Pick a set of buyer questions, record the full AI answers and citations, and run the same set again after a documented change. One screenshot can illustrate what happened in one test. A repeatable sample lets your team compare platforms, find wrong answers, and decide what to fix.
Define each metric before sampling
| Metric | Count as success | Report alongside it |
|---|---|---|
| Brand mention | The answer names the correct brand | Exact wording and ambiguous names |
| Recommendation | The answer explicitly presents the brand as an option for the task | Constraints, caveats, and alternatives |
| Owned-domain citation | An answer cites or links to a page on your canonical domain | The full URL and whether the page supports the claim |
| Factual accuracy | The answer's checkable claims match verified product or service facts | Examples of outdated or incorrect claims |
For each rate, divide the number of valid answers meeting the definition by the number of valid answers in that sample. In an illustrative sample of 40 answers, eight brand mentions are a 20% mention rate. Report timeouts or missing responses separately. A valid answer without a brand mention belongs in the denominator.
Run a test you can repeat
- Collect actual questions. Ask sales and support for examples, then split them by language, industry, and purchase stage. Keep brand questions separate from unbranded category questions.
- Record the conditions. Keep the platform, mode, locale, query, approximate time, and account state in the test record. Save the answer and linked sources where you are permitted to do so.
- Track one change at a time. For example, add missing specs to a product page and record when the update went live. Keep the prompt set stable for the next pass and note other changes during that window.
- Read the answers as well as counting them. Confirm which page supports each claim. Compare rates within each platform and report raw counts when samples are small.
For platform-specific diagnostics, see how to check brand visibility in ChatGPT Search. Use the SEO and GEO comparison to keep the answer sample separate from search impressions, site visits, and inquiries.
Common questions
How often should we sample?
Run a baseline first. Choose a regular interval your team can keep and run the same set after meaningful content changes. Record the number of tests and the dates of both windows.
What should we do about an incorrect answer?
Record the exact claim and citations. Fix incorrect information on pages you control and check authoritative documents for discrepancies. Repeat the same question before reporting a change.
Does an increase in mentions mean we generated more inquiries?
Track mentions and inquiries as separate outcomes. Review tagged referrals, submitted forms, and sales qualification before attributing an inquiry to a source.