Branded and scenario prompts measure different things

When a company name is already in the prompt, the system does not need to decide which brand fits the situation. It can explain, compare, or clarify a preselected named subject. That is useful for testing recognition and answer completeness, but it is not a test of whether the brand emerges in a natural choice.

A scenario prompt starts with a person’s task: choose a tool, service, or next step. No name is supplied, so a brand’s appearance becomes a separate observation. It still does not prove that the recommendation is right, but it answers a different practical question.

This test can show which characteristics the system associates with the brand, which limitations it mentions, and whether it gives a checkable next step. It is useful for a named subject audit: do the description, conditions, audience, and links match what the editorial team considers important?

A strong result here may only mean that the brand answers a direct request about itself well. It does not say whether the system would choose it among competitors, and it does not replace a test without the hint.

A scenario set reveals which tasks and attributes lead to a brand mention. Record more than appearance: list position, explanation of the choice, presence of a link, fit with constraints, and whether the claim can be checked on the site.

If the brand does not appear, that is not automatically a failure. The prompt may not describe the real audience, the constraints may be too narrow, or alternatives may fit the task better. Treat absence together with the prompt and the intended choice criterion.

Start with pairs: a direct prompt about the brand and a scenario prompt for the same task without the name. Record language, country, service mode, date, and account when they affect the run. Do not rewrite the wording after the first run merely to make the comparison look better.

Store the full response and visible URLs in the result record. Mark mention, link, scenario usefulness, and factual error separately. A single “visibility” column hides important differences and creates false precision.

A team may see the brand in a branded response and call it a recommendation. Or it may miss the brand once in a scenario and call the page invisible. Both conclusions exceed the observation. Use repeated runs and write the success rule before running them.

It is especially risky to combine rates from different sets: they have different total counts and task difficulty. The sum of branded and scenario appearances is not an overall AI-visibility rate.

For each topic, save both prompt versions, the measurement goal, and the reader action you expect. After the run, write separate conclusions for recognition and choice. If results conflict, publish the conflict as an observation instead of forcing one attractive formula.

The practical recommendation is simple: never treat a branded test as a substitute for a scenario test. They are two layers of one editorial diagnosis, both needed to understand not only “does the system know us?” but also “does it help a person choose?”