A search-term report should not end with another export. Cursor was built for code, but its agent plus groas makes it an ads cockpit: read the live account, save the analysis to a file, and push an approved change in the same session.

By the end of this tutorial, you will have a working Google Ads connection, two saved reports, and a reviewed set of negatives pushed live:

  • reports/waste.csv: search terms to review before blocking.
  • reports/weekly.md: a weekly account summary you can keep and compare.
  • A groas change log showing what changed and why.

Before starting, have Cursor installed, a groas account with Google Ads connected, and permission to change the selected ad account. Open the workspace where you want the reports to live. Say you are spending $20k/month. This is the setup I would use before letting an AI tool touch it.

Why use a code editor for Google Ads?

Google’s own Google Ads MCP server is read-only. It can expose account data, but it cannot close the loop by modifying bids, pausing campaigns, or creating assets. The assistant finds wasted search terms. You still add the negatives yourself.

Cursor with groas connects the three parts: read the account, save the finding, and apply the approved change. The useful difference is not another chat window. It is the file between the recommendation and the action.

I used to tell clients to export search terms to Sheets for review. The review still matters. In Cursor, the reviewed file becomes the input: reports/waste.csv goes in, approved negatives go out, and the weekly summary stays at reports/weekly.md, where git can version it.

Use the editor when the file earns its keep. Otherwise, it is another window to babysit.

1. Add groas to Cursor’s MCP configuration

Open ~/.cursor/mcp.json for global use, or .cursor/mcp.json inside the client repo for a project-specific configuration.

Add the groas entry without replacing any existing server entries. If the file is new, paste this configuration:

{
  "mcpServers": {
    "groas": {
      "url": "https://mcp.groas.ai/mcp"
    }
  }
}

Save the file.

Expected result: Cursor has a groas server entry available in its MCP settings.

Common mistake: Replacing an existing mcpServers object and losing the other connections. Add the entry; do not bulldoze the file.

2. Authenticate the account that owns the connection

Open Settings > MCP, find groas, and click its connection control. At the time of writing, the unauthenticated status reads Needs authentication.

Sign in with the groas account where you connected Google Ads.

Expected result: The status changes to connected, and Cursor lists groas tools available in Agent mode.

Common mistake: Signing in with a personal groas account that has no ad accounts linked. The server can connect perfectly and still have nothing useful to show. Check the account before troubleshooting the editor.

Optional: use an API key instead

If browser sign-in does not fit your setup, open Account settings > API Keys in groas and create a key. Add it as a header to the same server entry, following the Cursor connection instructions. Save and restart Cursor.

Keep the key out of the repo. Use ~/.cursor/mcp.json rather than a project configuration that gets committed to git.

Expected result: The same groas tools appear without the browser sign-in flow.

Common mistake: Versioning the credential alongside the reports. The reports belong in git. The key does not.

3. Prove the connection with a read-only request

Switch to Agent mode and open a chat. Replace the bracketed account identifier before sending this prompt:

For Google Ads account [ACCOUNT NAME OR ID], list campaigns for the last 30 days with spend, conversions, and cost per acquisition. Confirm which account you used. Do not change anything.

Approve the groas read call when Cursor requests it.

Expected result: Cursor returns campaign rows with the selected account’s real numbers. Confirm the account name or ID before continuing.

Common mistake: Starting with a budget edit. If the connection or account selection is wrong, find out on a read, not on a write.

Use Agent mode throughout this tutorial. Ask mode explains; Agent mode calls tools and writes the files this workflow depends on.

4. Set write approvals before preparing changes

Open Settings > Cursor Settings > Agents > Approvals & Execution and select Allowlist. Review which tools you have already allowed to run without intervention.

I use Allowlist for client accounts because I want writes to stop for review. Do not pre-approve groas write tools for this walkthrough. A broad groas:* allowlist entry permits the whole server, not just reads.

Cursor’s security documentation distinguishes workspace file access from terminal commands and MCP calls. Approval behavior also depends on the selected mode and tool permissions, as the MCP documentation explains. Treat these settings as guardrails, not a hard security boundary; the run-mode documentation roundup makes that distinction explicit.

Expected result: Your configuration does not authorize unattended groas writes, and you know where to inspect a requested tool call before approving it.

Common mistake: Assuming that choosing an approval mode cancels an existing broad allowlist. Check the permissions themselves. The label is not the control.

5. Pull the waste report into a named file

Keep every working prompt specific about three things: the account, the date range, and the file path.

Replace the account identifier and send:

For Google Ads account [ACCOUNT NAME OR ID], pull 90 days of search terms. Write every term with spend over $50 and zero conversions to reports/waste.csv with columns term, campaign, spend, clicks, conversions. Summarize the top 10 by spend in the chat. Do not change anything in the ad account.

Expected result: reports/waste.csv appears in the workspace with the requested columns, plus a short chat summary.

Open the file. Check one term’s spend against Google Ads using the same account and date range. If the numbers match, the connection is doing something more useful than producing fluent prose.

Common mistake: Asking only to “find wasted search terms.” Without a date range or destination, you can get a polished answer and no reusable artifact.

The $50 threshold identifies candidates for review. It is not permission to block them.

A miniature paper factory diverts wasted search terms into a bin while useful terms continue along a conveyor

6. Save the weekly summary instead of screenshotting it

Use the same account identifier:

For Google Ads account [ACCOUNT NAME OR ID], build a weekly summary for the last 7 days. Write it to reports/weekly.md with these sections: Spend, Conversions, Cost per acquisition, Top 5 campaigns, Top 5 search terms by spend, and 3 changes you would make today with the reason for each. State the account and date range in the report. Use only my live account data. Do not change anything in the ad account.

Expected result: reports/weekly.md exists with the six requested sections and numbers for the stated period.

Open it in Cursor’s sidebar. Read the recommendations, not just the totals. These are proposals, not instructions you have already approved.

The report reads like a note rather than a dashboard. That is the point. You can diff it next week, commit it, and show a client what the account did without another Sheets export.

Common mistake: Asking for “a summary of my ads” and leaving the result in chat. Next week, the screenshot becomes the reporting system. Always name the file.

A weekly Google Ads summary saved as a Markdown file beside a stack of approved negative keywords

7. Review the file, then push only approved negatives

Open reports/waste.csv before requesting any write. Delete every row you do not want blocked.

A term with zero conversions is not automatically a bad term. I keep the business review here, not after the write: a client may still want a term for brand reasons even when the report makes it look wasteful.

This walkthrough uses account-level negatives. Keep only terms you intend to block across the account, not terms that are merely wrong for one campaign. If that scope is wrong, stop rather than approving the batch.

Save the file, then send:

For Google Ads account [ACCOUNT NAME OR ID], read reports/waste.csv. Prepare to add only the terms in that file as account-level negative keywords. Before making changes, show the exact terms, proposed match types, and scope, then wait for my confirmation. Do not add anything else. After the approved write, report the number added and identify any failures with the reason.

Read the proposal. Confirm it only if the terms, match types, account, and scope match your intent. Then inspect and approve the requested groas write calls.

Expected result: Cursor reports how many negatives were added and names any failures. Verify the account-level negatives in Google Ads against the reviewed file and approved proposal.

Common mistake: Approving a batch you have not read. The CSV is the review surface, not a magic safety device. Saving it prepares the input; approving the proposed write authorizes the change.

This is where groas goes beyond a read-only connector: the reviewed report becomes an action, without rebuilding the list somewhere else.

8. Check the change log and test undo

Open the groas change log after the write. Compare its entries with the approved proposal and Cursor’s completion report.

Through groas, the agent can read campaigns, search terms, budgets, keywords, and ad status. It can also make approved changes, with each change logged alongside its reason and one-step undo available. That write-and-review loop is what matters here, not how many tools appear in Cursor’s sidebar.

Test undo once on a single test negative before relying on it for a budget change. Inspect the corresponding account state afterward.

Expected result: The log accounts for the changes you approved, and the test reversal behaves as expected.

Common mistake: Treating a successful chat response as the audit trail. The response is a receipt. Check the account and the log.

A checkpoint holds an ad change for approval beside a change log and undo lever

When I would skip Cursor

Cursor is overkill for a quick question. Its working environment is files and code, which means someone still has to open the editor, read the output, and review changes.

I would skip it when:

  • The team works entirely in Sheets and chat.
  • Nobody responsible for the account wants to use an editor.
  • The task is a one-off question such as “Why did spend spike yesterday?”
  • There is one account, one campaign, and no useful reason to keep a repo.

Use ChatGPT or Claude with the same groas connection instead. Ask in plain language and keep the approval habit. The chat route is covered in connecting Facebook Ads Manager to ChatGPT without the CSV shuffle; the connection-and-approval pattern also applies to Google Ads.

I keep Cursor for work where files earn their keep: weekly reports I diff, waste lists I review line by line, and site plus ads changes that belong in the same session.

Cursor pays when the file is the workflow. When the answer is the workflow, stay in chat.

FAQ: what this setup does—and does not—require

Can Cursor actually change Google Ads?

Yes. The write path runs through groas, rather than stopping at a report. In this tutorial, the sequence is deliberately slower than “find waste and fix it”: pull the data, review the file, inspect the proposed negatives, approve the write, and check the log.

Keep that separation. A recommendation and permission to execute it are different things.

Do I need to know how to code?

No, but you need to tolerate an editor. You paste a JSON configuration, connect the account, and type prompts in Agent mode. You do not write application code.

You do name file paths and read files before approving changes. If that already sounds like work you will avoid, use the groas connection in ChatGPT or Claude instead. Do not build a review workflow around a tool nobody will open.

9. Verify the workflow, then make reporting a habit

Check these three things in order:

  1. Verify the files. Open reports/waste.csv and reports/weekly.md. Confirm the selected account, the relevant date ranges, and the figures against Google Ads. The waste file covers 90 days; the weekly report covers 7.
  2. Verify the live changes. Compare the account-level negatives in Google Ads with the terms, match types, and scope you approved. Reconcile any failures Cursor reported.
  3. Verify the audit trail. Check that groas records each change with its reason and that the single-negative undo test worked.

If those checks disagree, stop before the next write. A file saved successfully does not prove an account changed correctly.

Once all three line up, the first thing I would change is the reporting cadence, not another campaign setting. Run the weekly-summary prompt every Monday and commit the file. Let the waste list build for a week before reviewing the next batch.

That file history becomes the audit trail. It is the reason to use an editor for ads in the first place.