How to measure AI visibility
Traditional analytics tell you how people find your site. They say nothing about the conversation happening one layer up — the moment a buyer asks ChatGPT, Claude, Gemini, or Perplexity a question in your category and the model answers by naming a few brands. If you are not one of them, you are invisible to that buyer, and no dashboard built for the blue-links era will tell you.
Measuring AI visibility means measuring that answer directly. Here is the framework we use.
Start with the questions, not the keywords
Keyword volume is the wrong unit. An AI assistant does not return a ranked list of pages for a keyword; it returns a synthesized answer to a *question*. So the first artifact is a prompt set: 25–100 real buyer questions, phrased the way a person actually talks, that should plausibly surface your brand.
A few rules that matter:
- Do not name your own brand in the prompts. "Is [your brand] any good?" is a vanity check — of course the model talks about you. The signal you want is whether you show up *unprompted* when someone asks a neutral question like "what's the best tool for X?"
- Cover the buying journey. Mix broad discovery ("best options for…"), comparison ("X vs Y"), and bottom-funnel ("cheapest…", "for small teams…") questions.
- Hold the set constant. The whole point is week-over-week comparison. A prompt set that drifts can't be trended.
The four metrics that matter
Run that prompt set across every assistant and parse each response. Four numbers fall out:
| Metric | What it answers | Why it matters |
|---|---|---|
| Visibility rate | Of all answers, what share mention you at all? | Your baseline presence. The headline number. |
| Share of voice | When you appear, how do you stack up against named competitors? | Presence is not enough if a rival is always the top pick. |
| Citation count | How many answers link a page on your domain as a source? | Citations are the modern top organic result — they drive both inclusion and referral traffic. |
| Sentiment | When you are named, is the framing positive, neutral, or negative? | A frequent but unflattering mention can cost you the deal. |
A useful refinement is the mentioned-but-not-recommended gap: answers where your name appears ("[you] is one option, but I'd go with [competitor]") but you are not the pick. These are the highest-leverage cases to fix, because you are already in consideration — you just need the content that tips the model.
Measure lift, not snapshots
A single run is a photograph; you want the film. Because the prompt set is fixed, re-running it on a schedule gives an apples-to-apples trend. When you ship a change — a new page that directly answers a question, a structured-data fix, a corrected fact — you can watch the specific question's outcome move and attribute the lift to the change.
This is the discipline that separates AI visibility work from guesswork:
1. Establish a baseline run. 2. Pick the highest-leverage gaps (usually mentioned-but-not-recommended, or a competitor cited where you have no equivalent page). 3. Ship one specific change per gap. 4. Re-run and compare. Did the line move?
Where Lumear fits
Lumear automates this loop: it runs your prompt set across ChatGPT, Claude, Gemini, Perplexity, Google AI Overviews, and Bing, parses every response for the four metrics above, captures exactly which sources each model cited, and ties visibility changes back to the recommendations you shipped. The prompt set stays constant so the trend is real, not noise.
The takeaway: you cannot improve what you do not measure, and you cannot measure AI visibility with SEO tools. Pick your questions, track the four numbers, and re-run on a cadence. That is the whole game.