1. How many answers is it built from?
A percentage without a count is not evidence. Under about twenty answers it is an anecdote.
2. What is the range?
The likely range shows where the real figure probably sits. A wide range means more answers are needed, not that the tool is broken.
3. Does the prompt name the brand?
Asking 'is Acme any good?' will always mention Acme. Those questions should be kept apart from the headline, which should measure whether AI brings the brand up unprompted.
4. Which surface was measured?
Consumer product or API, with or without web search, from which country. If the report does not say, ask.
5. Can you read the answers?
Every figure should open into the stored answers with the match highlighted. If you cannot see why a mention counted, you cannot defend it to a client.