Separate Mentions, Recommendations, and Citations
Being mentioned is different from being recommended. Define each metric and return to the original answer to interpret the result.
On This Page
Mentions, recommendations, and citations answer different questions. Before interpreting a metric, define what is counted and what forms its denominator.
Mentions: Does the Brand Appear?
A mention records the appearance of a brand name or verified alias. Exclude unrelated entities with the same name, then read the context: an answer may recommend the brand, describe a limitation, or correct an error.
A mention is an observation, not automatically a positive assessment.
Recommendations: Is the Brand Presented as a Choice?
Interpret recommendations in context. A suggested option, a conditional candidate, and a background reference are not the same.
Keep the constraints in the prompt. A recommendation for one budget or use case cannot be generalized to every customer.
Citations: What Material Supports the Answer?
Citations concern linked or referenced sources. A brand website being cited, a third party discussing a brand, and an answer recommending that brand are distinct relationships.
Check whether a link is accessible and whether its content supports the adjacent claim, identifies the correct entity, and covers the relevant period.
Define the Denominator
State the valid answer count, treatment of failed samples, prompt scope, and model scope. A failed collection should not silently become an observation that the brand was absent.
When a sample is small or volatile, raw counts and conditions are more useful than percentages presented with excessive precision.
Return to the Evidence
Investigate explainable changes: which prompt group moved, how descriptions changed, which sources were used, and whether the result repeats under comparable conditions.
Use the baseline method to preserve each observation. Current brand edits should not rewrite the interpretation of historical results.
