--- title: "Audit a website from inside Claude or ChatGPT: the SeeGeo MCP server and connector" description: "SeeGeo's MCP server puts a full AI-visibility audit inside Claude, ChatGPT, Cursor or any agent: connect once, ask for a grade, get the fixes, re-check." canonical: https://see-geo.com/blog/seegeo-mcp-connector language: en published: 2026-09-23 author: "SeeGeo Team" publisher: SeeGeo alternates: fr: https://see-geo.com/fr/blog/seegeo-mcp-connector de: https://see-geo.com/de/blog/seegeo-mcp-connector --- # Audit a website from inside Claude or ChatGPT: the SeeGeo MCP server and connector > SeeGeo's MCP server puts a full AI-visibility audit inside Claude, ChatGPT, Cursor or any agent: connect once, ask for a grade, get the fixes, re-check. **Quick answer: SeeGeo now runs inside the AI tools you already use. Connect `https://see-geo.com/mcp` as a connector in claude.ai or ChatGPT (one click, sign in, Allow), or add it to Claude Code, Cursor or any MCP client with an API key, and the assistant gains five tools: `audit_site` (the full graded audit, 20–40 seconds), `get_audit`, `crawler_check` (can GPTBot, ClaudeBot and PerplexityBot read this site — a few seconds, free), `list_audits` and `usage`. The point is not a new place to see a grade. It is that the assistant that is about to edit your robots.txt, write your JSON-LD or draft your About page can measure the site first, make the change, and measure again — without you copying anything between tabs. Free plan: 10 API audits a month. The key it gets is listed on your account page; revoking it disconnects it.** ## What the assistant can do once it is connected SeeGeo's audit reads a site the way AI engines do and grades it A–F across six categories: crawler access for 16 AI bots, technical foundation, structured data, content extractability, off-site presence and entity clarity. It returns the category scores and a ranked list of findings, each with a one-line fix. The [methodology](https://see-geo.com/methodology) is public, including the engine changelog that says what changed in each version. Over MCP that becomes five tools: | Tool | What it does | How long | |---|---|---| | `audit_site` | Runs the full audit of a domain and waits for the result: grade, six category scores, findings with fixes | 20–40 s | | `get_audit` | Reads an audit by id — status while it runs, the report when done, optionally the full evidence | instant | | `crawler_check` | Reads robots.txt for every AI crawler and fetches the homepage as GPTBot, OAI-SearchBot, ClaudeBot and PerplexityBot to see whether a bot wall blocks them. Does not count as an audit | 2–5 s | | `list_audits` | The account's recent audits with grade and links | instant | | `usage` | Plan and audits used this month | instant | Every result comes back twice: as readable text for the model, and as structured JSON (`structuredContent`) for clients and agents that parse rather than read. Since yesterday's first connector session, the structured half carries the top findings too — we watched an agent get a grade with no fixes because its client only read the structured part, and fixed that the same day. ## Why in the chat, and not on the site Because the work happens in the chat now. A founder asking Claude "why doesn't ChatGPT ever mention us?" used to get a plausible essay. With the connector, the same question produces an audit of the actual site, the actual findings, and a plan that starts from what is measurably wrong. Some patterns from the first sessions: **Measure, change, measure.** "Audit example.com. Fix whatever is critical in this repo. Audit it again." The assistant runs `audit_site`, sees "AI crawlers blocked by robots.txt" or "no Organization schema", edits the files it has in front of it, and re-audits to confirm the score moved. The ten-minute audit cache means a re-run returns the earlier result unless something changed — so a real re-audit after a deploy is the honest check. **Check before you publish.** `crawler_check` takes seconds and does not count against the plan. Before a launch, before moving hosts, after adding a WAF: "can the AI crawlers still read us?" is now a question with an immediate answer, from inside the deploy conversation. **Compare.** "Audit us and our three competitors, and tell me which category we lose on." Four audits, one table, and a reason. In our [3,000-question local study](https://see-geo.com/blog/chatgpt-local-favourite), ChatGPT's local answers were 82% place cards linking the business's own site; the site that link opens is what the audit grades. **Explain a finding.** The findings are short by design. In the chat you can ask what "entity clarity 74" means for a company with three product names, and the assistant has the finding's evidence to answer from, not a guess. ## Three ways to connect **claude.ai (and the Claude desktop app)** — Settings → Connectors → *Add custom connector* → URL `https://see-geo.com/mcp`. Claude discovers the OAuth server, registers itself, and sends you to a SeeGeo page that says *Let Claude use SeeGeo as you?* with an Allow button. Sign in if you are not already (there is no password; a magic link or Google). The tools appear in every chat where the connector is enabled. **ChatGPT** — Settings → Apps & Connectors → *Advanced* → *Developer mode* → *Create*, server URL `https://see-geo.com/mcp`, authentication *OAuth*. Same consent page, same tools. **Claude Code, Cursor, Windsurf, any MCP client** — with an API key from your [account page](https://see-geo.com/account): ``` claude mcp add --transport http seegeo https://see-geo.com/mcp \ --header "Authorization: Bearer sg_…" ``` or locally from npm, no server round-trip for the transport: `npx seegeo-mcp` with `SEEGEO_API_KEY` in the environment. The package is MIT and listed in the official MCP registry as `com.see-geo/seegeo`. Details and a Python example are on [/developers](https://see-geo.com/developers). What the connector actually receives, in every case, is an ordinary SeeGeo API key. The OAuth flow creates one named after the client — "Claude (connector)", "ChatGPT (connector)" — and lists it on your account page next to any key you made by hand. Revoke it there and the connector is disconnected; there is nothing else to clean up. It cannot read other accounts, change your plan, delete anything, or see your billing. It can run audits, read your reports and check crawler access, counted against your plan like any key. ## What we changed to be worth an agent's time Agents are unforgiving users: they do not scroll, they do not retry politely, and they give up on a 401. The connector work was the last of a series: - The public API is described by an OpenAPI 3.1 spec at [/openapi.json](https://see-geo.com/openapi.json), which is also what a Custom GPT imports as Actions. - Every content page has a markdown twin (`Accept: text/markdown` or a `.md` suffix) and the whole site is in [/llms-full.txt](https://see-geo.com/llms-full.txt), so a model researching GEO gets prose, not navigation. - The MCP server's default wait is 50 seconds — inside the 60-second tool timeout most clients enforce — and if an audit runs longer it returns the id to poll rather than failing. - Unauthenticated calls to `/mcp` now answer with the OAuth discovery hint (RFC 9728) instead of a dead end. We found this by reading our own logs: 26 connector attempts in one day, all bouncing off a bearer-key wall. - Findings are in the structured result, not only the text (see above). ## Limits, honestly The audit is a crawl of the public site plus a set of checks; it does not query ChatGPT or Gemini about you. That is what [tracking](https://see-geo.com/pricing) does, and it is not in the connector yet — the next tool to add is "what did the engines answer this week for my prompts". The free plan allows 10 API audits a month and 60 requests a minute; paid plans raise the audit cap to 100, 500 or 1,000. An audit of the same domain within ten minutes is served from cache, and the six-category rubric is ours, not a standard — read the [methodology](https://see-geo.com/methodology) before quoting a score to a client. ## Frequently asked questions **What is the SeeGeo MCP server?** An implementation of the Model Context Protocol that exposes SeeGeo's audit as tools an AI assistant can call: `audit_site`, `get_audit`, `crawler_check`, `list_audits` and `usage`. It runs remotely at `https://see-geo.com/mcp` (OAuth or API key) and locally as the `seegeo-mcp` npm package (API key). **How do I add SeeGeo to claude.ai?** Settings → Connectors → Add custom connector → `https://see-geo.com/mcp`. Claude registers itself, you sign in to SeeGeo and click Allow, and the tools appear in your chats. Disconnect by revoking the "Claude (connector)" key on your account page. **How do I add SeeGeo to ChatGPT?** Settings → Apps & Connectors → Advanced → Developer mode → Create, with `https://see-geo.com/mcp` as the server URL and OAuth as the authentication. The same consent page appears. A Custom GPT built from `/openapi.json` with an API key is the other route. **Does the connector cost anything?** No. The free plan includes 10 API audits a month, shared across the connector, the API and the npm server; `crawler_check`, `get_audit`, `list_audits` and `usage` are free. Paid plans raise the audit cap. **What can the connector see and do?** Only what its key allows: run audits, read the reports of the account that connected, check crawler access and read plan usage. It cannot change settings, delete audits, see other accounts or billing. The key is listed on your account page and revoking it disconnects the assistant immediately. **Is the audit the same as the one on the website?** Yes — the same engine, the same rubric, the same report URL. The connector adds no scoring of its own; it returns what the audit found, as text and as JSON.