A mention is a signal. Rank adds context.
Understand brand mentions, recommendation positions, evidence, sentiment, and missing answers without turning an AI sample into a misleading score.
What counts as a mention?
A mention records whether the answer identifies a tracked brand through its name, configured alias, or domain. The saved evidence quote helps you check what the model actually said.
Always inspect ambiguous names. A common word may refer to something other than your company, and a passing reference is not the same as a product recommendation.
When should a brand have a position?
Position describes a recommendation order that the answer actually establishes. If the answer presents an ordered list of recommended products, a brand may have a clear first, second, or third position.
If the answer only mentions a brand in prose, groups products without order, or discusses it in a caveat, its recommendation position may be unranked. That does not mean the mention disappeared.
Keep the denominator visible.
Suppose 10 questions are sent to three providers. That creates 30 possible provider answers. If only 20 are saved, a result calculated from those 20 answers describes collected coverage, not the entire planned set.
Questions with no saved answer are unknown. They should not be reported as “brand not mentioned.” Compare periods with similar question sets and provider coverage.
Use evidence before sentiment.
Sentiment is an interpretation of how the answer discusses a brand. Positive language can help explain a mention, but it does not prove buyer preference or a commercial outcome.
Open the full answer, locate the evidence quote, and check whether the language supports the recorded mention, recommendation position, and sentiment. Treat unclear cases as review work.
Report a pattern, with its limits.
A useful summary might say: “Our brand appeared in 12 of 20 collected answers across these questions and providers.” Include the dates and collection method.
Avoid claiming the result measures all AI users or all possible prompts. Answers vary by provider, model, context, available sources, and time.