geoseo-mcp

An open-source MCP server for traditional SEO and Generative Engine Optimization (GEO): Google Search Console, Bing, IndexNow, Google AI Overviews, and citation tracking across ChatGPT, Claude, Gemini, Perplexity, Grok and Brave, all in one server.

Try it in 30 seconds (no API keys)

# Audit any page: title, meta, headings, schema, freshness, GEO quotability, 0-100 score
uvx geoseo-mcp audit https://example.com

# Check a site's llms.txt against the spec
uvx geoseo-mcp llms-txt https://llmstxt.org

# Same audit after running JavaScript (for SPAs); needs the [render] extra
uvx --from 'geoseo-mcp[render]' geoseo-mcp audit https://your-spa.com --render

Sample output:

$ uvx geoseo-mcp audit https://example.com
https://example.com
  score: 15/100
  title: 'Example Domain'
  words: 27

  8 issue(s):
   ✗ meta_description: Meta description should be 70-160 chars  [length=0]
   ✗ h1: Exactly one H1 expected  [found=0]
   ✗ word_count: Pages under 300 words rarely rank or get cited  [words=27]
   ✗ schema: No JSON-LD schema found; LLMs and Google use this heavily
   ✗ canonical: Missing canonical URL
   ✗ open_graph: Missing core Open Graph tags (og:title, og:description)
   ✗ freshness: No published-date signal found; LLMs deprioritize stale content
   ✗ quotability: Few short, fact-dense sentences; LLMs prefer quotable lines  [quotable_sentences=0]

  3 check(s) passed. Use --json for the full report.

Other keyless commands: audit-site ./content, generate-llms-txt ./content https://yourdomain.com, status. Run uvx geoseo-mcp --help for everything.

Poke at all 47 tools in the MCP Inspector:

npx @modelcontextprotocol/inspector uvx geoseo-mcp

Running geoseo-mcp with no arguments starts the MCP server over stdio, which is what MCP clients expect.

Add it to your MCP client

Cursor

Or add it to ~/.cursor/mcp.json by hand (see examples/cursor-mcp.json).

Claude Code

claude mcp add geoseo -- uvx geoseo-mcp

# with keys
claude mcp add geoseo \
  -e GEOSEO_OPENAI_API_KEY=sk-... \
  -e GEOSEO_PERPLEXITY_API_KEY=pplx-... \
  -- uvx geoseo-mcp

Claude Desktop

Add to ~/Library/Application Support/Claude/claude_desktop_config.json (macOS) or %APPDATA%\Claude\claude_desktop_config.json (Windows):

{
  "mcpServers": {
    "geoseo": {
      "command": "uvx",
      "args": ["geoseo-mcp"],
      "env": {
        "GEOSEO_OPENAI_API_KEY": "sk-...",
        "GEOSEO_ANTHROPIC_API_KEY": "sk-ant-...",
        "GEOSEO_GEMINI_API_KEY": "AIza...",
        "GEOSEO_XAI_API_KEY": "xai-...",
        "GEOSEO_BRAVE_API_KEY": "...",
        "GEOSEO_PERPLEXITY_API_KEY": "pplx-...",
        "GEOSEO_SERPAPI_API_KEY": "...",
        "GEOSEO_GOOGLE_CLIENT_SECRET": "/absolute/path/to/client_secret.json",
        "GEOSEO_BING_WEBMASTER_API_KEY": "...",
        "GEOSEO_YANDEX_OAUTH_TOKEN": "y0_...",
        "GEOSEO_INDEXNOW_KEY": "..."
      }
    }
  }
}

Every key is optional; drop the ones you don't have. Other clients (Continue, Cline, Windsurf, VS Code) take the same command / args / env block. See examples/.

Then just ask:

"Which of ChatGPT, Claude, Gemini, Perplexity and Grok cite acme.com for 'best CRM for startups'? Run 3 samples each." "Audit my ./content folder and show the 5 worst pages." "Build the internal link graph and find orphan pages."

What it covers

MIT licensed and local-first: your credentials and data stay on your machine.

How citation checks work

multi_llm_citation_check(questions, target_domain) asks every configured engine each question several times (samples, default 3), because LLM answers vary from run to run. For each engine you get:

• citation_rate with a 95% Wilson interval, plus a per-question cited_rate and a stable flag
• mention_rate and unlinked_mention_rate: answers that name your brand without linking to you
• top_competing_domains: who gets cited instead of you
• subdomain-aware matching (docs.acme.com counts for acme.com)

Brand aliases are derived from the domain (acme.com → acme) and you can pass your own via brand_aliases. Requests run in parallel up to GEOSEO_MAX_CONCURRENCY (default 8).

Rough costs

Approximate list prices as of October 2026; check each provider's pricing page before running large batches.

A 10-question check with 3 samples across four engines is 120 calls, which is usually well under $2. Use samples=1 for quick spot checks.

Get credentials

• OpenAI: API key. Uses the Responses API with the web_search tool. Override the model with GEOSEO_OPENAI_MODEL.
• Anthropic: API key. Uses Claude with the server-side web search tool.
• Gemini: API key from AI Studio. Uses Google Search grounding (the same index AI Overviews use).
• Perplexity: API key. Uses the Agent API with the fast preset; set GEOSEO_PERPLEXITY_EFFORT or GEOSEO_PERPLEXITY_MODEL to change it.
• xAI Grok: API key from the xAI console, set as GEOSEO_XAI_API_KEY. Uses the Responses API with the web_search tool; override the model (default grok-4.3) with GEOSEO_GROK_MODEL.
• Brave Answers: API key with the Answers plan, set as GEOSEO_BRAVE_API_KEY. Answers come from Brave's own index.
• SerpAPI: API key. Free plan: 250 searches/month. Used for AI Overviews tracking.
• Google Search Console: see docs/setup-gsc.md. One-time OAuth flow.
• Bing Webmaster: API key from Bing Webmaster Tools → Settings → API access.
• Yandex Webmaster: OAuth token. Register an app with both webmaster:hostinfo and webmaster:verify (hostinfo alone only lists sites), then get a token via https://oauth.yandex.com/authorize?response_type=token&client_id=<id> and set GEOSEO_YANDEX_OAUTH_TOKEN. Docs.
• IndexNow: any random 32-char hex string hosted at https://yourdomain.com/<key>.txt. Spec.

Tools that need a key you haven't set return a clear error; everything else keeps working.

Configuration

Example: only the keyless audit tools, for a lean agent context:

"env": { "GEOSEO_TOOLSETS": "audit,graph" }

Streamable HTTP

uvx geoseo-mcp serve --transport http --port 8000
# MCP endpoint: http://127.0.0.1:8000/mcp

It binds to localhost by default. There is no built-in auth, so put it behind a proxy you trust before exposing it anywhere.

JavaScript rendering

audit_page reads raw HTML by default. If a page looks like a JS shell (an empty SPA root or almost no text), it adds a js_rendering finding. Install the extra and pass render=True (or --render on the CLI) to audit the rendered DOM:

pip install 'geoseo-mcp[render]' && playwright install chromium

Tools (47)

• Status: geoseo_status, list_llm_engines
• Google Search Console: gsc_list_sites, gsc_performance, gsc_inspect_url, gsc_submit_sitemap
• Bing Webmaster: bing_list_sites, bing_query_stats, bing_page_stats, bing_url_info, bing_submit_url, bing_crawl_issues
• Yandex Webmaster: yandex_list_sites, yandex_site_summary, yandex_query_stats, yandex_url_info, yandex_submit_url, yandex_crawl_issues
• Indexing: indexnow_submit_url, indexnow_submit_urls
• LLM citations: perplexity_query, perplexity_citation_check, openai_query, claude_query, gemini_query, grok_query, brave_query, multi_llm_query, multi_llm_citation_check
• Google AI Overviews: aio_check, aio_citation_check
• On-page: audit_page, audit_site, find_broken_external_links, lint_schema_page, lint_schema_site
• Link graph: internal_link_graph, suggest_internal_links
• llms.txt: generate_llms_txt, validate_llms_txt
• Trends: trend_init, trend_stats, snapshot_gsc, snapshot_llm_citations, snapshot_serp_aio, trend_gsc, trend_llm_citations

suggest_internal_links uses TF-IDF by default (offline, free). Pass mode="openai", "gemini" or "local" (sentence-transformers) for embedding similarity; vectors are cached in SQLite by content hash.

For Yandex tools, pass a normal URL as site; it's matched to the Yandex host id (https:example.com:443), following the main mirror. yandex_url_info only covers pages in "Monitoring important pages" (Yandex has no arbitrary-URL inspect endpoint).

Snapshots are append-only and stored at $GEOSEO_DB. Schedule the snapshot_* tools weekly to build a private time series of your AI-search visibility.

Roadmap

Tracked in issues. Shipped in 0.4: broken-link prospector, schema linter, Grok and Brave citation engines, embedding-based link suggestions, Yandex Webmaster.

Contributing

PRs welcome. Each engine is a module in src/geoseo_mcp/engines/ that exposes query() and returns the standard {engine, question, model, answer, citations, cited_domains, usage} shape; tools live in src/geoseo_mcp/tools/. See docs/architecture.md.

git clone https://github.com/Rachit8484/geoseo-mcp.git && cd geoseo-mcp
uv sync --extra dev
uv run ruff check src tests && uv run pytest

License

MIT. See LICENSE.