SEO & GEO

seo-geo-mcp-server

An MCP server that lets an AI agent audit a page for SEO and GEO (Generative Engine Optimization) — on-page tags, structured data, robots.txt, sitemaps, hreflang, and whether ChatGPT, Claude, Perplexity and Gemini can actually crawl and cite you. No API keys required.

[ci] [npm] [MCP] [TypeScript] [License: MIT]

Built on the v2 MCP SDK: the server speaks the 2026-07-28 protocol revision and still accepts 2025-era clients (Claude Desktop, Claude Code, Cursor) from the same factory — one build, both eras, nothing to configure.

Ask Claude "how is this page doing, and will AI assistants cite it?" and it runs a full audit and hands you a graded report with prioritised fixes — instead of you pasting a URL into six different web tools.

> Audit https://example.com/guide and tell me if ChatGPT can cite it
  seo_audit(url="https://example.com/guide", include_geo=true)
  Overall: A (92/100) · indexable: yes    Meta tags & social preview   95 (A)    Heading structure           100 (A)    Structured data              75 (C)
  GEO readiness: B (85/100)    ✅ AI crawler access        25/25    ✅ Server-rendered content  20/20    ❌ Authorship & entity       2/10  1. Add author and Organization markup with `sameAs` links to official profiles.

Why this exists

Two gaps, one server.

The SEO gap: the on-page checkers are all web UIs. None of them let an agent run the audit, read the result and fix the code in the same loop.

The GEO gap: "generative engine optimization" tooling is mostly rank-tracking dashboards behind a subscription. The mechanics that actually decide whether an AI assistant can cite you are cheap to check and almost never checked:

• Can the AI search crawlers reach you? Blocking GPTBot stops training. Blocking OAI-SearchBot stops you being cited. Most sites that meant to do the first have accidentally done the second. This server separates them.
• Does your content exist without JavaScript? Googlebot renders JS. GPTBot, ClaudeBot, PerplexityBot and CCBot largely do not. A client-rendered page can rank fine in Google and be invisible to every AI assistant.

It is the agent-facing companion to the tools at ortamarco.me, and shares its core (the connect-time SSRF guard and host validation) with domain-security-mcp-server.

Tools

Audits

GEO

On-page

Technical

Every tool is read-only, declares an outputSchema and returns structuredContent (validated by the SDK) alongside human-readable Markdown (default) or JSON (response_format="json"), plus actionable error messages.

Honesty notes

This server deliberately refuses to overstate two things that most GEO content gets wrong. Both are surfaced in tool output, not buried here:

• llms.txt is not an adopted standard. It is a community proposal from September 2024. No major AI vendor has documented that its crawlers read it from third-party sites, and Google has publicly said it does not. The tool reports presence and validates shape — and geo_audit deliberately does not score it. (llms-full.txt is a docs-tooling convention, not part of the proposal.)
• Some robots.txt blocks are advisory. Perplexity-User, ChatGPT-User and meta-externalfetcher are documented by their own vendors as ignoring or possibly ignoring robots.txt. Reporting those as cleanly "blocked" would be misleading, so they are listed separately as unenforceable.

Crawler tokens carry a provenance field distinguishing first-party vendor documentation from community aggregators, and vendors that publish no token at all (xAI/Grok, Microsoft Copilot) are named explicitly — because a missing rule cannot be read as either allowed or blocked.

Install

Requires Node.js 20.18+. Nothing to clone — every MCP client can run it with npx.

Use it with Claude Code

claude mcp add seo-geo -- npx -y seo-geo-mcp-server

Use it with Claude Desktop or Cursor

Add to claude_desktop_config.json (or ~/.cursor/mcp.json) — see examples/:

{  "mcpServers": {    "seo-geo": {      "command": "npx",      "args": ["-y", "seo-geo-mcp-server"]    }  }}

On Windows use "command": "cmd" with "args": ["/c", "npx", "-y", "seo-geo-mcp-server"]. Restart the client, then ask: "Audit the SEO and GEO of example.com."

Self-host (HTTP transport)

The same server speaks stateless Streamable HTTP for remote or multi-client use. One endpoint serves both protocol eras; there is no session state and no Mcp-Session-Id to carry.

TRANSPORT=http npx -y seo-geo-mcp-server# POST JSON-RPC to http://127.0.0.1:3000/mcp   ·   health at /healthz

It is safe by default: it binds to 127.0.0.1 and only accepts localhost Host and Origin headers, which blocks DNS-rebinding attacks from a web page. To expose it — for example behind Coolify or Traefik — opt in explicitly:

Binding to a non-loopback address without ALLOWED_HOSTS or MCP_AUTH_TOKEN works, but the server says so on stderr. With Docker (the image sets HOST=0.0.0.0):

docker build -t seo-geo-mcp .docker run -p 3000:3000 -e ALLOWED_HOSTS=mcp.example.com -e MCP_AUTH_TOKEN=change-me seo-geo-mcp

Develop

npm run dev      # tsx watch (stdio)npm test         # deterministic unit tests: SSRF guard, robots matcher, nesting guard,                 # gzip sitemaps, SPA detection, JSON-LD, HTTP transport defaultsnpm run smoke    # call all 17 tools over MCP, in BOTH protocol eras, and validate                 # structuredContent vs outputSchemanpm run inspect  # open the MCP Inspector against the built servernpm run build    # type-check + emit dist/

evals/ holds a 10-question LLM evaluation set (stable, verifiable) and instructions for running it — see evals/README.md.

How it works

src/├── index.ts        # transport selection: serveStdio | createMcpHandler + Express├── server.ts       # the factory: registers every tool on one McpServer├── schemas.ts      # Zod 4 outputSchema for each tool├── core/           # pure logic, no MCP coupling — reusable & testable│   ├── validate.ts     # host/URL validation and the address classifier│   ├── netguard.ts     # connect-time SSRF guard (every socket's address is checked)│   ├── fetch.ts        # fetch: guarded dispatcher, manual redirects, loop detection, byte caps│   ├── page.ts         # HTML loading + the shared parsed-document model│   ├── meta.ts         # title/description/canonical/robots, Open Graph, hreflang│   ├── content.ts      # headings, readability, word counts, image SEO│   ├── structured-data.ts # JSON-LD/microdata + Google rich-result requirements│   ├── robots.ts       # RFC 9309 parser and rule matcher│   ├── ai-crawlers.ts  # the AI crawler registry (token, purpose, compliance, provenance)│   ├── sitemap.ts      # discovery, index following, gzip, protocol limits│   ├── links.ts        # link classification + broken-link sampling│   ├── redirects.ts    # chain tracing + host canonicalisation│   ├── geo.ts          # llms.txt, JS-rendering detection, GEO scoring│   └── seo-audit.ts    # the composite audits (one fetch, every analyser)└── tools/          # thin MCP wrappers (Zod schemas, descriptions, formatting)

The core/ layer is deliberately free of any MCP types, so the same logic can power both this server and a web UI.

Security

The tools fetch URLs the caller names and URLs the audited site names (links, og:image, hreflang alternates, sitemap children, redirects), so every outbound connection is screened against server-side request forgery:

• Private, loopback, link-local (cloud metadata), shared (CGNAT), multicast and reserved addresses are refused in every spelling, including IPv4 embedded in IPv6 ([::ffff:127.0.0.1]).
• The check runs at connect time, on the address the socket is about to use, so DNS rebinding and names only an internal resolver knows are refused too. Redirects are followed by hand, every hop is screened, and loops are reported.
• Untrusted input is bounded: 2 MB of HTML, 512 levels of nesting (what browsers keep), a backtracking-free robots.txt matcher, 10 MiB per sitemap (gzip output included), six link checks at a time, and at most 200 items per list in a result.

Found a problem? Please open a private security advisory.

License

MIT © Marco Orta