Why most AI visibility rankings are hard to compare
AI answers change with the prompt, engine, model version, market, language and date. A ranking that does not state these is not reproducible, and two rankings that differ on any of them are not comparable.
The minimum a fair benchmark fixes
Before any number is published, the benchmark should declare:
- The prompt set: the exact non-branded questions, and how they were chosen. Branded prompts measure something else.
- The competitor universe: which brands are counted, and how brand names and aliases are matched.
- The market and language for every prompt.
- The engines, and the measurement window in which all of them were queried.
- How many answers each figure is based on.
Report unavailable states, not zeroes
Some runs return no AI answer, a refusal or a technical failure. These are reported as their own states. Counting them as zero visibility would penalize brands for engine behaviour that has nothing to do with them.
Keep the metrics separate
AI Share of Voice, Discovery Visibility, Recommendation Rate, Owned-Site Citation Coverage, Sentiment and Prominence answer different questions. A benchmark that blends them into one /100 score hides which of them moved. Publish them side by side, each with its own definition.
Make the result checkable
Publish the methodology with the results, date every figure, and keep the answer evidence behind each number so that a brand can check how it was counted. A benchmark that cannot be checked is an opinion.
Use a benchmark as a starting point
A category view shows where a brand stands on one scope at one time. To act on it, track the brand's own prompts and competitors over repeated windows, as described in the SOV Tracker methodology.