Rate my site

Part of CLAUDE SEO GOD MODE. Follows RULES.md at the repo root: no invented data; every finding tagged OBSERVED / INFERRED / RECOMMENDED / UNKNOWN.

Mode A: scripts available (Claude Code, Desktop with this plugin/skill, or the MCP server)

1. Normalize the URL (add https:// if missing). Rate only sites the user asked about. The crawl is polite: robots.txt is honored, about 30 pages, 0.3 s delay.
2. Run one of:
   • MCP tool seo_rate_site (site, max_pages), or
   • python <repo root>/scripts/rate_site.py <url> --out <folder>. The repo root is two folders up from this file (${CLAUDE_PLUGIN_ROOT} when installed as a plugin). Use --max-pages 50 for bigger sites.
   • Pick --out: the user's current project or working folder if there is one, as seo-rating-<host>/. Otherwise omit it (a temp folder).
3. Read the printed scorecard. Files written: report.html (visual report), scorecard.md, rating.json.
4. Reply in this shape:
   • Headline: one sentence with the overall score, grade, and the single biggest issue.
   • Score table exactly as produced (Overall, SEO, AEO, GEO, grades), plus the "Potential" line.
   • Top 5 fixes in plain language: the point gain, the pages affected, and one sentence on why it matters.
   • Fix kits: if the scorecard has any, show the 1–2 most useful ones, with the [CONFIRM] placeholders left as they are.
   • What's working: 3–5 bullets, so the owner knows what not to break.
   • Can't see from outside: one short line (rankings, traffic, index status, backlinks, AI citations).
   • Next step offer: open the visual report, draft the fixes, run the deeper audit (workflows/initial-audit.md), or connect Search Console (skills/ai-search-answer-workflow).
5. Mention report.html by its file name, or as a link if the file is inside the user's folder. Don't paste temp-folder paths into the chat. If the user wants the report and it's in a temp folder, copy it into their folder first.
6. Keep it scannable. Don't paste the full "All checks" table unless asked.

Never change the scores, invent extra numbers, or claim rankings or traffic. Verify any surprising top fix against the live page (curl or view-source) before presenting it as fact. If it's wrong, say so and note it as a rater false positive (see CONTRIBUTING.md § Self-improvement loop).

Mode B: no Python or no network for scripts (e.g. a claude.ai chat)

Rate manually with the same rubric (references/rating-rubric.md), using whatever page-fetch tool is available. Fetch at least: the homepage, /robots.txt, the sitemap, http:// plus the www/non-www variant of the homepage, a made-up URL (to test for soft 404s), and 3–5 key pages from the navigation.

• Score only the checks you could verify. Mark the others UNKNOWN and leave them out of the denominator.
• Start with: "Manual rating from N fetched pages (no crawler run); scores are approximate." Tag scores INFERRED.
• Never guess performance numbers.
• Use the same reply shape as Mode A.

Interpreting scores

• SEO (60% of overall): crawl & index, on-page, content, structured data, performance. Performance is an INFERRED estimate when PageSpeed is unavailable; set GOOGLE_API_KEY for real data.
• AEO (20%): can answer engines quote you? Snippet eligibility, question headings, direct answers, FAQ match, structure, raw HTML.
• GEO (20%): can generative AI search find, trust, and cite you? AI crawler access (robots.txt plus a live CDN/WAF probe), entity clarity, location, sameAs, About/Contact, contact details, sourced claims, Bing access.

Google says AI Overviews and AI Mode run on core Search ranking and need no special markup, and that placement isn't guaranteed. The AEO/GEO checks are readiness signals, not Google requirements.