An SEO audit that refuses to grade myths

A red warning feels like a fact, even when it isn't.

Plenty of SEO tools flag keyword density, a "domain authority" number or a title that's a few characters too long. A client sees red and asks for a fix. Developers spend a sprint on something that was never a ranking signal, and the real problems wait.

The rule catalogue has a section of things to reject.

Each one is either disproven or widely misunderstood, and shipping it would damage the tool's credibility:

• keyword density
• third-party "domain authority" style scores presented as a Google signal
• the meta keywords tag
• title or description length as a ranking factor (it's a truncation warning, nothing more)
• word count as a ranking factor
• AMP as a ranking factor
• llms.txt as a visibility factor (detect it, never score it)
• a Lighthouse SEO score treated as a ranking

If a finding can't be defended, it doesn't ship.

Every rule in the catalogue lists its method, threshold, severity, data source and citation. When someone asks "says who?", the report answers. That turns an argument about opinions into a conversation about evidence.

A number without context is a guess.

The dashboard won't show a score on its own. It shows a band (excellent, good, fair, poor) and how much of the site the measurement actually covered. 117 of the 145 rules run on raw HTML. The other 28 need a rendered page or Lighthouse. If those didn't run, the report says so instead of pretending.

Each rule must prove it can pass, fail and stay out of the way.

There's a small set of deployed fixture sites with defects planted on purpose. Every rule needs a fixture for pass, fail and not applicable. A test also stops the three pillars (technical, relevance, authority) from importing each other, so a change in one can't quietly move another's score.

The model is a reading aid, not a source of truth.

The AI layer is opt-in. It reviews meaning-level questions, like whether two pages compete for the same query. Every answer must point to evidence spans that are validated against the page. It never invents target queries, rankings or competitor facts. The cost of every run is logged, and each run has a ceiling.

This site ships one.

My portfolio has an llms.txt, for developer tools and AI assistants that read it. It isn't there for rankings. My engine would report that it exists, and give it zero points. That's exactly the point.