Most people check this once, by asking ChatGPT about their company, and get a flattering paragraph back. That measures nothing. A buyer does not ask about you; they ask for a recommendation in a category, and the only answer worth knowing is whether you are in it.
What follows is the manual version. It takes about an hour the first time and about twenty minutes a week afterwards, and the discipline is the same as measuring anything else in search: fix the questions before you know the answers, and never change them because you did not like one.
The steps
- Write ten to twenty buyer questionsPhrase them the way a stranger would, without your name in them: 'best invoicing tool for freelancers', 'how do I track whether my SEO fixes worked'. Building that set properly is most of the work, and it is the part people skip.
- Ask each one in a fresh conversationA conversation carries context. If you asked about your own company three messages ago, the answer is contaminated. New chat, one question, record the answer.
- Record every name in the answer, not just yoursWho was named, in what order, and whether a source was cited. Your absence is only half the finding; who was named instead is the half you can act on.
- Repeat the whole set, same wording, weeklyOne run is an anecdote. The same twenty questions asked every week for six weeks is a rate, and the rate is the thing that can move. Answers vary between runs even with identical wording, which is why the rate matters more than any single answer.
- Read it as a rate and compare it with search'Named in 6 of 20 answers this week' is a measurement. Put it beside your share of voice in the same questions and beside your Search Console impressions, and the gaps between the three are your list.
What not to conclude
Being unnamed does not mean ChatGPT has never seen your site. It means that for that question, in that run, it chose others, and the reasons are mostly about whether your site says plainly what it is for and whether anyone else says so too. Answer engine optimisation is the honest summary of what is known about changing that, which is less than the people selling it suggest.
A failed request — a rate limit, a timeout, a refusal — is not an answer that named nobody. Record it separately and leave it out of the denominator, or a bad week for the API becomes a bad week for your brand on paper.
Questions people ask
Does the answer change if I am logged in?
It can. A logged-in account with memory enabled may carry what it knows about you into a new conversation. Use a browser session with memory off, or a clean account, so the answer reflects what a stranger would get rather than what a returning customer would.
How many questions is enough?
Enough that one strange answer cannot move the rate by much. Twenty is a reasonable floor for a single product; below ten the week-to-week noise is larger than any real change you are likely to make.
Is there a way to do this without the spreadsheet?
Yes, and that is what the AI visibility part of Rankelle is: the same fixed question set, asked weekly across ChatGPT, Perplexity, Gemini and AI Overviews, with failed requests kept out of the rate and the rivals listed beside you. The method above is the whole product; the tool only removes the Tuesday.
Try it on your own site, free
The whole audit of every website you add, no card and no expiry. Pay only when you want the proof.