The short answer
Pick the questions your customers would ask, ask the same questions to each assistant on a schedule, and count how often your site is named or cited. Treat the result as a sample: answers vary from one run to the next, so report counts and a range, not a single score, and do not read a cause into them.
What you can and cannot measure
There is no report from an AI assistant that tells you how often it mentioned you, the way Search Console reports impressions. Everything you can know comes from asking the assistant yourself and from the visits that reach your site.
That makes AI visibility a sampling problem. You observe some answers to some questions at some times. It can show you how often you are named. It cannot show why, and a single answer shows very little.
Choose the questions first
Start from the questions a customer would type or say before they know your name: the problem they have, a comparison between options, a question about a price or a method. A question that contains your brand name measures awareness, not discovery, so track those separately.
Keep the list short and fixed. If you change the questions every week you cannot compare one week with another. Ten to thirty questions that matter beat three hundred that do not.
Ask repeatedly and count
Assistants can give a different answer to the same question on the same day. One run is an anecdote. Ask each question several times, on the same schedule, without a logged-in account or earlier conversation that could tilt the answer, and record for every answer whether your site was named in the text, listed as a source, both or neither.
Report counts: named in 7 of 20 answers. A percentage looks precise, and with 20 answers it is not: it can move a lot with a handful more. Show a percentage only once you have enough answers behind it, and show the range around it.
- Same questions, same schedule, same assistants.
- No account or history that personalises the answer.
- Record named, cited, both or neither, per answer.
- Keep the raw answers so you can check a number later.
Add the sources you can see
Two other signals are cheaper and less noisy than asking. First, whether AI crawlers may read your pages: robots.txt says which of them are blocked, and a block is a decision that should be on purpose. Second, visits: in Google Analytics 4 you can look at sessions whose referrer is an assistant. That counts people who clicked through, not people who read your name in an answer and went no further.
On Google, check whether the results for your important queries show an AI Overview and whether it names you. If you already rank on the first page and the overview does not name you, that is a specific question to answer for that page.
Where the evidence stops
Read the numbers with their limits
A count that goes up after you changed a page does not show that the change did it. Answers vary, assistants change, and other sites change too. Note the dates of your changes next to the counts, and say what you can show and what you cannot.
Do not compare a score across assistants as if it were one scale. Each assistant draws on different sources and behaves differently. Track each separately.
What to do with the result
Use it to find gaps, not to rank yourself. Questions where competitors are named and you are not point at pages to write or improve. Questions where you are named but not cited point at sources to make easier to quote. Crawl blocks you did not intend are the one thing you can simply fix.
In Lastmod
Lastmod asks ChatGPT and Google AI Mode the questions you approve, a few times a week, and counts how often your website is named or used as a source. A percentage shows only once enough answers were measured, and an interval around it shows how far it can still move. Lastmod also reads your robots.txt per AI crawler, records AI Overviews that skip a page you already rank for, and, with Google Analytics 4 connected, shows sessions from assistants per page. It does not say why you are named and does not link a change to the count.