Tools
What each number means
Tells
- Tells, total
- Findings from impeccable, restyle's own rules and axe that mark the page as default AI output or broken.
- High severity
- Tells a visitor notices in the first seconds: gradient heroes, gradient text, the centred-pill-three-cards hero.
- Second-order clichés
- The looks that replaced the first wave: cream, serif and terracotta; near-black with one acid accent; hairline broadsheet.
Accessibility
- axe serious + critical
- axe-core violations rated serious or critical, counted per rule.
- Contrast failures
- Text elements below WCAG 2 contrast (4.5:1 body, 3:1 large text), measured on computed colours.
Lighthouse (lab, mobile)
- Performance
- Lighthouse performance score, 0-100, one lab run with simulated mobile throttling.
- Accessibility
- Lighthouse accessibility score, 0-100 (an automated subset; not a full audit).
- Best practices
- Lighthouse best-practices score, 0-100.
- SEO
- Lighthouse SEO score, 0-100.
- LCP
- Largest Contentful Paint: when the biggest element above the fold finished rendering.
- CLS
- Cumulative Layout Shift: how much the layout jumps while loading. Under 0.1 is good.
- TBT
- Total Blocking Time: main-thread time blocked by long tasks, the lab stand-in for INP.
Typography
- Distinct font sizes
- How many computed font sizes the page uses. Descriptive: a deliberate scale has 5 to 8.
- Size entropy
- Shannon entropy (bits) of text length over font sizes. Descriptive: low means one size does nearly all the work.
Palette
- Hue entropy
- Shannon entropy (bits, 12 bins) of chromatic pixel area by hue. Descriptive.
- Chromatic colours
- Distinct colours with chroma above 0.04 covering more than 1% of the screenshot. Descriptive.
Model-based
- Distinctiveness
- 0-1, image-embedding distance from a reference set of default AI-generated sites. Higher is less generic.
- Prototypicality, structural
- 0-1 share of this niche's layout and content conventions the page keeps (phone number, hours, prices...).
- Prototypicality, niche match
- 0-1 zero-shot probability that the screenshot reads as its stated niche.
How the model-based scores are computed
Distinctiveness embeds the page's screenshots with an image model and measures how far they sit from a reference set of landing pages generated by AI coding tools with default prompts. It is reported as 0 to 1, where 1 means further from every reference page. This run used the refset reference set and the Xenova/clip-vit-base-patch32@q8 embedding model. Before the redesign, the closest reference was s28 at similarity 0.78.
Prototypicality has two parts. The structural score is the share of this niche's expected conventions the page keeps, checked in the DOM. The niche match is a zero-shot image classification: the probability that the screenshot reads as its stated niche rather than a neighbouring one. Visitors trust pages that look like their category, so a redesign should be distinctive without losing these.
Limits
- Lighthouse numbers come from one lab run with simulated mobile throttling, not from real visitors. Treat differences of a few points as noise.
- INP cannot be measured in the lab; TBT stands in for it and only approximates interaction delay.
- axe and the contrast check find a subset of accessibility problems. A clean result is not a full accessibility audit.
- Distinctiveness depends on the reference set: a page can be far from default AI output and still be unremarkable.
- The judge is a single model judgement per pair, not a panel of people.