seo-audit

name: seo-audit description: "Audit and optimize any website's SEO + GEO (AI/LLM visibility) + Core Web Vitals, then fix what's broken. Runs a portable, zero-dependency hard-gate audit over a build output directory (Astro/Next/Hugo/Jekyll/plain HTML) and a live-URL crawl/GEO audit (robots.txt AI-crawler policy, sitemap, llms.txt, on-page JSON-LD, canonical, security headers). Includes a battle-tested LCP playbook (7.5s → 1.5s mobile on the reference site): render-blocking CSS, critical-CSS split, font preload discipline, third-party JS deferral. Use when asked to: check/improve a site's SEO, raise a Lighthouse/PageSpeed score, fix slow LCP / Core Web Vitals, make a site discoverable by AI agents (ChatGPT/Claude/Perplexity/Gemini), add structured data, set up robots/sitemap/llms.txt, or review a site before launch. Distilled from the reference site's production build-time SEO gates." compatibility: Claude Code, Claude Desktop, Cursor keywords:

• seo-audit
• seo
• geo
• ai-visibility
• llms.txt
• structured-data
• json-ld
• robots.txt
• sitemap
• canonical
• lighthouse
• pagespeed
• core-web-vitals
• lcp
• render-blocking
• critical-css
• font-loading
• LCP优化
• 性能优化
• open-graph
• schema.org
• technical-seo
• generative-engine-optimization
• ai-crawler
• indexnow
• hard-gates
• SEO体检
• SEO优化
• 网站SEO
• AI可见性
• 结构化数据
• 站点审计
• 上线前检查
• 搜索引擎优化
• GEO优化
• 爬虫策略 metadata: author: zeze source: a production site's build-time audit gates + worker SEO layer openclaw: homepage: "https://github.com/Cosmofang/seo-audit" author: "zeze" runtime: node: ">=18" permissions:
  • "Reads files under the build-output directory you point it at (audit-seo.mjs)"
  • "Makes outbound HTTPS requests to the live URL you provide (audit-live.mjs)"
  • "Writes nothing — reports are printed to stdout / saved only where you redirect them"

You audit a website's technical SEO and GEO (Generative Engine Optimization — being found, cited, and recommended by AI assistants), report concrete problems ranked by impact, and apply fixes. This is an actionable harness, not just advice: two zero-dependency Node scripts do the measuring, the reference files tell you exactly what to fix and why.

The rules here are distilled from a production site whose build fails if any gate is violated — that discipline is why it scores high. Treat the gates as hard constraints, not suggestions.

When to use

• "Check/improve my site's SEO", "raise my PageSpeed/Lighthouse SEO score", "review before launch"
• "Make my site visible to AI / ChatGPT / Claude / Perplexity", "add llms.txt", "fix robots for AI crawlers"
• "Add structured data / JSON-LD / schema", "set up sitemap / canonical / Open Graph"
• "My LCP is slow / fix Core Web Vitals / PageSpeed says my site takes 7 s to load"

What you have

• scripts/audit-seo.mjs — on-disk auditor. Runs the 12 hard-gate checks over a directory of built HTML + its local CSS/JS/images. Framework-agnostic. Node ≥18, no install.
• scripts/audit-live.mjs — live-URL auditor. Checks the things only visible on a deployed origin: robots.txt policy (incl. per-AI-bot allow/deny), sitemap.xml, llms.txt, homepage JSON-LD (@graph-aware), canonical, HSTS / Vary / Cache-Control.
• references/hard-gates.md — the 12 gates: exact thresholds, the general rule, and the Astro+Cloudflare reference implementation.
• references/structured-data.md — copy-paste JSON-LD recipes (Organization, WebSite, Breadcrumb, Article, Product, FAQ) using the nested @graph pattern.
• references/geo-ai-visibility.md — the GEO layer: robots AI-crawler allowlist (exact user-agents), llms.txt format, AI-oriented schema, IndexNow.
• references/lcp-playbook.md — Core Web Vitals deep-dive: the measured levers that took the reference site from 7.5 s → ~1.5 s mobile LCP (render-blocking CSS, critical-CSS split, font discipline, deferring non-LCP DOM, third-party JS, CLS guardrails, CI lock-in). Use when the problem is speed, not markup.

Workflow

1. Locate the build output (don't audit source — audit the shipped HTML)

SEO lives in the rendered HTML. Find the build dir: Astro dist/, Next out/ (or .next after next export), Hugo public/, Jekyll _site/, Vite dist/, or a plain folder. If it doesn't exist yet, run the project's build first. Confirm with the user if ambiguous.

2. Run the on-disk audit

node scripts/audit-seo.mjs --dir <build-dir>
# options: --strict (warns→errors, CI mode) · --json · --max-page-kb 500 · --max-img-kb 500

Read the output: ✗ = ERROR (genuinely hurts ranking / breaks crawlers / Core Web Vitals — fix these first), ⚠ = WARN (best-practice miss). The heuristic score is a rough dial, not a Lighthouse number.

3. Run the live audit (if deployed)

node scripts/audit-live.mjs https://www.example.com

This is where GEO shows up: which AI crawlers are allowed/blocked, whether llms.txt exists, what JSON-LD @types the homepage actually ships.

4. Report and fix

• Group findings by severity; fix ERRORs first, then high-value WARNs.
• For each fix, open references/ for the exact target value and the reference implementation, then edit the source (templates/layout/config) — not the built HTML (it's regenerated).
• Re-run the audit to confirm green. For CI, wire audit-seo.mjs --strict into the build so regressions fail the pipeline.

The 12 hard gates (cheat-sheet — full detail in references/hard-gates.md)

1. Exactly one <h1> per page.
2. Viewport meta width=device-width, initial-scale=1.
3. Semantic landmarks — <main> + <nav> + <footer> present.
4. <title> present (≈10–60 chars) + meta description present (≈50–160 chars). Length is a soft warn — longer is a valid deliberate GEO choice.
5. Canonical — absolute-URL <link rel="canonical">, host matches the deploy origin (build-time, never runtime).
6. Open Graph — og:title + og:image.
7. Images — every <img> has width+height+alt; non-hero loading="lazy"; hero fetchpriority="high"; each file ≤500 KB (WebP/AVIF).
8. No inline executable <script> (allow only application/ld+json/json/importmap) and no on*= handlers → strict CSP script-src 'self'.
9. No external resource refs (fonts/img/css/js) — self-host for CSP + speed.
10. Page weight — HTML + same-page CSS + JS ≤500 KB (images budgeted separately).
11. Structured data — JSON-LD present; site-wide Organization + WebSite.
12. URL hygiene — lowercase, trailing slash, ≤3 path depth, one route source of truth.

GEO layer (references/geo-ai-visibility.md): robots.txt explicitly allows the major AI crawlers, ship llms.txt, enrich Organization schema with knowsAbout, ping IndexNow on deploy.

If the complaint is LCP / PageSpeed performance (gates pass but the site is slow): open references/lcp-playbook.md. Diagnose first — find the actual LCP element (often text, not an image) and measure with DevTools applied throttling, not Lantern/simulated. Then work the levers in impact order: render-blocking CSS → critical-CSS split → font preload discipline → defer non-LCP viewport DOM → eager hero image → third-party JS on idle/interaction → IntersectionObserver-deferred init. Keep CLS at 0 by reserving space for everything you defer, and lock wins in with a Lighthouse CI gate + per-page JS byte budget.

Notes

• 404/50x pages are exempt from canonical/description/OG/JSON-LD (they're noindex by design) — the auditor already skips them.
• The auditor uses conservative regex extraction, not a full DOM — it's for audit signals. A clean run is strong evidence, not a formal guarantee.
• Don't relax a threshold to make the audit pass. Fix the page.

Purpose & Capability

seo-audit is an actionable SEO + GEO auditing harness. Two zero-dependency Node scripts do the measuring; three reference files tell you exactly what to fix and why. It turns the build-time SEO discipline of a production site into a portable, framework-agnostic gate you can run on any site.

Does NOT:

• Audit source files — it audits rendered/built HTML (run your build first)
• Modify your site — it reports; you (or the agent) apply fixes to source templates
• Replace Lighthouse — its score is a heuristic dial for audit signals, not an official number
• Send your URL or content anywhere except the live origin you explicitly pass to audit-live.mjs

Instruction Scope

In scope (will handle):

• "Check / improve my site's SEO", "raise my Lighthouse/PageSpeed SEO score", "review before launch"
• "Make my site visible to AI / ChatGPT / Claude / Perplexity", "add llms.txt", "fix robots for AI crawlers"
• "Add structured data / JSON-LD / schema", "set up sitemap / canonical / Open Graph"
• Running either auditor and reporting findings ranked by severity, then applying fixes via references/

Out of scope (won't handle):

• Off-page SEO, backlinks, keyword-ranking tracking, or paid-search work
• Content writing / copywriting beyond meta title & description guidance
• Auditing a directory that hasn't been built yet (build first, then point the auditor at the output)

Behavior on missing input:

• audit-seo.mjs with no --dir defaults to dist; if the directory has no HTML it reports zero pages (no crash)
• audit-live.mjs with no URL prints usage and exits with code 2

Credentials

No credentials required. This skill uses no API keys, tokens, or accounts.

No hardcoded secrets exist anywhere in the scripts.

Persistence & Privilege

Writes: nothing by default. Both scripts print reports to stdout; --json still prints to stdout. Output is persisted only where you redirect it (e.g. > report.json).

Does NOT write:

• No files inside your project, the skill directory, or your home directory
• No shell-config or credential files
• No cron jobs or background processes

Privilege: runs as the current user, no sudo or elevated permission. Requires only Node ≥18 (global fetch).

Uninstall: delete the skill directory — there is no other state to clean up.

Install Mechanism

Standard install (clawHub)

clawhub install seo-audit

Manual install

cp -r /path/to/seo-audit ~/.claude/skills/seo-audit/

Verify install

node scripts/audit-seo.mjs --help 2>/dev/null || node scripts/audit-seo.mjs --dir . --json | head
node scripts/audit-live.mjs            # should print usage and exit 2

Both scripts are zero-dependency (Node ≥18). No npm install step is needed.

Version: 1.1.0 · Created: 2026-06-09 · Updated: 2026-06-10 · Changes: see CHANGELOG.md