Build your own audit prompt set.

Fewer than about forty and a single unusual result skews the whole picture. Fifty across a spread of treatments and phrasings gives a score stable enough to compare month over month.

Engine outputs vary between runs. A single result tells you very little. Three runs and a majority verdict removes most of that noise without becoming a research project.

No. This is deliberately manual — open each engine, paste each prompt, record what comes back. It is tedious rather than difficult. Budget two to three hours for fifty prompts across four engines.

Four things per prompt: were you named, who was named instead, in what position, and which source the engine cited. That last column is the actual diagnosis — everything else is symptom.