How to measure AI visibility: questions, engines, ranges and change

AI visibility is how often AI answer engines name your brand when a buyer asks a question about your category. You measure it by asking a fixed set of buyer questions to each engine, many times, and reporting the share of answers that name you with a range that says how sure that share is.

In short

• Measure a rate, not a rank: AI answers vary every time, so “position” has no fixed meaning.
• Count only questions that don’t mention you. Questions with your name in them measure recall, not discovery.
• Ask every engine your buyers use, the same way, with web search on, and keep the searches and sources behind each answer.
• Report every rate with its 95% range and its sample size, and call a change only when the range of the difference excludes zero.
• Measure competitors on the same answers, so a gap is a like-for-like difference.

This guide covers:

• What is AI visibility, exactly?
• Which questions should you measure?
• Which engines, and how should they be asked?
• How many answers make a reliable number?
• How do you compare yourself with competitors?
• When has your visibility really changed?
• What does the number not tell you?

What is AI visibility, exactly?

AI visibility is the share of answers, to questions buyers ask without naming you, in which an engine names your brand. It is a rate between 0% and 100%, measured on a defined set of questions, engines and dates.

That definition rules out three common shortcuts. A single answer is not a rate. A “visibility score” that blends mentions, sentiment and position into one number can’t be checked. And a ranking position assumes a stable list, which AI answers don’t have.

Which questions should you measure?

Measure the questions a buyer asks before they know your name: the category, the comparison and the problem. “Best project management software for agencies” measures discovery; “is Taskmoor any good” measures recall, and you’d expect to be named in nearly every answer.

A useful question set has three properties:

1. It is fixed. The same questions every run, so runs are comparable. Add questions in a new set rather than changing the old one.
2. It covers the buying journey. Category questions, comparisons, pricing, security and integration questions, and the problems your product solves.
3. It separates branded from unbranded. Questions that mention you are measured apart and kept out of the headline figure.

Drafting the first set from your own site is a reasonable start, as long as a person reviews it. AmpliRank drafts questions from your website and you approve each one before it is asked.

Which engines, and how should they be asked?

Ask every engine your buyers use, through each engine’s own API, with web search switched on, the same way for every question. Today that means ChatGPT, Claude, Gemini, Perplexity and Google AI Mode.

Web search matters because buyers’ answers come from it. With search on, engines run their own searches before answering (a step Google calls query fan-out) and cite the pages they read. Keeping those searches and citations with every answer is what lets you explain a result instead of only reporting it.

One limit is worth stating plainly: API answers are not identical to what a signed-in person sees in an engine’s app, which adds history, location and account context. That is a reason to report a rate with a range, not a reason to stop measuring.

How many answers make a reliable number?

Enough that the range is narrower than the change you want to detect. As a working rule, about 200 answers give a 95% range of roughly 6 points either side of a 30% rate; 50 answers give about 12 points either side.

Answers	Observed rate	95% Wilson range
20	30%	15% to 52%
80	30%	21% to 41%
200	30%	24% to 37%
500	30%	26% to 34%

The Wilson interval is the standard choice for proportions like this because it stays sensible at small samples and near 0% or 100%, where the simpler textbook formula gives impossible ranges. To halve the width, collect about four times the answers. Why one AI answer isn’t a measurement works through the numbers.

How do you compare yourself with competitors?

Measure competitors on exactly the same answers you are measured on. If 192 answers name you 60 times and a competitor 84 times, the 13-point gap compares like with like: same questions, same engines, same day.

Measuring each competitor on a separate sample adds its own noise to every comparison. Pair the comparison over the same questions and engines, and call a gap only when its range excludes zero. Otherwise report it as “no clear gap”, with the range.

When has your visibility really changed?

Your visibility has changed when the range of the difference between two runs excludes zero. Two headline numbers moving apart is not enough: 25% one week and 31% the next, each from 192 answers, is still within the noise.

Two practices keep change honest:

• Same questions, same engines. A run that asked different questions, or skipped an engine, isn’t comparable with the last one.
• Observed, not caused. When a rate moves after you publish a page, report the change as observed. Many things change in an engine’s answers each week, and one run can’t attribute the move to your page.

What does the number not tell you?

The rate tells you how often you are named; it doesn’t tell you why, or what to do. For that you need the evidence behind each answer:

• The searches the engine ran, grouped into the sub-questions it actually researched.
• The sources it cited, especially in answers that leave you out.
• Your own pages: whether any of them answers each question the engines searched.

Those three turn a number into a list of work: the page to write, the source to earn a mention on, the question to add to the next run. That list, with the evidence attached, is what a weekly brief should carry.

Where to start

Write down fifteen questions a buyer asks before they know your name. Ask each one in the engines your buyers use, several times, and record whether you are named, who is named instead, and which sites are cited. Our method shows how each range is computed, and you can start measuring with the same questions in AmpliRank.

Sources

• Edwin B. Wilson, Probable Inference, the Law of Succession, and Statistical Inference, Journal of the American Statistical Association 22 (158), 1927, accessed 12 November 2026.
• Google Search Central, AI features and your website, accessed 12 November 2026.