Run the check yourself once a month: about ten buying questions worded the way your customers ask them, put to ChatGPT and Perplexity while logged out, with the answers logged in a spreadsheet.
Plan on half a day per check. Until that spreadsheet exists, you'll find out you're being cited the way we did: because someone tells you.
The benchmark
In August 2026, at a first meeting, a brand manager told us that ChatGPT had cited us: "
I mostly asked ChatGPT for the best Shopify agencies and it came up. I think you were first or among the first, so I contacted you."
Three metrics turn her "I think" into a measurement:
- presence rate: the answers that cite your brand as a share of all the answers you collected;
- link rate: the share of answers that point to your domain;
- share of voice: your citations compared with those of three competing brands you name in advance.
The rule
Start by hand, before you buy anything.
In the audits we run, we budget two to three hours for ten to fifteen questions put to three assistants, which works out to about four minutes per answer read and logged (a planning estimate, not timed).
At that pace, ten questions asked three times each on ChatGPT and Perplexity give 60 answers, so four hours.
For paid tools, we gave ourselves one rule: no subscription until the manual check shows real citations to track.
The entry price is modest, but in our view you learn the most by reading the answers yourself, where you can see which comparison the assistant read, which sentence from your page it picked up, and who it ranks ahead of you.
The steps
- The questions. Write about ten, split into four types: the problem ("how to choose a carry-on bag"), the category ("best French leather sneaker brand"), the comparison ("brand A or brand B") and the brand ("brand A reviews").
- The runs. Ask each question three times per assistant, in a new conversation each time. According to a SparkToro study published in January 2026 and run on ChatGPT, Claude and Google's AI, there is less than a one-in-a-hundred chance of getting the same list of brands twice. So record presence and set rank aside.
- The conditions. Open a private browsing window, with no account, in French and from France. With an account logged in, you're partly measuring your own history.
- The spreadsheet. Log one row per answer: the date, the question, the assistant, whether your brand appears, any link to your site, the competing brands named and the sources shown.
- The cadence. Keep the same day of the month and the same questions, so two checks stay comparable. We settled on this frequency for this site.
- The reading. Calculate the three metrics for each assistant, and add a note to your logbook for every release that could move them. This check measures presence, and what that presence earns shows up elsewhere, in GA4.
The trap
The trap is to look for the cause of an absence on your own site alone.
In May 2026, Muck Rack classified more than 25 million links cited by ChatGPT, Claude and Gemini: about 84% lead to content the brand in question neither produced nor paid for (press, encyclopedias, forums, other companies' sites). The study comes from a PR software vendor, an interested party, and it leaves Perplexity out: read that figure as an order of magnitude.
We saw this in summer 2026 on this site. The third-party rankings that assistants draw on in their answers didn't mention it anywhere, and an absence like that gets fixed where those rankings are published.
The sources column then becomes the most useful one in the spreadsheet, since it lists the pages where you can get a mention or a correction.
The sentence to present
In this fictional example, ChatGPT cites your brand in 6 answers out of 30 and the closest competing brand in 18, and the answers where you're missing rely on two comparisons and a forum thread.
Copy the sentence, then rewrite it with your numbers for your leadership team:
"Across our ten buying questions, each asked three times, ChatGPT cites us one time in five and cites the closest competing brand three times in five. When it doesn't cite us, it relies on two comparisons and a forum thread: I suggest we deal with those three pages before commissioning new articles."
