Why you need a method
When someone asks ChatGPT which company to hire, they get a written answer instead of a list of links. Your usual analytics tools say very little about what happened in that conversation, and the answer can change from one attempt to the next. A single screenshot proves nothing.
To know whether your visibility is improving, you need what any serious measurement needs: clear definitions, a fixed sample and conditions you can repeat.
Mention, citation and recommendation
These are three different signals, and they are worth recording separately.
- Mention: the engine names your brand in the text of the answer. It shows the engine knows you, even if it does not suggest you.
- Citation: the engine links to one of your pages as a source. It tells you which content it uses, and it can bring you visits.
- Recommendation: the engine puts you forward as an option for what the user asked. This is the signal closest to a customer.
Why one number is not enough
You can have one signal without the others. A brand can be cited as the source of a fact while nobody recommends it as a supplier. It can also be recommended with a link to a directory instead of its own site. A single visibility percentage hides those differences.
Choose your prompts carefully
Your prompts are your sample. They should sound like what a real customer asks, so start from phone calls, the messages you receive, Search Console data and your real leads. Cover at least these types:
- Supplier: which company to hire for a service in an area.
- Problem: what to do in a specific situation, without naming the service.
- Comparison: which option to choose, using criteria such as price or experience.
- Brand: what the engine knows about your company and whether it is correct.
Repeat every prompt
Models generate each answer with some variation, and many queries rely on web searches that change too. A prompt run once is an anecdote. Run it several times and note how often you appear.
Lock the list before you start and always use the same number of repeats. If you change the prompts or the sample between rounds, you are no longer comparing like with like. The useful figure is frequency: how many answers mention, cite or recommend you, and how that moves over time.
Work in clean sessions
Memory, chat history and your account all affect what the engine says. If you test from your own profile, after talking about your company, you will see a flattering version your customers never see.
- Start a new conversation for each repeat.
- Turn off memory and personalisation, or use an account with no history.
- Keep the same language, location and search mode.
- Note the date and the model you used.
Measure each engine separately
ChatGPT, Gemini and Perplexity do not retrieve or present information in the same way. The sources they consult, how they cite and how much they lean on web search all differ. An average across engines hides that, so store results by engine and by prompt type. That shows you where to act: one engine may need more of your own content, and another may need you present in the sources it reads.
What to expect from the measurement
Nobody serious can guarantee you citations in AI answers. The engines decide each answer and change without warning. What we can do is measure carefully, find out which sources carry weight and work on those.
In each round we record the engine, the prompt, the repeat, the three signals and the sources cited. Then we compare that with data no model controls: Search Console, phone calls and real leads. If visibility goes up and contacts do not arrive, the work needs another look.