hainrixz/claude-seo-ai

claude-seo-ai

The SEO + AI-search optimization toolkit for Claude Code. Audit any website on two independent axes — classic Search SEO and AI Visibility (GEO/AEO) — and optionally apply the safe fixes for you.

Two scores, never blended. A page can rank well in Google yet be uncitable by ChatGPT, Perplexity, Google AI Overviews, Gemini, and Claude — or the reverse. claude-seo-ai measures both and tells you exactly what to fix.

🇪🇸 Resumen en español al final · guías completas en docs/es/.

---

Why

AI answer engines now sit beside classic search as a primary discovery channel, and the rules for being cited differ from the rules for ranking. Most SEO tools treat AI search as a footnote and never write fixes back into your code. This one is built for 2026–2027: honest, agent-native, and able to both diagnose and (opt-in) fix.

It is original work — inspired by the patterns of community SEO tooling but copies no branding, text, or names from any other project. MIT-licensed.

Install

As a Claude Code plugin (recommended):

/plugin marketplace add Hainrixz/claude-seo-ai
/plugin install claude-seo-ai@claude-seo-ai
/reload-plugins

Cross-agent (Cursor, Codex, Gemini CLI, Windsurf…) via Vercel Skills:

npx skills add Hainrixz/claude-seo-ai

The plugin works fully offline (Tier 0, no keys). See Data tiers for optional rendering and API integrations. The skills-only channel ships the Markdown but not the scripts/ acquisition layer — see docs/en/distribution.md.

Usage

/claude-seo-ai:audit    [--pages N] [--max N] [--render static|auto|js]
                                  [--ua default|googlebot|bingbot|gptbot|oai-searchbot|claude-searchbot]
                                  [--vertical ecommerce,docs] [--environment production|preview|staging|local]
                                  [--feed ] [--out ]

/claude-seo-ai:geo      [--pages N] [--render static|auto|js] [--feed ]
                                  [--probe ""] [--gsc-ai-export ]

/claude-seo-ai:score   [findings.json | run-dir | latest[:host]]

/claude-seo-ai:compare   […] | --baseline [latest] --against 
                     | --staging  --prod  |  --gap ""

/claude-seo-ai:fix      [--target auto|local|shopify|wordpress|webflow|wix|ghost|hubspot|bigcommerce|instructions]
                                  [--category M5 --category auto] [--include-proposed]
                                  [--project ] [--dev-url ] [--run ] [--lang en|es]
                                  [--dry-run] [--publish] [--rollback ] [--force]

audit, geo, score and compare are read-only and never touch your files. fix previews everything by default and writes only after you confirm each change. --render defaults to static, so a run never launches a browser unless you ask. Power users can call any module directly, e.g. /claude-seo-ai:seo-schema-jsonld.

Every audit is persisted under ${CLAUDE_PLUGIN_DATA}/runs/// — which is why score, compare and fix work from a report on disk instead of “the last audit this session”. Full flags and the run layout: docs/en/usage.md.

Works on any site

The audit is the same everywhere. What changes per platform is what the platform already generates for you, what it refuses to let anyone change, and whether a fix can be written back at all — so a four-layer detector (platform · framework · CMS plugins · hosting) writes a profile.json for every run, and the fixer routes each finding to the adapter that owns that surface.

Status is the header of each platform’s knowledge card: stable = exercised end to end with a rollback · beta = implemented but thinly exercised, changes stay PROPOSED · instructions-only = no write API exists, so every fix is a click path for a human.

Anything the detector cannot name still gets the full audit — it simply routes every fix to instructions, which renders the exact snippet and the panel or file it belongs in, in EN or ES. That is a real outcome, not a failure.

Credential setup, per-platform quick starts, what never goes live by default, and the full UNVERIFIED list: docs/en/platforms.md.

Two scores, never blended

Every audit reports two 0–100 scores with letter bands (A–F) and a one-line interpretation (details):

• Search SEO — weighted toward indexability, Core Web Vitals, on-page, schema.
• AI Visibility (GEO/AEO) — weighted toward answer extractability, schema, fact density, AI-crawler access, entities.

Severity gating caps an axis at 40 when a severity: 5, established, failing finding lands in an active category (e.g. site-wide noindex) — a directional failure never caps. Conditional verticals re-normalize so a blog isn’t penalized for lacking Product schema. needs_api and manual_review checks are excluded from the math and counted separately as score confidence. An axis with no scored finding reports unscored, never F.

How it works

A skill-first, three-layer design — the scripts acquire and prove, the model judges:

1. Directive — one of the five command skills (audit/geo/score/compare/fix).
2. Orchestration — seo-orchestrator runs the deterministic pipeline in one command (crawl → platform profile → 25 checks → first report), then dispatches four read-only specialist subagents in parallel with an absolute-path envelope, and merges everything into a persisted, schema-validated report.
3. Execution — 23 focused seo-* module skills, preloaded into their agent, each emitting findings that conform to schema/finding.schema.json — with observed evidence and a runnable verification.reproduce command.

See docs/en/architecture.md.

What it audits

A complete 2026 suite, grouped:

The deterministic half of that is a registry of 25 checks emitting 151 finding ids, which run with no model in the loop and never throw: a check that fails is isolated per page and its error is recorded in checks.json.

Get cited by AI engines

The GEO/AEO modules score how citable your content is — and they lead with the one thing a vendor actually documents.

• Google’s eligibility gate is Search eligibility. Per Google’s “AI features and your website” documentation, a page can appear in AI Overviews / AI Mode only if it is indexed and eligible to be shown with a snippet. That is the whole gate. noindex, nosnippet, max-snippet:0 and data-nosnippet on the primary content remove you; Google-Extended does not — it is a training/grounding token that leaves Search-based AI features untouched. These M14.ai_eligibility.* findings are the only established ones on the AI axis.
• AI-crawler access. Whether retrieval bots (OAI-SearchBot, Claude-SearchBot, PerplexityBot, Amazonbot, DuckAssistBot…) can reach and render your page, with a per-vendor source for every user-agent row. See references/ai-crawlers.md.
• Content-Signal in robots.txt (search / ai-input / ai-train) is parsed in both its global and per-group placements. It is a preference, not an enforcement: the IETF draft expired in 2026-04 and no AI vendor documents compliance, so those findings are directional at most — and a Content-Signal line is never presented as a substitute for a Disallow.
• Agentic commerce readiness for stores (M18): the UCP merchant profile at /.well-known/ucp, an Agentic Commerce Protocol product feed linted against the nine required fields, the "79.99 USD" price form, GTIN check digits and duplicate item ids, plus feed-vs-page price and availability consistency.
• Agent readiness (M22): can a scripted agent understand and operate this page from the HTML alone — named controls, labeled form fields, real links instead of href="javascript:". Directional by construction, capped at severity 3, and what static HTML cannot answer is listed verbatim rather than skipped.
• --probe "" runs a WebSearch and reports presence rate, best rank and the competitor tally. It is severity 0 and never scored, and it carries the script’s own disclaimer verbatim: this is web-search presence, not AI Overview / AI Mode / ChatGPT / Perplexity citation data — no vendor exposes that. Indexing is a documented precondition; presence here is necessary, not sufficient.
• --gsc-ai-export imports the Search Console Generative AI report you exported by hand (there is no API for it). Impressions only — no clicks, no citations, no per-answer attribution — and also severity 0.

The opt-in fixer

The fixer (skills/fix) is disable-model-invocation: true — Claude can never trigger writes on its own. Only /claude-seo-ai:fix does, and only the one writer subagent (seo-fixer-writer) has Write/Edit; every auditor is read-only by tool allowlist.

• AUTO (deterministic, additive, verifiable): meta viewport/charset/lang, Tier-1 JSON-LD, robots.txt AI directives, canonical, hreflang sets, OG/Twitter cards, image dimensions, XML sitemaps, llms.txt (disclosure-gated).
• PROPOSED (per-item accept): generated ``/meta description, answer-block/TL;DR rewrites, internal-link insertions, heading restructuring, generated image alt text.
• ADVISORY (never written): content rewrites, Core Web Vitals/performance, rendering strategy, redirects, link-building, Merchant Center/GBP backend.

Dry-run by default · previews everything (unified diff for files, the request + variables + before values for APIs, the exact command for CLI ops) · every change badged [none] / [staged] / [LIVE] · one-shot confirmation tickets that expire in 15 minutes · backups under ${CLAUDE_PLUGIN_DATA} · idempotent · re-verifies each change (a stale cache is pending_cache, never a pass) · rollback from the captured before state · never writes to .git/secrets/lockfiles (enforced by two PreToolUse hooks, one for the file tools and one for Bash).

One guarantee is writer protocol rather than code, and this repo does not blur that line: being git-aware — git status --porcelain first, refuse a dirty working tree unless --force, prefer a seo-fix/ branch — is instructed in skills/seo-fix-apply and skills/fix, and no adapter runs git. Everything else in the list above is enforced by code: the adapters, the ticket TTL and the two PreToolUse hooks.

Nothing goes live by default: Shopify pushes to an unpublished theme and a live push from the shell is a hard deny; Webflow and HubSpot stay staged until a separately-confirmed publish; Ghost never flips a draft; WordPress never touches post status or content. Where a platform has no staging surface, you are told before you confirm, not after.

And nothing is dropped quietly: the plan accounts for every finding in the report — considered, planned, held back for want of --include-proposed (with the ids), advisory, and the ones no adapter owns.

Honesty guardrails

This tool refuses to ship SEO myths:

• llms.txt and the other AI discovery files (M21) are scored 0 — reported, never counted. Google Search ignores llms.txt; adoption is ~10 % and no major vendor documents honoring it for retrieval (it is still useful as IDE-agent context).
• Only what a search engine documents may be established — on the AI axis, that is M14 eligibility and nothing else. And only an established failure can cap a score: a severity-5 directional finding never does.
• Probes are never scored. --probe and --gsc-ai-export are severity 0, excluded from the score and from every cap, and reported beside it rather than inside it.
• A thin run says so. When less than 50 % of an axis’s always-on weight carried a scored finding, the band is marked provisional and the unmeasured categories are named. A one-page run cannot look like an A.
• Unknown-language pages are manual_review, not a pass. The heuristics ship EN and ES lexicons; on a page in any other language the language-dependent checks are skipped and say so, instead of scoring an unreadable page as clean.
• FAQPage/HowTo flagged as deprecated-for-Google-rich-results (still valid for AI extraction) — never counted as a rich-result win.
• Keyword density is not optimized for; stuffing is flagged as a negative.
• No “AI-specific keyword” magic — citation comes from extractable structure, verifiable facts, and authority.
• Lab vs field Core Web Vitals are clearly distinguished (only CrUX field p75 drives the score).
• needs_api / manual_review are never a silent pass and never override a check that did decide.
• No fabrication, ever — never invents statistics, citations, dates (no backdating), credentials, or sameAs identity links; and no naked percentages.
• Every finding carries a confidence tier (established / directional / speculative); leaked/inferred guidance (e.g. NavBoost) ships only as directional.

Data tiers

Tier 0 is not a degraded mode — it is the acquisition layer. WebFetch is deliberately not used to acquire pages: inside Claude Code it returns a small-model summary, never raw HTML, headers or status codes. Nothing is ever installed on your behalf, and the tier is measured from what actually happened, not declared.

See .mcp.json.example for opt-in render MCPs (plus @shopify/dev-mcp, a docs oracle that never writes) and docs/en/mcp.md.

Acquisition layer, CI action & headless runner

The skills are Markdown, but the scripts/ are what make the audit real: zero-dependency Node ESM (Node ≥ 18, no install step), each with main(args) → {result, code}, uniform exit codes (0 ok · 1 usage · 2 runtime · 3 gate tripped), and no --help — run one with no arguments and it prints its usage line.

# Acquire one page (status chain, headers, robots directives, parsed DOM) into a run
node scripts/snapshot.mjs https://example.com --out ./runs --render auto --json

# Crawl: robots-aware discovery, URL-template detection, sampled by template
node scripts/crawl.mjs https://example.com --out ./runs --pages 12 --per-template 2 --depth 3

# The whole deterministic pipeline: acquire → profile → 25 checks → report.json + report.md
node scripts/audit.mjs https://example.com --out ./runs --pages 5 --format json

# Re-score a persisted run, and diff two of them
node scripts/score.mjs   --run ./runs/example.com/
node scripts/compare.mjs --baseline latest --against ./runs/example.com/ --format md

# Gates and tests
node scripts/check.mjs        # syntax + JSON + skill frontmatter + version alignment
node tests/run.mjs            # 1000+ assertions over synthetic fixtures, no network

Run it headless in CI with the bundled composite GitHub Action — same deterministic subset, a job summary, the scores as step outputs, and exit 3 when a gate trips:

- id: seo
  uses: Hainrixz/[email protected]
  with:
    url: https://example.com
    pages: '5'
    render: static
    fail-under-search: '70'
    fail-under-ai: '60'

A ready-to-copy workflow is in .github/workflows/seo-audit-example.yml.

Project structure

.claude-plugin/   plugin.json (+ userConfig credential prompts) + marketplace.json
skills/           34 skills — 5 commands (audit, geo, score, compare, fix)
                  + 23 seo-* audit modules (M1–M22)
                  + orchestrator, vertical-detect, crawl-render, seo-score, platform-detect, fix-apply
agents/           5 subagents (4 read-only auditors + 1 writer), each preloading its module skills
hooks/            two PreToolUse guards: guard-write (file tools) + guard-bash (shell)
scripts/          zero-dep Node acquisition layer + runner
  ├─ lib/         html, fetch, robots, sitemaps, store, renderers, route-map, adapter, credentials…
  ├─ checks/      25 deterministic checks emitting 151 finding ids
  └─ adapters/    local-files, shopify-theme/admin, wordpress-rest/wpcli, page-api (+providers), instructions
references/       scoring model, AI crawlers, schema, CWV, routing + 24 platform knowledge cards
schema/           finding + report JSON Schemas, JSON-LD templates
action.yml        composite GitHub Action for CI
docs/en, docs/es  bilingual guides (usage, architecture, scoring, mcp, distribution, platforms)
assets/           pixel-art images
tests/            node:test suites (unit + e2e) over synthetic fixtures

About

Built by Enrique Rocha — I help teams ship AI: consulting, automations, and agents. This is a community, MIT-licensed project: use it, fork it, open issues and PRs (see CONTRIBUTING.md).

• 🌐 tododeia.com
• 📸 Instagram @soyenriquerocha

License

MIT · Claude mascot artwork generated for this project in pixel-art style.

---

🇪🇸 Resumen (Español)

claude-seo-ai es la herramienta open-source de SEO + búsqueda con IA para Claude Code. Audita cualquier sitio en dos puntajes independientes — Search SEO clásico y Visibilidad en IA (GEO/AEO) — con hallazgos reproducibles, y opcionalmente aplica las correcciones seguras por ti (meta tags, JSON-LD, robots.txt para crawlers de IA, hreflang, sitemaps…), siempre con confirmación previa.

• Instalar: /plugin marketplace add Hainrixz/claude-seo-ai → /plugin install claude-seo-ai@claude-seo-ai. Multiagente: npx skills add Hainrixz/claude-seo-ai (solo las skills en Markdown; la capa scripts/ va en el canal de plugin).
• Usar: /claude-seo-ai:audit · :geo · :score · :compare · :fix. Los cuatro primeros son de solo lectura; fix previsualiza todo y escribe solo tras tu confirmación cambio por cambio.
• Cada auditoría se persiste en ${CLAUDE_PLUGIN_DATA}/runs///, así que score, compare y fix trabajan sobre un informe en disco, no sobre «la última auditoría de esta sesión».
• Funciona en cualquier sitio: un detector de cuatro capas (plataforma · framework · plugins · hosting) enruta cada corrección al adaptador dueño de esa superficie — Shopify, WordPress, Webflow, Wix, Ghost, HubSpot, BigCommerce, árboles de código o instructions. Nada se publica en vivo por defecto.
• Honestidad: llms.txt y el resto de M21 se informan pero puntúan 0; las sondas (--probe, --gsc-ai-export) tienen severidad 0 y nunca entran en la puntuación; una ejecución delgada marca su banda como provisional; needs_api / manual_review nunca son un aprobado silencioso; nunca inventa estadísticas, fechas ni enlaces de identidad.
• Sin claves funciona (Tier 0: los scripts de Node sin dependencias son la capa de adquisición). Render de SPAs y Core Web Vitals reales son opcionales (Tier 1+).

Guías completas en español: docs/es/ — uso · arquitectura · puntuación · plataformas · MCP · distribución. Hecho con cariño por tododeia.com.