Each question goes to the model cold, with your brand name deliberately left out of the prompt. That is the entire measurement: whether the model volunteers you unaided. Mention the brand to set context and it will dutifully talk about you, which would make the tool a flattery machine rather than a check.

The answer is then scanned for your name on a word boundary, so "Clay" doesn't match inside "Clayton". A second, separate call extracts the products the answer actually named — that is a model call rather than a regex because vendor names have no lexical pattern, and a capitalised-word heuristic returns sentence openers and little else.

Rate limiting is charged per question, not per request, because each question is a real model call. Asking three in one go costs three.