How to check whether ChatGPT actually cites your site

AI visibility has become a line item on a lot of marketing plans without anyone agreeing on how to measure it. Here is the method we use internally. It is manual, it takes about an hour, and it will tell you more than most dashboards.

Build a prompt set, not a keyword list

People do not type keywords into assistants. They describe a situation. So instead of "best crm for small business", write the twenty things a real buyer would actually ask, in their words, with their constraints. "We're a 12 person agency on Google Workspace, what CRM won't be overkill" is a prompt. "crm software" is not.

Twenty is enough to see a pattern and few enough that you will actually finish. Write them down somewhere you can rerun them next month.

Run each prompt three times, in a fresh session

This is the step everyone skips and it is the one that matters. Assistant answers are not deterministic. The same prompt can return three different sets of sources, and if you run it once you will draw a conclusion from noise. Use a logged-out or temporary session so your own history is not shaping the answer, and record what came back each time.

• Were you named in the answer text, whether or not you were linked?
• Were you linked in the citations?
• Which competitors appeared, and how often across the three runs?
• Which page of yours did it pick? It is often not the one you would have chosen.

Read the sources it did pick

The most useful output of this exercise is usually not your own score. It is the list of pages that got cited instead of yours. In our experience they are rarely the best-written pages on the topic. They are the ones that state things plainly, put the specifics in a scannable structure, and answer the question in the first hundred words rather than after four paragraphs of throat-clearing.

Comparison tables, dated pricing, explicit numbers and clear definitions get pulled repeatedly. Long narrative introductions almost never do.

Expect the numbers to be unstable

Month to month, our own tracked mention rate moves by five to ten points on prompts where nothing about the underlying page changed. Model updates, index refreshes and retrieval changes all move it. Treat a single month as a data point and not a verdict, and be suspicious of any tool that reports this to two decimal places.

When to bother automating it

If you have fewer than about thirty pages worth tracking, do it by hand quarterly and spend the money elsewhere. Automation earns its keep when you need the same prompt set run consistently over months and compared against your Search Console data, because the interesting question is not whether an assistant mentions you. It is whether the pages that get mentioned are also the ones that earn clicks.