A brand was named – but did the reader get a useful recommendation?

A mention is an observable line in which a system names a brand, product, or organization. It may be accurate, neutral, incidental, or placed in the wrong category. Preserve the complete answer and mark the exact passage; do not assign an intention to recommend.

Usefulness starts with a different check: can the reader tell which job the option fits, what they will get, and when the choice may fail? This evaluates visible content, not the model's hidden process.

Split the answer into five fields: the person's scenario; reason to choose; condition of fit; limitation or risk; and next action. If a field is absent, record 'not stated'. Missing detail is not automatically an error, but it shows what the reader lacks for a decision.

This coding is more useful than a single 'good answer' label. It explains why two texts containing the same name can give a reader very different confidence and reasons to act.

Save the exact prompt, service, mode, model, language, region, date, account state, search availability, full answer, and visible links. Do not merge ChatGPT, Perplexity, Gemini, Claude, Alice, GigaChat, Grok, and DeepSeek into one metric: they are different observation environments.

Repeat one scenario under a rule written in advance. One answer shows a state at one moment; it does not prove a stable recommendation or the effect of a particular site edit.

An answer may contain a name, an explanation, and a link, but these are three separate audit rows. Record what was said, whether an address was visible, and which passage on the opened page supports the statement. A domain near an answer does not prove that it supported a particular sentence.

The next action should also be visible and concrete: open a page, compare conditions, check a limitation, or continue searching. 'Learn more' without a route does not resolve uncertainty.

Instead of an artificial percentage, use four statuses for each field: aligned with the public description; acceptably simplified; distorted; or unsupported by visible evidence. The status describes the text, not the system as a whole. In disputed cases, preserve the quote and the reason for the decision.

Before running the audit, write the public reference in ordinary language. It should be available to readers: category, audience, main differentiator, and limitations. A hidden internal formulation turns the check into a subjective comparison.

A team cannot edit another system's answer. It can make the public context clearer: connect the job, benefit, limitation, and support on one page; remove contradictions between the site and external material; and show an example instead of repeating a slogan.

After an edit, rerun the same prompts and compare the matrix fields. A changed portrait is an observation. Do not attribute it to one edit without comparable periods, repeats, and alternative explanations.

This column does not measure a model's internal knowledge, answer probability, position, referral, or purchase. It describes what a reviewer saw in a saved answer. An unseen support remains unknown, and no logged referral does not mean no influence.

Editorial conclusion: audit usefulness as the chain 'scenario → reason → condition → limitation → action'. It helps a person read an answer critically, but it is not a promise of higher AI visibility.