A website AI-search-readiness diagnostic produced scores of 72, 85 and 61 for the same page. The developer moved verifiable checks such as robots.txt, sitemaps, HTML and TLS into code, leaving narrowly scoped qualitative questions to the LLM. Checks inspect actual content and allow-rule exceptions rather than merely accepting HTTP 200, and distinguish absent files from invalid ones. The pipeline excludes irrelevant input such as comments and validates structured outputs. Reduced variation is a reported implementation outcome, not an independently benchmarked accuracy guarantee.