Website Score

Website PM Score

Evaluate B2B SaaS websites against first-principles product marketing criteria. Generate an actionable PM Score (0-100) that identifies specific gaps and provides prioritized, wireframed recommendations. Knowledge type: website-score per .claude/rules/ontology.md.

Research substrate

Default substrate: Exa (per .claude/rules/exa-protocol.md, auto-loaded). Primary tool: web_fetch_exa for clean page extraction. Migration window: prefer plugin namespace mcp__plugin_exa_exa__web_fetch_exa once installed; legacy mcp__exa__web_fetch_exa still mounted as fallback. Citation: every Exa-derived claim uses [VERIFIED: exa_search, {url}, accessed {YYYY-MM-DD}]. Quality gate: ≥3 sources per major claim, ≥50% [VERIFIED], no fallback to WebSearch without flagging the data gap. Worked examples + tool catalog: .claude/skills/meta-skills/exa/.

When to run

Invoke when the user says: evaluate this website, score this website, website audit for [URL], PM score for [company], product marketing score, website assessment, homepage analysis, landing page audit, rate this website, what's wrong with this site, review [URL], website teardown, messaging audit. Do NOT invoke for: landing page copy writing (use /landing-page-copy), competitor research without scoring (use /competitor-research), positioning work (use /product-messaging), ICP research (use /icp-behavioural).

Two assessment modes — pick by depth needed:

• Quick — 5 categories scored, normalized to /100. Best for lead-magnet output, prospect triage, fast diagnostic.
• Full — 9 categories scored, raw /100. Best for active client engagements, comprehensive teardown, before/after planning.

Visual phase map → the premium reference. Full scoring rubric → the premium reference.

Inputs

Required:

• website URL — primary site to evaluate (must be accessible/fetchable).
• assessment mode — Quick (5 categories) or Full (9 categories). User specifies or inferred.

Recommended (improve quality):

• competitor URLs — enables competitive teardown mode (element-by-element comparison).
• vertical context — adjusts expectations for industry.
• specific focus areas — weights certain categories higher.
• known positioning — helps evaluate differentiation accuracy.

If website URL missing, ask. If mode unclear, confirm Quick (lead magnet) vs. Full (comprehensive). If competitor URLs invalid, drop competitive context section without blocking.

Steps

1. Validate inputs — confirm URL is fetchable, mode is set, competitor URLs (if any) are valid.
2. Fetch primary pages — homepage required; Full mode also pulls pricing, features, customers, about. Capture hero section verbatim. Use web_fetch_exa for clean extraction.
3. Apply 5-second test — first viewport only. Check for tool anchor OR task anchor (one suffices). Frameworks → the premium reference ("The Amorphous Software Problem"). Output pass/fail with rationale.
4. Competitor fetch (if competitor URLs provided) — fetch homepages, apply same 5-second test, note comparative observations.
5. Score Quick categories (both modes) — Product clarity /12, Audience specificity /12, Problem articulation /15, Differentiation /15, Action clarity /10. Each score requires verbatim evidence from fetched pages.
6. Score Full categories (Full mode only) — Outcome clarity /15, Trust & proof /10, Risk reduction /6, Product visualization /5. Same evidence rule.
7. Calculate total + interpretation — Quick: sum to 64, normalize to /100. Full: raw /100. Apply score interpretation table (the premium reference): 85-100 Excellent / 70-84 Good / 55-69 Needs work / 40-54 Weak / 0-39 Critical.
8. Per-ICP matrix (if ≥2 personas provided) — score per persona × per category instead of single number. Per-persona evidence + recommendations + priority. Full protocol → the premium reference ("Per-ICP scoring mode").
9. Prioritize recommendations by tier — Fix this week (blocking copy-only), Fix this month (structural/new content), Fix this quarter (strategic/research-required). Tier criteria → the premium reference ("Recommendation tiers").
10. Create before/after wireframes — hero rewrites, key section additions, CTA improvements. Verbatim "before"; concrete copy in "after". Box-drawing format per template.
11. Estimate impact — apply benchmarks from the premium reference and the premium reference. Provide ranges, not precise numbers; cite basis.
12. Add competitive context (if competitor URLs provided) — element-by-element comparison matrix (hero anchor, status quo named, differentiator, gap assessment).
13. Self-evaluate against guardrails → the premium reference ("Anti-hallucination guardrails" + "Pre-delivery quality checklist"). Mark inferences explicitly. Note "Not quantified" for missing metrics.
14. Write output per template → the premium reference (full structure: score summary, 5-second test, category analysis, prioritized recommendations, competitive context, iteration prompts).
15. Offer iteration prompts post-delivery → the premium reference (refinement / expansion / quality offers + auto-update protocol).

What good looks like

Evaluations (binary pass/fail before declaring "done")

• All applicable categories scored (5 for Quick, 9 for Full).
• Every score has verbatim evidence quote from a fetched page.
• Scores sum correctly; interpretation matches range.
• 5-second test result includes pass/fail + grounding check (tool anchor / task anchor presence).
• ≥1 recommendation per tier (Fix this week / month / quarter).
• Before/after wireframes for hero changes use box-drawing characters and verbatim "before".
• No invented metrics — every claim quantified is on the site or marked "Not quantified".
• No assumed pages — only score what was fetched.
• Output title is # Website PM Score: [Company Name] exactly.
• Per-ICP matrix present if ≥2 personas were provided.

Integration with other skills

Recommended chains:

• Prospect qualification: company-context → website-score (Quick mode for triage).
• Active client teardown: competitor-research → website-score (Full) → landing-page-copy.
• Multi-persona client: icp-behavioural → website-score (per-ICP matrix) → product-messaging.

Pre-slim original

Pre-slim SKILL.md (777 lines, v2.0) archived at .claude/skills/_archive/website-score/SKILL-pre-slim-20260429.md. See the premium reference ("Changelog") for the v2.3 entry documenting the slim.

Sourced patterns — technical SEO discipline

Add these technical-SEO checks to the audit rubric. Each is a sharp threshold with a Google-enforced consequence:

• Core Web Vitals thresholds (Google ranking signal).
  • LCP (Largest Contentful Paint) < 2.5s
  • INP (Interaction to Next Paint) < 200ms
  • CLS (Cumulative Layout Shift) < 0.1 Pages failing any of these are penalized in mobile + desktop ranking. Check via PageSpeed Insights.
• Title tag: 50–60 characters with keyword near the beginning. Above 60 truncates in SERP; below 50 underuses real estate.
• Hreflang reciprocity rule. Multi-language pages require self-referencing hreflang entries AND reciprocal links between languages. Missing either invalidates the cluster — Google ignores the entire setup.
• E-E-A-T signals checklist. Experience (author has used the product/service), Expertise (credentials visible), Authoritativeness (citations + inbound links), Trustworthiness (HTTPS, clear contact info, transparent ownership).
• "Important pages within 3 clicks of homepage" rule. Pages 4+ clicks deep are effectively invisible to most users + crawlers. (Paired with /site-architecture Step 1.)
• One <h1> per page; logical hierarchy (no H1→H3 jumps). Multiple H1s confuse content-importance signals.
• JS-injected schema caveat. web_fetch and similar tools cannot detect JavaScript-injected schema. For pages using client-side schema rendering, validate via Google's Rich Results Test (Google does execute the JS) — not via fetch tools. (Paired with /schema-markup Step 5.)

Priority hierarchy when auditing: Crawlability → Indexation → Speed → On-Page → Authority. Fix in that order; later-stage fixes don't matter if earlier stages are broken.