How to measure AI mentions and citations without inventing a visibility score

An AI answer can name your product without linking to it. It can link to a page on your site without spelling out the brand. It can also mention the brand while explaining why it is unsuitable. Counting all three as a single success hides the difference between being present, being cited and being recommended.

This guide proposes a manual observation method you can reproduce with a spreadsheet and the AI Visibility Calculator. The formulas describe your sample only. They are not an industry-standard ranking score, an estimate of all AI users or evidence that a website edit caused a change.

Define the question set before you collect answers

Start with the decisions your intended customer needs to make. For an illustrative meeting-notes product, useful questions might concern multilingual calls, exporting action items or comparing tools for a small team. Keep questions that include your brand in a separate group. Asking an engine directly about your product makes a mention much more likely than an unbranded selection question does.

Write down a fixed set of prompts and label their intent. Keep engines separate and record whether browsing was enabled. Where visible, preserve the model or product version, date, location, account context and whether the conversation was new. These details do not remove variability; they make it easier to understand what changed between runs.

Use definitions another reviewer can follow

For this method, a mention means the answer names the brand or an agreed, unambiguous variation. A citation means the answer includes a link to a URL on the site being measured. Count each answer at most once for each metric. Three links to your domain in one answer still produce one cited answer. A third-party review about your brand is a different category from a citation to your own site.

Decide how to handle ambiguous product names before scoring. If your brand shares its name with a common word, require enough context to identify it. Add an uncertainty column rather than forcing every ambiguous reference into yes or no. Have a second person review disputed cases if you can, and preserve the explanation for the final label.

Track unsuccessful requests separately. A network error is not an answer. A completed response that says it cannot help may reasonably remain in the sample under a predefined rule. The important point is to document exclusions and keep the same rule across periods; otherwise the denominator can quietly change to make a report look better.

Calculate two rates with one clear denominator

The two categories can overlap, so adding the percentages does not produce a total visibility rate. They do not need to be nested either: a link can appear without the brand name. Enter whole-number counts in the calculator, add the sample context and download the report. Keep the underlying responses with it; a percentage alone cannot explain why an answer was counted.

Always show counts beside percentages. Moving from one to two mentions in ten answers doubles the mention count, but it is still one additional observation. A small or selectively chosen prompt set can move sharply. Without a representative sampling design, do not attach a claim about the entire market to a precise-looking decimal.

Compare equivalent samples and examine the misses

Repeat the same prompts under comparable conditions before interpreting a trend. If you add prompts, report the original set and the expanded set separately for the transition. Keep engine-specific reports rather than allowing a larger sample from one service to dominate a pooled average. Note product or model changes alongside changes you made to the website.

Then read the answers. An incorrect price, an outdated limitation and a missing citation suggest different follow-up work. Check whether your own public pages state the relevant fact clearly, whether supporting documentation exists and whether it is accessible. Improving a page can be worthwhile even if a later AI answer does not cite it. Avoid attributing every movement to the last edit.

Keep business outcomes in a separate report

A mention rate is not click-through rate, and a citation rate is not a conversion rate. Where analytics can identify referred visits, review those visits and their outcomes separately. Google says traffic from its AI features is included in Search Console’s Web performance reporting; that does not turn a hand-collected answer sample into a dedicated AI traffic report. See Google’s AI features documentation for its stated reporting scope.