Connect an AI Visibility Tool to Claude to Automate GEO Reports
Six setup steps and one prompt: connect the Promptwatch MCP server to Claude, map nine GEO report sections to named MCP tools, and run the same report every week.
One server URL, one key, one prompt. That is the whole stack for a GEO report Claude can rebuild every week. The hard part is not the connection. It is deciding which tool call fills which section of the report, so the output reads the same on week 12 as it did on week 1.
This guide uses Promptwatch as the data source. It has a hosted MCP server, read-only keys, and a tool for every section in the table below.
What you need
- A Claude account. Custom connectors run on Free, Pro, Max, Team and Enterprise. Free allows one custom connector (Anthropic).
- A Promptwatch paid plan. MCP and API access are listed on every paid plan. Essential is $95/mo, Professional $245/mo, Business $579/mo; agency plans run $199, $399 and $799.
- Professional or higher if you want the crawler section. Essential lists no crawler-log allowance. Professional lists 25M.
Setup in six steps
- Create a read-only project key. Promptwatch issues org, project and read-only keys. A project key sees one project. Read-only hides the write tools, so Claude never receives them.
- Open the connector form. In Claude: Customize > Connectors > Add > Custom > Web.
- Enter the server. Name it, then paste
https://server.promptwatch.com/mcp. Review the detected authentication. Promptwatch accepts OAuth or a Bearer key. - Add the key. For a fixed key, use Request headers:
Authorization: Bearer <key>. Pick "No sign in" if the header does all the work, or a sign-in option if you use OAuth. - Team or Enterprise only. An Owner or Primary Owner adds the connector under Organization settings > Connectors. Members then click Connect. A shared key gives every member that key's access, which is why step 1 says read-only.
- Enable it per chat. Open a conversation, click "+", turn on the connector.
Claude connects from Anthropic's cloud, not your machine. The server must be reachable on the public internet. Promptwatch's hosted endpoint is. A self-hosted server behind a firewall would not be.
Report sections mapped to tools
Nine sections. Each one names the tool that fills it, so Claude doesn't improvise.
| # | Section | Promptwatch MCP tool | Answers |
|---|---|---|---|
| 1 | Visibility trend | getVisibilityTimeSeries, getBrandVisibilityOverTime | Did we go up or down, and on which prompts |
| 2 | Competitors | getTopCompetitors, getCompetitorHeatmap | Who took the share we lost |
| 3 | Cited pages | getCitationTopPages, getCitationDomainsOverTime | Which URLs the engines used |
| 4 | Offsite sources | listRedditCitations, listYoutubeCitations | Which threads and videos get cited |
| 5 | Sentiment | getSentimentTimeSeries | How the answers describe us |
| 6 | AI crawlers | getCrawlerTrend, getTopCrawlerPages | Did the bots read the pages that matter |
| 7 | AI-referred traffic | getVisitorReferrerStats, getVisitorTopPages | Which assistants sent visits, to which pages |
| 8 | Content gaps | listContentGapPrompts, getContentGapRecommendations | Prompts we don't answer yet |
| 9 | Next actions | listActionItems | What is already queued |
Rows 6 and 7 are the ones most trackers can't fill. In our catalog, LLM Pulse and Mentionable both list MCP access, and both listings show no AI crawler-log product. A Claude report built on either has no row 6. Promptwatch covers all nine from the same connector.
Optional row 10 for ecommerce: getShoppingTopProducts and getShoppingProductsOverTime for ChatGPT Shopping. Shopping data needs Business ($579/mo).
The prompt
Store it somewhere the team can copy it. Run it unchanged.
Use the Promptwatch connector. Build the weekly GEO report for this project,
last 7 days vs the previous 7.
Sections, in order, using only these tools:
1 getVisibilityTimeSeries 2 getTopCompetitors 3 getCitationTopPages
4 listRedditCitations, listYoutubeCitations 5 getSentimentTimeSeries
6 getCrawlerTrend 7 getVisitorReferrerStats 8 listContentGapPrompts
9 listActionItems
One table per section. Top 5 rows max. Flag any number you calculated
rather than read. End with three lines: what moved, likely cause, next step.
Fixed tool list, fixed row limits, fixed comparison window. Change any of the three and week-over-week comparisons break.
Running it every week
Claude does not get a scheduler from the connector. The repeatable part is the saved prompt: open a new chat on the same day each week, enable the connector, paste.
When the report has to land without anyone opening Claude:
| Option | What it does | Plan note |
|---|---|---|
| Promptwatch Slack agent | Posts to a mapped channel | Org owner connects it |
| Looker Studio connector | Live dashboard for stakeholders | Professional, Business, agency |
| REST API v2 | Scripted pulls into your own pipeline | Same keys and scopes |
For which surface fits which job, see Promptwatch MCP vs REST API.
Optional write step
With a write key, Claude can call createReport to create a report in the project, or createDashboard to create a dashboard. Useful if stakeholders live in Promptwatch rather than in chat. Keep it a separate, deliberate run. The weekly read stays on the read-only key.
Guardrails
- Read-only by default. The shared connector never holds a write key.
- Spot-check the headline visibility figure, one competitor share and the crawler total against the dashboard. Claude summarizes tool output. It can misread a table.
- Protect the response budget. Plans cap responses as well as prompts: Essential lists 50 prompts and 6,000 responses. A reporting chat should read data, not add tracked prompts.
- One project per chat. Mixing projects in one conversation mixes numbers.
Verdict
MCP access is common now. Full report coverage is not. Most trackers can feed Claude a visibility table. Promptwatch feeds it the visibility table plus crawler logs, citation pages and AI-referred traffic, which is the difference between a report that says what moved and one that says why. Start with Professional if you want all nine rows. Compare the rest of the field on the leaderboard, or go straight to promptwatch.com.