Build a fixed list of the questions your buyers actually ask, run them across the tools that matter to you on a regular schedule, and record which businesses get named and which sources the answers lean on. That gives you a sample you can track for direction over time. It does not give you a share-of-voice percentage, and any tool reporting one is estimating from a sample of its own without always saying so.
Why this has to be sampled
Answers vary between users, between sessions, and between one minute and the next. There is no ranking to check and no index you can query. Two people asking the same question can get different sources, and neither result is wrong.
That rules out measurement in the sense anyone means it. What remains is sampling, done consistently enough that the direction is informative even when any single reading is not.
The method
- Write twenty to thirty questions your buyers genuinely ask, in their words. Mix category questions, comparison questions and problem questions.
- Fix the list. Changing the questions between rounds destroys comparability, which is the only thing this exercise offers.
- Pick your tools and stick to them. Note which, because they behave differently.
- Run each question in a fresh session with no prior conversation, so memory and personalization do not contaminate the result.
- Record four things: whether you were named, who else was, which sources were cited, and whether anything said about you was wrong.
- Repeat monthly or quarterly, on the same day, in the same way.
What to record
| Date and tool | Both change the result, so both have to be on the row |
|---|---|
| Named or not | The simplest signal, and the one worth trending |
| Who else appeared | Tells you who the consensus currently favors in your category |
| Sources cited | The most actionable column. These are the places worth being described well |
| Accuracy | Whether what was said about you was correct. Errors here are worth fixing first |
The sources column is the one that earns the exercise. It tells you which directories, publications and third-party pages are shaping how your category gets described, and those are places you can act on.
Reading the results honestly
- Trend the direction over several rounds. One round tells you almost nothing.
- Watch the source list more than the mention count. It changes more slowly and it is more actionable.
- Treat factual errors as urgent. A wrong service area or an outdated claim repeated by a model is a live problem.
- Label it as a sample every time it is reported. Including in your own internal decks.
On the tools that offer to do this
Several products now report AI visibility as a percentage. Some are useful, and the useful ones tell you their methodology: how many prompts, how often, from where, across which models.
Ask for that before trusting a number. A percentage produced from an undisclosed sample is a sample with a decimal point on it.
A spreadsheet, thirty fixed questions and an hour a month gives you more usable insight than most paid tools, because you chose the questions and you know exactly how the numbers were made.