Connect an AI Visibility Tool to Claude to Automate GEO Reports
Read, propose, or write: how far to let Claude act once your AI visibility tool is connected, and the key scopes that keep a weekly GEO report from turning into an unreviewed publish.
Connecting an AI visibility tool to Claude is quick. Deciding what Claude is allowed to do once it's connected takes more thought, and most setups skip that decision by accident. The same connection that lets Claude read last week's visibility can, with the wrong key, let it create reports, add tracked prompts or push content to a live site.
This post sets out three levels of agency for a Claude-built GEO report, using the Promptwatch MCP server as the worked example, and the scope discipline that keeps each level where you put it.
The connection itself
Claude reaches outside tools through custom connectors built on remote MCP. Per Anthropic's help center, you add one under Customize > Connectors > Add > Custom > Web: a name, the remote server URL, then the authentication Claude detects. You can choose "Sign in now", "Sign in when needed" or "No sign in". Fixed credentials such as an API key go under Request headers.
For Promptwatch, the URL is https://server.promptwatch.com/mcp. It accepts OAuth or a Bearer key, so the header is Authorization: Bearer <key>. MCP and API access are listed on every paid plan.
Two details from Anthropic's docs shape everything below. Connectors run from Anthropic's cloud, so the server has to be public, which a hosted endpoint like this one is. And on Team and Enterprise, an Owner or Primary Owner adds the connector in Organization settings > Connectors, members connect individually, and a shared API key gives every member that key's access. Whatever key you put in the org connector is the agency level for your whole team.
Level 1: read
At this level Claude monitors and summarizes. It takes the weekly copy-paste out of your hands and nothing else.
The key is a read-only project key. Promptwatch issues org, project and read-only keys; a project key sees one project, and a read-only key hides the write tools from the client. Claude doesn't get a "denied" error when it tries to write. The write tools never appear in its tool list.
A read-level report prompt calls a fixed set of tools:
Using Promptwatch, report on the last 7 days vs the 7 before.
getVisibilityTimeSeries for the overall trend and the 5 biggest movers.
getTopCompetitors for who gained on those movers.
getCitationTopPages and getCitationDomainsOverTime for which sources changed.
getCrawlerTrend and getTopCrawlerPages for AI crawler activity on the cited pages.
getVisitorReferrerStats for AI-referred visits by assistant.
Tables only, top 5 rows, flag anything you inferred.
The crawler and visitor calls are what make this more than a mention report. Agent Analytics shows whether ChatGPTBot or ClaudeBot actually read the page that lost its citation, and visitor analytics shows whether the visibility change moved traffic. Crawler logs are on Professional ($245/mo) and up; Essential ($95/mo) lists no crawler-log allowance.
Level 2: propose
Here Claude reads and then recommends. You still hold the decision. It takes the first draft of the "so what" section off your plate.
The key doesn't change. Still read-only. What changes is the prompt. Add three more calls:
listActionItemsto read what Promptwatch's action board already has queued, so Claude doesn't propose duplicate work.listContentGapPromptsandgetContentGapRecommendationsfor prompts with no matching page and the suggested fix.listOffsiteMentionOpportunitiesfor places worth getting mentioned when the citations for a prompt are third-party sources.
Then ask for a ranked proposal list with the evidence behind each item. The output is text in a chat. A strategist reads it, keeps some items, kills others, and acts in the Promptwatch app or hands the work out. Nothing has changed in the project yet.
This is where most teams should stay for a while. A proposal you can audit teaches you how well Claude reads your data before you let it change anything.
Level 3: write
At this level Claude executes. The write tools on the server that make sense for reporting are createReport, which creates a report in the project, and createDashboard, which creates a dashboard. Both need a write key.
The same server also exposes content tools such as createContent and publishContentLive, which drive Content Agents publishing to a connected CMS. Those don't belong in a report loop. A weekly report that ends in a publish call has quietly turned into a content pipeline nobody reviewed.
Scope the write level narrowly:
- Keep the write key out of the shared org connector. The org connector holds the read-only key, always.
- Whoever owns the write step holds the write key separately and uses it only for the run that needs it.
- Connectors are enabled per conversation from the "+" button, so the write connection stays off unless someone deliberately turns it on.
- Never add tracked prompts from a reporting chat. Plans cap responses as well as prompts (Essential lists 50 prompts and 6,000 responses), and a report shouldn't spend allowance.
Which level does what
| Level | Key | What Claude does | What it takes off your plate | Who signs off |
|---|---|---|---|---|
| Read | Read-only project key | Pulls and summarizes | Exporting and charting | Analyst checks numbers |
| Propose | Read-only project key | Ranks next actions with evidence | First draft of recommendations | Strategist picks actions |
| Write | Separate write key | Creates a report or dashboard in the project | Assembling the deliverable | Owner of the write key |
Repeating it without a scheduler
The connector doesn't schedule anything. The repeatable unit is the saved prompt, run in a fresh conversation each week. If the report has to arrive with nobody at the keyboard, use delivery that doesn't depend on Claude: the Promptwatch Slack agent, the Looker Studio connector (Professional, Business and agency plans), or REST API v2 called from your own script.
Guardrails regardless of level
Check the headline numbers against the dashboard every week. Claude summarizes tool output, and a summary can be wrong even when the data is right. Keep one project per conversation. And widen scope one level at a time, only after the lower level has produced reports you trust.
For the full list of what the server exposes and which tools a read-only key hides, see Promptwatch MCP read tools, write tools and scopes.
The tool choice
The levels only work if the tool behind Claude has data worth acting on and a scope model that holds. Many trackers in our directory now list an MCP server, but a report that can only say "you were mentioned" gives Claude nothing to propose. Promptwatch gives it crawler logs, cited pages, AI-referred traffic and an action board on one connector, with read-only keys that remove the write tools outright. Start there, at Level 1, on promptwatch.com.