather than wrong?

Score it as vague, then check your own page. If you asked for price and got "flexible pricing", the likeliest cause is that your own page says something close to that.

How long does this audit take?

Building the fact sheet is a couple of hours and is the useful half. Fifteen prompts, three runs, and scoring is roughly a day spread over a fortnight. It is the cheapest serious thing a business can do about this.

Sources and change log

Anthropic's own documentation for the web search tool, the model overview including knowledge cutoffs, and the pricing page on the day you need a figure. Read them directly and note the date.

Our observations come from the clawlaw.in programme from July 2026, including the 6 August 2026 interim checks and the 17 September 2026 Claude response audit, and from the aiknowsus.com audit of September 2026. The 6 August observations concern pages that were public and can be checked from outside; counts across our capture files cannot be, until the captures are published.

Change log. First published 29 September 2026 with the rubric and the log format and no accuracy count. When the audit is scored, the two figures will be added here with their denominators and dates, and this line will remain.

Disclosure. AI Knows Us sells AI visibility measurement, and this page recommends measurement.

What to do first

Write fifteen facts and the URL on your own site for each one. Then fetch two of those pages with scripts disabled and see whether the fact is actually in the text a machine receives. That check has already caught a missing price list in our own programme, and it takes twenty minutes.