Published methodology

How the score is calculated

The score is deterministic: the same page scored twice gives the same number. Nothing about it is a model's opinion — the optional AI read in paid reports is evidence next to the score, never part of it.

1. 1

   Collect the page the way a crawler gets it

   One request for the URL with an honest user agent, redirects followed by hand and re-validated at every hop, a 15-second per-hop timeout and a 4 MiB page cap backed by structural DOM limits. Then robots.txt, llms.txt, Markdown and the sitemap are collected inside one 60-second evidence budget. A 200 that returns HTML where plain text belongs is treated as missing, because SPA hosts answer every path with the app shell.

2. 2

   Extract facts, never prose

   The HTML is parsed once into plain facts: title, meta, headings, links, images, JSON-LD blocks, microdata types, and the visible text with scripts and templates removed. Nothing is judged at this stage.

3. 3

   Run 31 independent checks

   Each check reads those facts and awards part of its own weight. Checks never read each other, so one failure cannot cascade. A check returns points, the evidence it found, and — when it did not score full marks — the fix.

4. 4

   Sum into 6 categories

   Weights are chosen so that the failures that actually remove you from AI answers cost the most. Being blocked in robots.txt for answer engines is the single heaviest check on the page.

5. 5

   Total and grade

   The five core category weights add up to exactly 100. Agent readiness can add up to 8 bonus points, and the displayed total is capped at 100. The grade is a band on that total, while the verdict names your weakest core category.

6. 6

   Order the fixes

   Every check that lost points contributes its fix. They are sorted by recoverable points first and impact second, so the list answers "what buys the most score for one change" rather than "what is most interesting".

Back to the scanner