Written 12 October 2026.
“We appear in 40% of AI answers” means nothing on its own. Without five labels, nobody can check it, repeat it or compare it with anything.
AI answers change with the date, the wording, the engine and even from one run to the next. So a number about AI visibility is only as good as the record of how it was made. Here’s what I look for, and what I put on my own numbers.
1. The date
Engines change their models and their sources without notice. A number from March describes March. If there’s no date, you can’t tell whether it still holds.
2. The exact prompt
“Best accounting software” and “best accounting software for a plumber” are different questions with different answers. In my personas test, adding one sentence about who was asking changed which brands were named. If the prompt isn’t published word for word, the number can’t be rerun.
3. The engine and the model
“ChatGPT” isn’t one thing. Free and paid versions use different models, and a company can change the model behind a product at any time. I record the exact version where the product shows it, such as Gemini Flash-Lite in the personas test, and say so when it doesn’t.
4. The number of runs
The same prompt can give different answers twice in a row. In the personas test, Sage Accounting was named in one generic ChatGPT run and not the next. One run shows what’s possible. It doesn’t show what’s typical.
5. One change at a time
If you rewrite a page, add schema and build links in the same week, and your AI visibility goes up, you can’t say which one did it. Change one thing between measurements, or say plainly that you changed several and can’t separate them.
Also worth recording
- Location. Answers differ by country.
- Account state. Signed in or out, with or without memory and chat history.
- The tool. If the number comes from a third-party tracker, name it. Different trackers use different prompt sets and methods, so their numbers aren’t interchangeable.
What I will not claim
- That a number with all five labels is right. Labels make a number checkable, not true.
- That these are the only things that change an answer. They’re the ones I can control and record.
The method note shows how I apply these to my own studies, and every study links its CSV so you can check the rows yourself.