How to measure AI citations with a repeatable set of questions

Start with a question the client wants answered: when someone asks an assistant about this service, does the business appear, and what does the answer say about it? A useful measurement records the question, the conditions and the result so another person can understand the comparison.

Build a small, relevant question set

Use questions from the client’s sales conversations, support requests and research. Include the markets and languages that matter. Keep branded questions separate from questions that do not name the client.

Record the exact wording. If you change the question set halfway through a reporting period, note the change rather than presenting the new result as a continuous trend.

Choose the assistants and modes deliberately. For a web-citation study, enable web search and record the setting. An answer from an API configuration is not automatically representative of the consumer app, and a sampled question set is not a census of what every buyer sees.

Record what happened in each run

Keep the date, assistant, visible model or mode, question and response. Extract the cited URLs, but retain the answer so an unexpected result can be reviewed.

Separate at least three outcomes:

• A mention: the answer names the business, with or without a link.
• A citation: the answer links to a source associated with the business.
• A referral visit: someone follows a link to the site, where it may be recorded by analytics.

A crawler request is a fourth kind of evidence. It shows a fetch, not that the page was cited or that a person visited.

Define the numbers before comparing them

For citation rate, state the denominator. “The client was cited in 12 of 40 completed runs” is clearer than an unexplained visibility score. Track failed runs separately; do not silently count a timeout as an answer without a citation.

For competitor comparisons, use the same questions and conditions. Explain how multiple brands in one answer are counted. If you classify tone or sentiment, review the relevant passage: a neutral mention, a recommendation and a warning are different results.

Repeat the questions over time. Investigate changes that recur rather than rewriting content in response to one missing link. A citation to a competitor’s page is a research lead, not proof that copying its format will produce the same outcome.

Put the sample beside the other evidence

Preferium’s AI-visibility tracking covers a prompt panel across supported assistants, with the plan determining how many are included. Google AI Overviews and AI Mode have their own observations. Keep those separate from ordinary search rankings and referral analytics.

A report can then say what changed in the sampled answers, which pages were cited and whether visits or enquiries moved too. It should also say what the sample cannot explain. A dated, reproducible observation is more useful than a single score with an unclear meaning.