How do you measure a brand’s visibility in AI answers?

Mentions, recommendations and citations: separate the observations

Your name can appear in a list without the assistant recommending your offer. Conversely, an answer can use a page on your site as a source without highlighting your brand. I recommend recording these separately:

• The mention: does your brand name appear in the answer?
• The recommendation: is your offer suggested to meet the expressed need?
• The citation: does a link actually lead to your site? Open it to check the page.
• Accuracy: do the business, products and conditions described match what you offer today?

An answer citing an old price calls for a different action from one that does not mention you. If you are new to this topic, start with the differences between SEO and GEO.

Prepare ten questions close to your customers’ needs

Here is an example for a fictional ceramics shop. These are test questions, not results obtained for a business. Adapt the product, budget and location to your activity.

1. How do I choose a ceramic cup for everyday coffee?
2. What should I check before buying handmade tableware online?
3. What is the difference between a stoneware cup and a porcelain cup?
4. Where can I buy handmade cups made in France?
5. Which shops offer handmade mugs for under €40?
6. Where can I find a dishwasher-safe handmade cup?
7. What ceramic gift should I choose for someone who loves coffee?
8. Which ceramics studios sell their work in Lyon?
9. What does [shop name] offer?
10. What alternatives to [shop name] are there for handmade tableware?

Keep questions that already name your brand in a separate group. They mainly check how it is described. Questions without your name are better suited to observing whether a new customer might discover you.

Keep conditions consistent for comparison

Choose services your audience is likely to use, then maintain a tracking table. For each attempt, record the exact question, date, service, model when shown, language, country and whether web search is enabled.

Use a new conversation for each question and record personalization settings. This reduces some context differences without making answers perfectly reproducible. Save the answer and cited links so you can review what you observed.

I would start with a small series repeated regularly, for example monthly. Keep multiple attempts for the same question if answers vary. A loading failure should be recorded as a test error, not as your brand being absent.

Count appearances without inventing a universal score

For each service and question group, you can calculate two simple indicators:

• Mention rate: the number of answers naming your brand divided by the number of valid answers observed.
• Citation rate: the number of answers containing at least one link to your site divided by that same number of answers.

Purely illustrative example: 2 answers linking to your site out of 10 valid answers give a citation rate of 20% in that sample. This does not mean your site appears in 20% of all that assistant’s answers. With ten observations, one extra result already changes the rate by ten percentage points.

Count an answer only once for each indicator, even if it contains three links. Keep the number of attempts next to the percentage and record incorrect descriptions separately. If you observe competitors, apply exactly the same rules; the comparison remains limited to your questions.

Supplement your tests with available data

Google announced dedicated Search Console reports for generative features in Search and Discover, with a worldwide rollout stated for August 31, 2026. They show, among other things, impressions, pages and countries. They describe Google, not every assistant. See the official announcement and report scope.

Also observe visits and actions on your site when your analytics can identify them. A mention may produce no click. A click may produce no sale. Keep those stages separate rather than attributing every revenue change to AI visibility.

To prepare access to your site’s data, follow the guide to connect Search Console and GA4 without exposing your credentials.

Use observations to choose an improvement

If an answer describes your offer incorrectly, first check the information on your own pages: product name, uses, price, conditions and update date. If cited sources are external, record exactly which ones before trying to get outdated information corrected.

If your site is absent, look at what the cited pages give readers: explanations, comparisons, examples or answers your page leaves open. These are ideas to work on, not proof that a change will guarantee a citation.

Choose an action useful to your visitors, record its date and observe what follows. These tests help you understand your presence; they do not replace an audit of your site’s technical condition.