The server fetches the page (SSRF-guarded), strips it to readable text plus headings and CTA labels, then scores it against a fixed six-area CRO rubric so two runs are comparable. Every finding quotes the page and proposes one concrete fix. Judges copy and structure, not rendered design.
The rubric is fixed on purpose. Ask a model to 'review this page' and you get whatever it feels like mentioning, which makes two runs incomparable. Scoring the same six areas every time turns it into something you can re-run after changes and actually track. Note the limit we state rather than hide: this reads copy and structure, not the rendered design.
