Audit scoring rubric — seven dimensions, one weighted site-health score

The scorecard is the audit's headline claim, so it must be reproducible: two auditors handed the same audit.json measurements should land within ±5 points per dimension. This file makes that possible — every dimension has observable anchors at 40 / 70 / 90, and the procedure below forbids scoring without evidence.

The seven dimensions

Weights sum to 100. Conversion focus carries the most weight because the audit's promise is a better business outcome, and the conversion path is the shortest lever; brand expression and accessibility carry less not because they matter less but because their business effect routes through the other dimensions.

Anchors

Score each dimension by finding the band the evidence fits. Anchors are observable states, not adjectives — if you cannot point at the measurement or screenshot that satisfies the anchor, you are in the band below.

brand-expression — does the site express its own brand, or a template's?

• 40 — brand-color share of painted pixels under 5%; design tokens are framework defaults while the actual brand color lives outside the token layer (T-tokens-unused); typography is a bare system stack with no display register; the page could belong to any company in the vertical.
• 70 — recognizable palette applied to primary CTAs and header (brand-color share roughly 5–15%); one consistent display face; but at least one unresolved brand tension remains (a palette color used as text only, a fragmented radius vocabulary, a temporal mark carried past its campaign).
• 90 — brand color deliberately deployed across surfaces (share ≥15%, or a documented restraint strategy where whitespace itself is the brand move); motif vocabulary coherent (≤2 radius values in active use); typography carries a distinct display register; no T-* brand tensions fire, or each firing tension is a deliberate, explainable choice.

visual-hierarchy-craft — does the eye know where to go, and does the craft hold up close?

• 40 — type scale is ad-hoc (typeScale.kind === "ad-hoc", no consistent ratio); three or more competing focal points in the first viewport; the VISION pass names a dated pattern dominating the hero (stock-photo hero + double CTA, carousel-as-homepage, wall-of-cards); spacing is inconsistent between sibling sections.
• 70 — a consistent scale ratio holds on the home page and there is a single clear focal point, but secondary pages fall apart, hierarchy is carried by font-size alone (no weight/color/space support), or section rhythm breaks below the fold.
• 90 — modular scale held across every audited page; deliberate spacing rhythm; the hero communicates what-this-is / who-it's-for / the ask within one viewport; alignment and optical balance survive zooming into any individual section.

conversion-focus — does the page funnel attention to one ask?

• 40 — four or more distinct CTA labels for the same underlying action (ctaFragmentation on one equivalence bucket ≥4); no primary CTA in the first viewport, or the primary CTA competes with ≥2 equal-weight alternatives; pricing or what-it-is is not reachable within one click of home.
• 70 — one canonical CTA voice, visible in the first viewport, but the journey stalls: no mid-page reinforcement of the ask, proof elements are generic ("trusted by thousands"), or the nav contains competing actions (T-nav-conflict).
• 90 — a single canonical CTA voice with clear primary/secondary hierarchy; a stated value proposition in the hero; proof elements (customers, numbers, testimonials) sourced from real content rather than boilerplate; the ask restated at natural decision points down the page.

accessibility — can everyone use it?

• 40 — a body-text/background pair in the captured palette computes below 3:1; over 30% of images carry empty alt (T-img-alt-empty); no <main> landmark; content-free link labels present ("click here", "read more" as bare links).
• 70 — body text meets 4.5:1 but at least one accent-on-background pair in an interactive role fails AA; alt coverage 70–95%; landmarks present but heading order skips levels or <h1> count ≠ 1 on some page.
• 90 — every captured palette pair meets AA in the role it is actually used in; ≥95% of content images carry meaningful alt; full landmark set (banner / main / contentinfo / nav); exactly one <h1> per page with ordered heading levels; no content-free link labels; animation respects prefers-reduced-motion where motion exists.

technical-seo — can search engines crawl, index, and represent it?

• 40 — sitemap absent or invalid; titles missing or duplicated across pages; no meta descriptions; no canonical; <h1> count ≠ 1 on audited pages; http does not redirect to https cleanly.
• 70 — crawlable and indexable (robots.txt sane, sitemap declared and valid, canonicals present, https enforced) but metadata quality is weak: titles truncate past ~60 chars or bury the value term, descriptions are generic, Open Graph incomplete, JSON-LD absent or Organization-only.
• 90 — full crawlability; unique, intent-bearing titles ≤60 chars; descriptions that read as answers to the page's query; complete Open Graph on shareable pages; valid page-type JSON-LD on every audited page; redirect chains are a single hop to the canonical host.

content-llm-visibility — will an AI answer engine find, understand, and cite it?

• 40 — the site's key facts (what it is, who it's for, pricing) live only in images, JS-rendered widgets, or are simply never stated; key pages carry under ~150 words of extractable prose; no structured data; headings are slogans ("Dream bigger") rather than topics.
• 70 — key facts stated in crawlable text and each page answers its primary question somewhere, but the answer is buried mid-page; schema covers a single entity type; no llms.txt; few headings map to questions a user would actually ask.
• 90 — each audited page answers its askable question in the first extractable prose block; schema.org covers the organization plus the page-type entities (Product / Service / FAQ / Article as applicable); headings scan as questions or topics; llms.txt present (or a deliberate, stated equivalent); content is deep enough to cite — specific claims, numbers, named capabilities rather than adjectives.

performance — does it load fast enough not to lose the visitor?

Classify each metric against Google's thresholds first, then band:

• 40 — any Core Web Vital classifies poor on mobile.
• 70 — nothing poor, but at least one mobile metric sits in needs-improvement; desktop is good; the causes are identifiable (render-blocking chain, unsized images, heavy third-party script).
• 90 — every measured metric classifies good on both mobile and desktop; images sized and lazy-loaded below the fold; no render-blocking chain beyond the first stylesheet.

Scoring procedure

1. Band first. Pick the anchor band the evidence fits. An exact anchor match takes the anchor's score; evidence between two anchors interpolates within the band. A profile worse than the 40-anchor scores in the 0–39 range at the auditor's judgment — but the judgment must still name the observations that put it there.
2. Evidence floor. Every scored dimension cites at least two evidence references in audit.json#scorecard.dimensions.<id>.evidence — measurement keys, finding ids, or tension ids. No evidence → the dimension is not scored (see rule 3), never guessed.
3. Not-measured handling. When a dimension's inputs could not be taken (e.g. CWV unmeasurable because the origin blocks lab agents), set "score": null and record the reason in notMeasured. The overall renormalizes over the measured weights, and the report states the omission in the scorecard hero — a missing dimension is visible, not silently averaged.
4. Overall. overall = round( Σ(scoreᵢ × weightᵢ) / Σ(weightᵢ over scored dimensions) ). Integers only, no decimals anywhere in the scorecard.
5. Judgment discipline. Wherever an anchor requires judgment ("weak metadata quality"), the report prints the observed value next to the judgment ("title: 'Home' — 4 chars, no value term") so the reader can re-derive the call.

Severity is orthogonal to score

Findings carry P1/P2/P3 severity (P1 = actively losing business or excluding users; P2 = material drag; P3 = polish). Severity prioritizes the fix list; anchors set the score. A dimension can score 70 and still carry a P1 (one severe defect in an otherwise competent surface), and the scorecard must not be back-derived from severity counts.