Five red flags (walk away)
Promises rankings or fixed positions
Nobody controls model answers. The more confident the pitch, the more suspicion it deserves.
No raw answer evidence
Scores and screenshots without traceable raw answers cannot be checked or reproduced.
No stated test conditions
Which model, which questions, when, and how counted — if any of these is unclear, the results are not comparable.
Static scores passed off as real AI recommendation rates
A rule-check score says whether signals are complete, not whether a model recommends you. Blending the two is a common packaging trick.
Filling your gaps with "industry common sense"
Inventing untested data from industry habits is not auditing — it is storytelling. Seeds, prices and audiences extracted are counted as extracted; the missing must be marked.
Three things to demand
Raw answer evidence
Every conclusion traces back to a real answer, with model and collection time noted.
A written methodology
How the question set is built, how category and brand questions are layered, and what the counting rules are — in writing.
Before/after comparison under identical conditions
Re-test with the same conditions after fixes. Whether it improved is decided by the comparison, not the promise.
Five questions to ask (each one filters a vendor)
Do you count mention rate or recommendation rate, and how do you distinguish them?
If they cannot answer the layering, the report is questionable.
Are category questions (no brand in the prompt) and brand questions counted separately?
Mixing the two is the number-one source of distorted data.
Can you show me the raw answers, matched conclusion by conclusion?
Ask on the spot; watch the speed and completeness of the reaction.
Which model and date produced this? Would the conclusion hold on another model?
Tests whether the vendor honestly states sample boundaries.
Which conclusions were measured, and which are projections?
Good providers separate fact from projection by default; vague answers deserve a discount.
Further reading: the same material is also published in our Zhihu knowledge base (self-published, not independent evidence). Open the knowledge base
Free tools, no sign-in needed