Checking the evidence
How to measure AI visibility without a made-up rank
Record what appeared, where it appeared and how you checked. Keep brand mentions, links to your site and business enquiries separate. A repeated, defined sample can describe that sample; one screenshot cannot establish your visibility across an entire market.
Why repeat the check?
In SparkToro's January 2026 study, repeated brand-recommendation prompts produced varying lists and orderings. The researchers found appearance frequency more informative than treating answer order as a stable rank. The study's prompt selection and provider context limit how broadly its findings apply. Read the study and its methodology.
For your own checks, save the exact question, engine, date, market context and response. Repeat the same version before comparing periods. Changing the question while drawing a continuous trend hides a method change.
What should the sample register contain?
| Record | Why it matters |
|---|---|
| Prompt and buyer intent | Shows what question was tested and whether it matches a real need |
| Engine, mode, locale and time | Distinguishes consumer search, API output and changing conditions |
| Attempted, valid and failed runs | Makes the denominator inspectable |
| Brand mention and owned-site citation | Separates a name appearing from a link to your evidence |
| Raw response and source URL | Lets someone verify the interpretation |
A worked example
Illustrative numbers: suppose 30 valid answers include six mentions of a business, and three of those answers link to its website. The mention rate is 6/30, or 20%. The owned-site citation rate is 3/30, or 10%.
Those figures describe the valid answers in that test. They do not mean 20% of customers saw the brand, or that 10% of enquiries came from AI. Report failed attempts separately and explain what counted as a valid response. Small, correlated prompt samples cannot support confident market-wide claims.
What can Google report directly?
Google announced dedicated generative-AI performance reports in June 2026 and noted worldwide rollout on 31 August. These report impressions and supporting breakdowns for Google's AI features. The data also appears in the overall performance report, so do not add it to the total a second time. Google's report announcement.
Keep that provider-reported view separate from third-party prompt experiments. If your reporting integration cannot retrieve the dedicated report, show that limitation instead of inventing a substitute.
Where do enquiries fit?
Track a separate enquiry record with landing page, referrer or campaign tags when available, customer-reported discovery and qualification outcome. Missing attribution stays unknown. An AI citation is evidence of appearance in an answer, not proof that it caused a sale.
What should you ask an agency to show?
Open one observation behind the chart. Check the saved question and answer, follow the citation, and compare the date and method with the previous period. Then ask which practical decision the result supports. If the agency cannot open the underlying evidence, the chart is not ready to guide spending.
Sources and review
- SparkToro AI consistency study. Repeated-prompt observations and limitations, 27 January 2026.
- Google generative-AI performance report announcement. Provider reporting, updated 31 August 2026.
References checked on 6 September 2026. Provider guidance can change. Our checklists and illustrative examples are recommendations, not observed client results. Send a correction.
Revision note
6 September 2026: initial researched version. No claim of historical client results.