Cindy Zhu.
Free guides
Marketing Claude

Read-only Google Ads analyst

Hey, it's Cindy 💡 This setup deliberately starts with an analyst that cannot spend money. Google’s current official Google Ads MCP server is read-only, so Claude can query account data.
btw the waitlist for AI System Club is now open! If you're keen on using AI to automate your content, video editing, job search, business or build AI agents, feel free to sign up (only if you want to hehe). Ps. the waitlist gets first dibs:)

What you need

  • a Google Ads account you are authorised to access
  • a 22-character Google Ads developer token
  • a Google Cloud project ID
  • OAuth credentials or Application Default Credentials
  • Python and pipx
  • Claude Code or another local MCP-compatible host

Official source: https://developers.google.com/google-ads/api/docs/developer-toolkit/mcp-server

Read-only does not mean no setup. The developer token, Cloud project, OAuth, and local server are required.

Connect the official server

Add this to your MCP configuration. Replace both placeholders locally. Never record or publish the real token.

{
  "mcpServers": {
    "google-ads-mcp": {
      "command": "pipx",
      "args": [
        "run",
        "--spec",
        "git+https://github.com/googleads/google-ads-mcp.git",
        "google-ads-mcp"
      ],
      "env": {
        "GOOGLE_PROJECT_ID": "YOUR_PROJECT_ID",
        "GOOGLE_ADS_DEVELOPER_TOKEN": "YOUR_DEVELOPER_TOKEN"
      }
    }
  }
}

Authenticate with the OAuth method approved for your account, restart the host, then ask:

What customers do I have access to through google-ads-mcp? Return customer ID and account name only.

If that fails, check the developer token, project ID, OAuth account, the Google Ads account permissions, and whether pipx is available.

What GAQL means

GAQL is Google Ads Query Language. It selects resources, segments, and metrics from the Google Ads API. The MCP server exposes:

Start by asking Claude to show the fields it queried. That makes the result auditable.

The first trust test

Paste:

Use google-ads-mcp in read-only mode.

For account [CUSTOMER ID], compare the last 7 complete days with the previous 7 complete days by campaign and device.

Return:
1. The exact date windows and timezone used.
2. Spend, conversions, conversion value, cost per conversion, and conversion value / cost.
3. Campaigns where spend increased but conversions did not rise proportionally.
4. Observations separated from hypotheses.
5. The GAQL fields and filters queried.
6. Missing tracking or business context that could change the interpretation.

Do not recommend a budget or bid change until you have shown the evidence. Do not imply you can edit the account.

Check three rows manually inside Google Ads. If the numbers disagree, stop and resolve date range, timezone, attribution, currency, conversion actions, or manager-account selection before trusting the analysis.

Ten useful read-only prompts

  1. Compare the last 7 complete days with the prior 7 by campaign and device.
  2. Rank campaigns by incremental spend, then show whether conversions moved with it.
  3. Find campaigns with cost and clicks but zero primary conversions.
  4. Split branded and non-branded Search performance using the naming rules I provide.
  5. Show conversion action mix by campaign and flag changes in what counts as a conversion.
  6. Find device segments whose cost per conversion is at least 30% worse than the account median. Show sample size.
  7. Compare weekday and weekend performance without recommending action.
  8. Find campaigns limited by missing or ambiguous data rather than poor performance.
  9. Produce an observation, hypothesis, evidence-needed table for the five largest anomalies.
  10. Create a weekly analyst memo with exact queried fields and a “human context required” section.

The evidence ladder

Claude earns trust in stages:

  1. Reproduce: the query matches the account UI for a small sample.
  2. Explain: it separates measured observations from possible causes.
  3. Repeat: the same prompt remains useful across several weekly reviews.
  4. Improve: your instructions capture business context it previously missed.

Do not add write access because one confident answer sounded right. The current official server is intentionally read-only.

Approval rule for any future automation

Keep campaign creation, publishing, spend, bids, budgets, audiences, and destructive actions behind explicit human approval. Use separate credentials and tools if you ever add write capability. A reporting server is not a permission system for future tools.

Sources

the waitlist is now open

AI System Club is launching soon

Build AI systems with me through recorded lessons you can follow at your own pace. Join the waitlist for first access and the founding price.

btw this list gets first dibs
the moment doors open :)