How to measure AI visibility without fooling yourself
One answer is an anecdote; here is a measurement you can defend in front of a CFO.
The short answer
Measure AI visibility by sending a fixed set of buyer prompts to each engine repeatedly, reporting each engine separately before blending, and comparing before and after with the same prompts on the same schedule. Single answers vary too much to trust, and branded prompts inflate the score.
Every number is an estimate of a rate: out of all the times buyers ask a relevant question, how often does an engine name you? Most measurement mistakes come from forgetting that.
Why is one AI answer not a measurement?
Engines are not deterministic. Ask the same question twice and you can get different brands, in a different order, citing different sources. A screenshot in Slack proves that an answer once existed. It does not tell you how often it happens. Run each prompt several times on a schedule and report the share of runs that mention you.
Which prompts should you track?
The fastest way to a flattering score is to track prompts with your brand name in them. Do not. Build the set from the questions a buyer asks before they have a shortlist:
- Problem questions: “how do I keep condensation out of a tent”
- Best-of lists: “best 2-person backpacking tent under $400”
- Comparisons: “Tarn 2P vs Halo 2”
- Brand questions, kept in their own group
Tag each prompt with its journey stage and persona. Ridgeline’s blended visibility is 38.4%, but it wins research questions and loses comparisons, and only the tags show that.
Should you measure each AI engine separately?
Engines retrieve from different sources and refresh at different speeds. Ridgeline sits at 51% on Perplexity and 29% on Gemini. Those gaps have different causes and need different fixes. Report each engine separately, then blend with weights you choose on purpose, for example by each engine’s share of your AI referral traffic.
How do you prove a change worked?
To show that a fix worked:
- Freeze the prompt set and the engines for the test.
- Record the ship date.
- Keep running on the same schedule for at least two to six weeks while engines refresh.
- Compare the rate before and after, and report the confidence. A two-point change on twenty prompts is usually noise; the same change on two hundred is not.
Lighthouse does this automatically when you mark a fix as shipped, and every number opens to the answers behind it.
Questions people ask
How many times should I run each prompt?
Enough that the rate is stable. For most categories, daily runs across a week give a usable weekly rate; launches and crises justify hourly runs.
Should I include branded prompts?
Yes, but in their own group. Mixed into the main score they inflate visibility and hide the questions you are losing.
