What an evidence-based audit actually contains
By Kestralia editorial team··5 min read
The minimum fields
Every finding should answer seven questions:
• **What** is wrong, in one sentence a non-specialist understands.
• **Where** it happens, with exact URLs.
• **How we know**, with the raw signal: HTTP status, header, canonical value, JSON-LD block.
• **When** it was observed.
• **How confident** we are, and why the confidence is not 100%.
• **What to do**, specific enough to execute.
• **How to verify** that it is actually fixed.
If a tool cannot produce these, it is producing content, not diagnosis.
Confidence is not decoration
A 5xx response is a fact: confidence 1.0. "Thin content" is a heuristic: 0.75 at best. "Possible cannibalization" is a model inference from token overlap: 0.65. Showing the number forces the tool to be honest about which kind of claim it is making.
Reproducibility
Run the same crawl on the same site twice with no changes and the findings should match. If the score jumps by ten points because of a timing difference, the score is noise. Kestralia groups findings by type, applies a documented prevalence factor and caps deductions per category, so one slow response does not randomly reorder the whole report.
The verification loop
The most important field is the last one. A finding marked "fixed" by a human is a claim. A finding that is absent from the next crawl is evidence. Products should be judged on the second, not the first.
Why this matters commercially
Agencies lose clients when a report cannot survive a single "how do you know?". Evidence-based reporting is slower to produce manually and much faster to defend. That is the entire opportunity.