When can a founder trust an AI visibility dashboard?
Trust a dashboard when you can open the answer behind every number: the prompt, the date, the engine, and the source it attached. Compare the same questions, the same engines, and the same window, and the score is a decision you can defend.
Start with the decision the view is supposed to support, then check that the prompts, the collection route, and the labels match that decision. A branded prompt tells you how an engine describes a company it was handed. An unbranded prompt tells you whether a buyer who never typed the name would hear it. For each result, keep the retained answer, the exact prompt, the date, the engine, the collection route, and the source URLs. A result with no citation stays in the record. Retrieved pages, attached citations, verified support, mentions, and recommendations each get their own field, so a single score can be opened back into those parts. The reliability view I ask an operator to show has separate counts: attempts, successful answers, failed collections, answers with no citation, answers with no brand mention, and the observations eligible for this comparison. Those groups can overlap. A failed collection stays labeled as a missing observation. The denominator is the set you meant to compare. Hold the comparison to what stayed fixed: question wording, whether the brand was supplied, engine, API or consumer product, collection provider, scoring definition, and window. A new model or a new prompt list starts a new series. The operator should be able to rebuild the displayed number from the note and the retained records. The Index below is the public aggregate. The vendor's retained records are how you audit the vendor's own score. Read the two populations before you read the two scores. Accept the dashboard for the job its evidence supports: a description check, a recommendation check, or a citation check. The AuthorityTech guide linked below is the tracker audit. Once the records rebuild the number, use it for the job those records describe.
The Index number
In this release the AI Visibility and GEO worth segment had 307 observed answer runs and 708 cited source domains across 13 distinct run dates.
Published November 5, 2026, updated November 14, 2026. Counts from Machine Relations Index release November 5, 2026, runs collected May 10 to November 5, 2026. Index table. Dataset: doi:10.5281/zenodo.23652423.
- Definition
- Machine Relations
- In practice
- AuthorityTech's checklist for when an AI visibility tracker is worth paying for
Related answers
FAQ
Does a failed collection count as an answer that ignored the brand?
A failed collection is missing evidence. A successful answer with no mention, or with no source attached, is an observed result. Keep those states in separate columns.
What do I do when the vendor chart and the Index disagree?
Compare the populations, dates, questions, routes, and definitions. The public Index is the aggregate. The vendor's raw answers are the audit of the vendor. Agreement on the population is what makes the two scores comparable.