On 20 September 2026, 5 of the 8 pages on Google’s first page for this question were selling a tracker. A tracker earns its fee at 200 questions across 6 markets. For a company with one product line and 3 buyer questions, it’s a spreadsheet with a subscription, and the spreadsheet on its own takes 20 minutes a month. The routine below is the one behind our free check. Do it once yourself first.
Write the 3 questions before opening a browser
The rule on our check page is one line: “Not your brand name”. A buyer who types your name has already found you. The check is whether you come up when nobody asks for you.
Write each question the way a buyer would say it out loud at the moment of choosing. Category plus one constraint that matters in your market: company size, sector, region or certification. “Payroll software for a 300-person construction group.” “Who audits ISO 27001 for a software company in the Netherlands.” “Contract warehousing for a chilled-food brand near Manchester.”
Three is enough for a first month. Two of them should be questions the site ought to win. The third should be one it will probably lose, so the sheet shows what an answer looks like with you absent from it.
Run them logged out, and turn Search on
OpenAI’s help page on web search, updated in August 2026, gives 2 reasons an answer is personal to the person asking. The first is memory: “If memory is enabled, ChatGPT may use relevant saved memories when rewriting a search query”. The second is location, which the page says is taken from the IP address to shape local results. Logging out removes the first. Running the check from the office, where your buyers aren’t, doesn’t remove the second, so note the city the run was made from.
So we use a private browser window, no login. The same page says web search works for “People who are not signed in”. Turn Search on rather than hoping it fires. The page describes a Search tool under View all tools. Then, in the answer, “Select Sources, when available, to view cited sources and other relevant links”. That panel is where the citations are.
Perplexity needs less setup. Its help centre, last modified 3 September 2026, describes numbered citations on every answer, so there is always a source list to read. Same 3 questions, same wording, same private window.
We copy each answer into a text file rather than a screenshot, so it can be searched later. Two assistants and 3 questions give 6 answers in about 10 minutes.
Record 5 fields per answer
The date. The question, verbatim. The vendors named, in the order they appear. The pages cited, as URLs. Whether any of those pages is yours.
That’s a 6-row sheet with 5 columns. Named and cited are separate columns on purpose. We covered why in what an AI citation is: a name can come from the model’s memory, and a citation is a page somebody fetched. Going from unnamed to named and going from uncited to cited are 2 different events, and the sheet should show which one happened.
We keep the answers, not just the sheet. When a competitor appears in month 3 and wasn’t there in month one, go back and read what the assistant said about them.
What changes between months, and what doesn’t
Expect the answers to move even when your site hasn’t. OpenAI’s page tells its own users that results and citations can be incomplete or wrong. Two runs on 2 days naming different vendors for the same question is noise. A signal is a change that holds for 2 consecutive months: your page appears in the citations for question one in October and November, having been absent in September. Or a competitor’s page is cited every month and yours never is, which tells you which page to read next.
Run the check on the same day each month, from the same kind of window, with the same wording. Change the wording and you’ve started a new series.
What this routine doesn’t see
It doesn’t see how often buyers ask. Six answers a month is a sample of 6, and the assistant may answer 500 real buyers differently. Server logs are a different question again: they record the fetches rather than the answers, with a different method, and we keep the 2 apart.
Bing’s side of the picture is different too. Copilot’s citations are reported in a console rather than observed one answer at a time, that console exists as of February 2026, and it’s a separate setup.
The routine sees one thing well: whether, on the day of the run, the assistant would have put the company in front of a buyer. For a site with 3 buyer questions, that’s the number we watch.
Do it yourself, then send us the 3 questions
Run the 20 minutes before anything else. If you want a second run from us, the free check is the same routine with Gemini added, plus what we’d change on your pages. The AI citations article repeats the ChatGPT and Perplexity run 30 days after delivery, so the before-and-after goes on record without anyone setting a reminder.