An audit that has to survive someone trying to refute it.

A structured diagnosis of a site against the ATLAS framework: technical foundation, search architecture, content and intent, entity clarity, and AI visibility. Every finding is scored against the published 100-point scale, so a result from one engagement means the same thing as a result from another.

It is not a template filled in from a crawler export. Every row in the output is a finding a person looked at, argued against, and could still defend.

Most of what an automated crawl reports is not wrong, but a meaningful fraction of it does not survive a second look — a canonical that looks broken until you check what it is consolidating onto, a redirect that looks like a loop until you trace where it actually lands. Atlas argues against its own findings before a client sees them.

On the one engagement this site can name: 158 findings were produced, and 44 were refuted outright before the client saw them. That is the discipline, not the exception.

Technical foundation carries 20 of the 100 points — the largest single weight, because it is the only layer whose failure is total. Search architecture and content & intent carry 15 each. Ecommerce, international, entity, authority and AI visibility carry 10 each, scored only where they apply to the site in question.

The weights are fixed and published, not adjusted per engagement, so a score from an Atlas audit can be compared to another Atlas audit rather than read in isolation.

Two steps most audits skip: hostile verification, where every finding is argued against rather than reported on trust, and instrument control, where the measuring tool itself is checked before a number from it is quoted. A tool that mismeasures produces a confident, wrong number, and that failure mode is worse than an honest "we do not know".

A prioritised defect list, each entry carrying an evidence ID, a verdict and the smallest change that resolves it — the same document format every Atlas engagement runs on, whichever of the five channels it came from. Where a figure cannot be evidenced, the audit says so rather than estimating.

Findings are ordered by what they cost against the published weight of the channel they sit in, not by how easy they are to fix. The output is the smallest change that moves the number — the same principle the two-week working loop runs on afterward.

The audit is the diagnosis. If the findings warrant ongoing work, it continues in two-week loops — diagnose, map, prioritise, build, measure, compound — run by the same person who wrote the audit. No account layer and no handoff to someone reading the document for the first time.