One way to measure visibility in AI responses is to choose a set of questions, repeat them with the same method and register the pages cited. The result is a comparable sample. If in October the brand was quoted in six answers and in November in nine, you can investigate what changed. These numbers represent the set observed and not the total of times the brand appeared for all users.
Start by separating name, quote, and visit. An answer can name a tag without including your address. You can also quote a tab of that tag without writing the name in the text. A visit only happens when someone opens the link. Using the same column for the three events makes it impossible to understand what has improved.
Set the questions and testing conditions
Separate questions of discovery, comparison and decision. For a billing software company, “how to send an invoice to the first customer?” is a discovery question. “What billing programs work in two stores?” is a comparison. “Does the program allow exporting the SAF-T file?” comes close to a decision and can already refer to a product.
Register the consulted service, language, location when relevant and type of session. Save the date and answer allows checking if the page was actually used as source. If you ask a question in a conversation that already mentioned the tag, identify the observation as a continuation of this conversation. To compare with a root query, open a new session.
You can test a recast as long as you save it as a variant. If you limit yourself to repeating a single sentence, you lose the variety of the actual search. If you change all questions every round, you can no longer compare the periods. A fixed part and an exploratory part allow you to cover both objectives.
Always report the denominator
Suppose you prepare 20 questions for a service. In five queries, the research experience has no response generated. Of the 15 answers available, seven mention the brand and four call the site. In this example, the brand was mentioned in 7 out of 15 answers and quoted with link in 4 out of 15. The round form should also keep the five consultations unanswered.
If the service responds to all the questions in the next round, the rate may drop even though there are more citations. Five citations in 20 answers represent 25%, below four in 15, about 27%. Showing the absolute numbers next to the percentage avoids assigning the brand a change that may come from the availability of the format itself.
The denominator also matters when comparing services. A search-oriented tool can bring sources in almost all answers. Another can present a text without links. A useful comparison starts by telling if you are counting answers, questions or citations.
Cross-check the results with what happens on the site
According to Google documentation on AI features, the data from these experiences are included in the Search Console performance report for the type of web search. An increase in this report may justify an investigation of the pages and queries. The aggregate series itself does not identify the visits that resulted from a quote in AI Overviews or AI Mode.
In analytics reports, group the references that arrive with an identifiable domain and observe the input pages. Some privacy applications or settings limit the available reference data. Keep a category for unknown origin and avoid assigning it to a tool just because there was an increase in the same week.
If the intention is to measure budget requests, set what counts as a valid request. A click on a contact button is less strong than a completed request. Register the two measures in your own columns to see if the page attracts visitors and allows them to proceed.
Turn an observation into an editorial task
A lost quote deserves to open the answer and read what has come to appear. If the replacement page includes recent business conditions, confirm the date and clarity of the conditions on the home page. If the change only happened in one of the rephrases, repeat the sample before editing an article that is already useful.
The report can close with a decision per group of questions. For unanswered questions on the site, open an editorial task. For full but hard-to-find pages, make a technical review. On the pages that receive visits without resolving the doubt, read the next route with the team that knows the customer. The time spent measuring serves to choose where to move first.
For an example of reviewing an existing site, see the Antarte SEO case study. It covers changes to the site and organic search results, a separate metric from citations in AI answers.