Validate an AI visibility report

Can you reconstruct the measurement?

Find the question inventory, engine and interface, date range, search settings, language and market context. Ask whether the displayed number describes all collected answers, a filtered subset or a composite instrument. A label such as visibility can refer to several different measurements. The report should define the event, eligible denominator and treatment of failed or missing responses before you interpret a trend.

For example, an illustrative report says that a fictional supplier has a 60% recommendation rate. You need to know whether this means six of ten answers, sixty of a hundred, or a weighted index component. Those are different summaries. Ask how the scorer distinguishes a recommendation from a neutral mention or a warning, and preserve the answers that support the count.

Are custom tests being confused with the Index?

Rankfor Index is a fixed product measurement with several components. A custom prompt in Answer Trail creates a separate record, and Dice Roller Test repeats a selected question. A report can contain evidence from all three if it names each clearly.

It should not imply that a custom repeat test replays the whole Index battery or that a mention percentage is the Index score.

Inspect Rankfor Index > Overview and Source & Language for the represented scan, then open the relevant saved custom records separately. Account-level Reports provides a reviewed export path through PDF, Markdown or CSV. Verify which charts, custom notes and source details the generated file actually contains. Rankfor’s public playbooks provide context for these different jobs.

Does the evidence support the headline?

Select several material claims, including an unfavorable finding and an exception to the main trend. Read the full responses and supporting pages. A citation beside a statement does not guarantee the page supports it; use the distinction described in citation-verifiability research. Mark unavailable pages, contradictory evidence and reviewer disagreement as unresolved review findings.

Check whether the comparison changed prompts, engines, search mode or language. An expanding sample can discover additional brands or domains without any brand intervention. The recommendation sampling study describes this dependence on the observation horizon. Ask for a like-for-like subset where possible, and label any break in the series when the method changes.

Which conclusions should pass review?

A finding can state what was observed, under which conditions, and what decision it informs. Claims about a campaign causing lift require more than a before-and-after screenshot. Look for competing explanations, control strategy where appropriate and a clear separation between delivery, answers and commercial outcomes. A crawler count should never silently become a lead count.

Finish the review with accepted findings, findings requiring correction and questions that remain unresolved. Assign each correction to a concrete artifact: the event definition, a missing denominator, an unsupported source link or an overconfident headline. This makes the review useful to the report’s author and gives decision-makers a transparent basis for accepting a bounded conclusion.