Evidence
How to Prevent AI From Inventing Website Problems
Preserve factual website evidence before asking AI to qualify a prospect or write outreach.
By Brian Woodson ·
Separate collection from interpretation
First collect current website content and observable behavior. Then ask the qualification layer to evaluate that material. When the same prompt is expected to browse, infer, judge, and write outreach without structured evidence, unsupported claims become harder to detect.
Require structured output
Store the exact finding, source URL, evidence excerpt or technical observation, business consequence, qualification decision, and rejection reason. A structured record makes every downstream message traceable.
Allow rejection
The model must be allowed to return “not qualified.” If the workflow expects every record to become outreach, it creates pressure to turn weak observations into invented problems.
Keep a verification gate
At lower tiers, a person verifies the evidence. At automated tiers, schemas, source capture, confidence rules, tests, and sampling preserve the same gate. Increasing automation should not lower the evidence standard.
Choose the prospecting system that fits.
Compare four evidence-first implementation levels and start with the lowest tier that meets the need.