How to audit your brand mentions in ChatGPT, Claude and Gemini

How many answers do you need before a number means anything?

More than most audits use. With three samples of ten questions, you have thirty answers per engine. "Named in 6 of 30" and "named in 9 of 30" are not reliably different; the gap is inside sampling noise.

Two rules:

• Never report a rate without its N. "We are named in a fifth of answers" hides whether that is 1 of 5 or 200 of 1,000.
• Do not call a change until it is larger than the noise. Repeat the audit and compare, but treat small moves as no change.

This is also why an audit and tracking are different things. An audit tells you where you stand once. Daily tracking builds N over weeks, so a change can be told apart from luck.

What did our own audit find?

We track 61 questions about our own category in our workspace and read them every day. For a fair picture, here is what the workspace shows for 2026-09-04 to 2026-10-04 (24 of 31 days sampled), read on 2026-10-04; the strip at the top of this page holds the headline. This workspace reads ChatGPT and Perplexity; Claude, Google AI Overviews and Gemini are not switched on for it, so they are unmeasured here, not zero.

• 4,524 answers read: 2,349 on ChatGPT and 2,175 on Perplexity.
• We were named in 197 of them.
• All 197 came from the three questions that contain our own name: "what is Attensira" and "Attensira reviews", each named in 66 of 66 answers, and "Attensira vs Profound", named in 65 of 66.
• On the other 58 questions, the category, comparison and problem questions, we were named in none of 4,326 answers. On "best AI search visibility tools", the most-named brand, Semrush, appeared in 86 of 112 answers; we appeared in 0.

That is the most common audit result there is, and the reason brand questions mislead. "ChatGPT knows us" was true. "Buyers asking ChatGPT for a tool hear our name" was not, on any question we track.

The same data also showed who decides the answers. Across those answers, the engines cited 100,380 links; the most cited single domains were OpenAI's help centre and Google's developer documentation, followed by competitors' own sites. Our own domain was cited 1,525 times, yet named in none of the category answers: being read is not the same as being named. That tells us where the work is: the pages that decide category answers, most of which are not ours.

What do you do with the results?

Sort the questions you lose by how much they matter, then look at who decides each one:

• Lost to a page of yours that says the wrong thing. Fix the page. See fixing what AI gets wrong about your brand.
• Lost to a roundup or comparison page. Get onto that page, or publish the comparison yourself. See comparison pages AI search will cite.
• Lost because your page does not answer the question. Write the page, with an answer block at the top.
• Lost because the engine confuses you with something else. Fix your naming. See entity fog.

Then re-run the same questions after each change, and compare against the readings before it.

Which mistakes ruin an audit?

Most audits we have seen, including our own first attempts, fail in one of a few ways.

Asking only branded questions. "What is [brand]" is the question you are most likely to win and the one that tells you least. Our own numbers above are the example: perfect on our name, absent everywhere else.

Asking signed in, with history. An assistant that has talked to you about your company before may bring it up again. Use a fresh session, or the engine's API, so the answer is closer to what a stranger sees.

Counting screenshots instead of answers. A screenshot of a good answer is one sample. Report the count, every time.

Treating a missing engine as zero. If you did not ask Gemini, you do not know how Gemini answers. Write "not measured", not "0".

Mixing countries and languages without noting it. Answers differ by market. If half your samples came from a US connection and half from a European one, say so, or keep them apart.

Stopping at the score. The useful part of an audit is not "named in 3 of 40". It is the list of pages that decided the 37 answers you lost. Without that list, you know you have a problem but not where to work.

Changing the questions between rounds. If the second audit uses different wording, you cannot compare it with the first. Freeze the question list, and add new questions as new rows rather than editing old ones.

When should you stop auditing by hand?

When the sheet gets past a few hundred rows a week, or when you start wanting to compare this month with last month. At that point the audit has become tracking, and doing it by hand costs more than the decisions it informs.

How often should you re-run the audit?

Re-run it after anything that should change the answers: a new page, a rewritten comparison, a new listing, a rename. Otherwise once a month is enough for a manual audit, because the engines change their sources more slowly than a weekly cycle can show, and a monthly rhythm leaves time to do the work the last audit found.

Keep every round. The value of an audit grows with each repeat: three rounds with the same questions give you a trend, a baseline for every change you make, and a record of which fixes moved which questions. One round gives you a snapshot that is out of date by the time you act on it.

How does Attensira run this audit?

The free audit is the one-off version: it proposes eight buying questions from your site, asks four engines (ChatGPT, Perplexity, Claude and Google AI Overview) each once, and shows per engine how many named you and who they named instead. One email unlocks the map of which pages decided each lost answer, and one rewrite of one of your pages.

A workspace is the daily version of the same sheet: its columns filled every day for every question you track, on ChatGPT, Perplexity, Claude and Google AI, with Gemini from Growth. It records named, linked, position and the brands named instead for every answer, and reports each number with its sample size. An engine you do not track is shown as unmeasured, never as zero. Then the crew does the work on the questions you lose, with your approval. That is the service of an SEO agency at the price of a tool.