How to Track Brand Visibility in ChatGPT, Gemini and Perplexity
Build a fixed set of 20–30 prompts a real buyer would type, run them across the major engines on a schedule, and record whether you were named, in what position, and what was said. Keeping the prompts fixed is what makes the trend meaningful — change them between runs and you are measuring nothing.
You do not need a tool for this. Most commercial AI visibility products are prompt-runners with a dashboard, and running the prompts yourself for the first few months teaches you more about why you are or are not cited than a score would.
Building the prompt set
- Use buyer questions, not your brand name. “Best travel CRM for a UK agency” tells you something; “what is Autratech” tells you almost nothing.
- Cover the funnel — a few problem-aware questions, a few comparison questions, a few supplier-selection questions.
- Include your weak spots. The categories you want to be known for but are not yet.
- Keep them natural. Write them the way someone actually types, not as keyword strings.
- Fix them. Once set, do not edit. Add new prompts as a separate group with their own baseline.
What to record
| Field | Why |
|---|---|
| Named yes/no | The headline measure |
| Position in the list | First of five is different from fifth |
| What was said | Being named inaccurately is its own problem |
| Who else was named | Your real competitive set, as the engine sees it |
| Sources cited | Shows which third-party pages are doing the work |
Reading it without fooling yourself
Three disciplines matter. Run each prompt in a fresh session, because conversation history changes answers. Accept that results vary run to run — a single appearance is noise, a pattern across three months is signal. And resist editing a prompt because you do not like the answer, which is the most common way these programmes quietly become marketing material.
What to do with the results
The sources column is where the value is. If the same three directories or comparison posts keep being cited for your category, those are your targets — being present and accurate in them is more tractable than hoping to be retrieved directly. If your competitors are named and you are not, read what is said about them: it usually describes a specificity you have not published.
Frequently asked questions
How often should we run it?
Monthly is enough. Weekly produces noise that looks like movement, and the underlying work takes longer than a week to show.
Should we buy a visibility tool?
Eventually, perhaps, if you are running many prompts across many engines. Start manually — the first few months are where you learn why answers come out as they do, and a dashboard hides exactly that.
Why do we appear one week and not the next?
Generation is probabilistic and retrieval varies. That is expected and is precisely why the method is a fixed set run repeatedly rather than a spot check.
What is a good result?
There is no benchmark, because no engine publishes distributions. Your baseline is your own first run. Progress is being named in more of your own prompts over time.
We set this up as part of the AI search visibility audit, and the prompt set is yours to re-run whoever does the work.
