Every claim in a UX Signal report is tied to something observed on the page. This is the whole procedure, including what it deliberately refuses to do.
The page is loaded in a real browser and photographed at desktop, tablet, and mobile widths. While it is open, code measures what a screenshot cannot settle: contrast ratios, tap-target sizes, sideways scroll, and how far the layout moves as it loads. Judgements come from the rendered page, never from a description of it.
Each capture is checked against the heuristic set for that experience type, weighted by how much the category moves a first impression. An e-commerce page is not scored on the same rules as a marketing site, and a logged-in product carries a rubric of its own.
The score starts at 100. Every confirmed issue deducts its points once, at its root cause, so a single problem appearing in three places is charged once rather than three times.
Findings without evidence are recorded but never scored. Strengths and manual checks appear throughout for context and cannot raise the number.
Not an average of category grades, and not a model's opinion of your site. Each confirmed issue carries a point value; the report shows exactly which categories those points came from, and they add back up to the difference between your score and 100.
985 pages auditedmedian score 7536% score 80 or above
Score distribution
Every completed page audit, bucketed by Signal score. No domains, no names — counts only, updated live from the same table your report is written to.
Where the web loses points
Mean stored score per category across every audit, worst first — the categories at the top are where the sites we scan are weakest today.
Checkout friction is life or death for a store and irrelevant to a marketing site. Pick an experience type and watch the ten categories re-rank by the weight the audit will actually apply.
Category weights · Marketing site
Rendered from the same configuration the scoring pipeline reads — this table cannot drift from what actually runs.
The 56 heuristics are not invented opinions. Each one operationalizes a principle the field already agrees on into a failure you can point at on a rendered page.
The Nielsen tradition — hierarchy, navigation as a map, matching the reader’s language, errors that explain their own fix — made checkable on a rendered page.
WCAG-derived basics: 4.5:1 contrast, accessible names on controls, meaningful alt text. We check the fundamentals; we do not certify conformance.
The mobile ground rules platforms converged on — 44px tap targets, no horizontal scroll, layouts designed for phones rather than surviving them.
What checkout and form studies keep re-finding: surprise costs abandon carts, account walls tax first purchases, every needless field loses someone.
The reasoning engine is a general model — Claude — but the judgement is this rubric, enforced by scoring code that only lets evidence-backed findings move the number. Where a rule has a number in it (contrast, tap-target size, sideways scroll, layout shift, load timing), code measures it in your page and the model rules from the measurement, not from how a screenshot looks. Every Pro and Team report also carries an unscored “How AI reads your site” section: the same machine-readability lens, turned on your page. Reports on a live page add stress tests: what moved while it loaded, and what breaks when the text gets longer.
A method is only credible if it names its own edges.
