How to measure AI visibility

Why single checks fail

Ask ChatGPT the same question twice and you may get two different lists of brands. A single check is a snapshot. To measure, you need many answers and a rule for what counts as a change.

The headline metric is AI visibility: the share of answers to relevant, unbranded questions that name your brand, per engine, market and week.

Step 1: Build a prompt set

Write down the questions buyers ask before they choose a provider like you, in their language. Good sources: sales calls, support tickets, Search Console queries.

• Mix intents: exploring (“how do I …”), comparing (“X or Y”), validating (“is X worth it”), problems.
• Start with 30 or more prompts. 30 prompts in 4 engines give up to 120 answers a day.
• Keep branded prompts (“Is [brand] any good?”) out of visibility: they always name you. Use them for sentiment and facts.
• Keep the set stable. Every change breaks the comparison with earlier weeks.

Step 2: Choose engines and markets

Buyers use different assistants. Track at least ChatGPT, Gemini, Google AI Overviews and AI Mode; add Perplexity and Claude if your buyers use them.

Answers differ by country and language. Treat each country and language pair as its own market, with prompts in that language, and report markets separately.

Step 3: Collect answers every day, as users see them

Read answers from the consumer interfaces where possible: what a person sees, with its sources. A bare model API without web search answers differently and has no sources. Where an engine is asked through its API, use web search so the answer has sources.

Store the full answer, the brands in the order they are named, and the cited and retrieved sources. Ask from the market's location.

Step 4: Report weekly numbers with n and a 95% range

One day of answers is noisy. Aggregate per week, per market and engine, and always show the sample size (n) and a 95% Wilson interval.

• Visibility: share of answers that name the brand. The headline.
• Named first: share of answers that name the brand first.
• Share of voice: the brand's mentions as a share of all tracked brands' mentions.
• Sentiment: how positively answers talk about the brand.
• Used as source: share of answers that cite or read one of your own pages.

Example: real data

Read it like this: Revolut was named in 96 of 324 answers (30%, range 25–35%). Monzo: 78% (74–83%). The two ranges don't overlap, so the difference is real for this week.

Step 5: Tell real changes from noise

Rule of thumb that holds up: a trend needs at least two full weeks, and a change is real only when the 95% intervals of the two weeks don't overlap. Otherwise the honest answer is “too early to tell”.

Never report day-to-day moves or a single position change as a signal.

Step 6: Look at the sources

Visibility says where you stand; sources say why. List the pages AI cites or reads for your prompts and mark those that name competitors but not you.

Step 7: Act, then measure again

Then compare whole weeks again, with the same prompts.

• Earned: get onto the comparison, review and list pages AI reads for your prompts.
• Owned: answer the buyer questions directly on your site, with facts that can be quoted.
• Correct: fix wrong facts about you at their source.
• Technical: let AI crawlers read your pages (robots.txt).

Common mistakes

• Reporting daily numbers or single positions as trends.
• Mixing branded prompts into visibility.
• Showing a percentage without n.
• Changing the prompt set and comparing with the weeks before.
• Comparing tools that use different prompts, engines or locations.

Checklist

1. 30 or more unbranded buyer questions per market, plus a few branded ones for sentiment.
2. ChatGPT, Gemini, Google AI Overviews and AI Mode; Perplexity and Claude if relevant.
3. Answers collected daily from the market's location, with sources.
4. Weekly visibility per engine and market, with n and a 95% range.
5. Trends only after two full weeks and without overlapping ranges.
6. A list of source pages that name competitors but not you.
7. One owner per next step; re-measure after each change.