Claude Code is not just for developers. Connect it to groas over MCP, and the terminal becomes a scriptable PPC and SEO operator: review an account, save the findings to disk, then apply the changes you approve.

By the end, you will have a connected workspace, three saved review prompts, and a reusable /audit command. The point is not to make marketing look like programming. It is to stop rebuilding the same review every Monday.

Before starting, have these ready:

  • Claude Code access: a supported paid Claude or Console account. The free claude.ai plan is not enough; follow the setup instructions for your account.
  • A groas account with your accounts and sites connected: groas Unlimited is $19 a month, flat, for unlimited teammates, ad accounts, and websites.
  • A project folder: use it to keep prompts and reports together. I separate Google Ads, Meta, and site work so the files stay readable.

That last point is organization, not security. A folder called google-ads does not stop a connected tool from touching Meta. Put account boundaries in the prompt, and check them before approving anything.

1. Install Claude Code and sign in

Run the installation command from the Claude Code setup instructions:

npm install -g @anthropic-ai/claude-code

Open your project folder in the terminal, then start Claude Code:

claude

Finish the browser sign-in.

Expected result: the terminal opens a Claude Code session tied to your account.

Common mistake: trying to sign in with only a free claude.ai account. Claude Code needs supported paid access; a working browser chat login does not establish that you have it.

If login loops, try starting from an empty folder. That helps separate project configuration trouble from account trouble. Get a basic session working before adding the marketing connection. Debugging two things at once is how a five-minute setup becomes your afternoon.

2. Connect groas over MCP and confirm the account list

Outside the Claude Code session, run:

claude mcp add --transport http groas https://mcp.groas.ai/mcp

Then check the registered connection:

claude mcp list

If groas needs authentication, open Claude Code and enter:

/mcp

Select groas and complete the sign-in flow. The MCP documentation covers connection and authentication troubleshooting.

Once connected, ask:

List my connected ad accounts and websites. Show their names and account or site identifiers. Do not change anything.

Expected result: groas appears as connected, and Claude returns the accounts and sites you recognize.

Common mistake: confusing connector scope with account permissions. The default local scope keeps the connector configuration specific to your project; project scope writes a shareable .mcp.json. Neither is a substitute for naming the account you intend to review. See the scope guide before sharing configuration through Git.

I start with local scope. Share the setup only when the prompts are stable, and never put a real API key in a shared file.

Do not continue until the account list is right. A beautifully formatted audit of the wrong business is still the wrong audit.

3. Run a Google Ads waste audit and save it

In your Google Ads project folder, paste this prompt. Replace the bracketed account value with an identifier from Step 2.

Audit only this Google Ads account: [ACCOUNT NAME OR ID].

Review the last 30 days for:
- Wasted spend.
- Search terms with over $50 spend and zero conversions.
- Signs of broken conversion tracking.
- Campaigns capped by budget.

Rank findings by wasted spend where it can be established.
Keep tracking concerns and budget constraints separate from confirmed waste.
State what you could not verify.

Save the review to audit.md. Include today's date, the account
identifier, and the reporting period.
Do not change anything.

Expected result: an audit.md file with a search-term table, a tracking section, and a list of budget-constrained campaigns. Each finding should identify what Claude actually found, not merely repeat the checklist.

Common mistake: asking for an audit without asking for a file. A chat answer is easy to scroll past. A saved report gives the next review something concrete to compare against.

Read the output against the 10-point Google Ads audit framework. Start with money out versus money tracked. If conversion tracking is broken, the bid recommendations built on it deserve suspicion.

Also keep the categories straight. A campaign hitting its budget cap is not automatically wasting money. It is a constraint to investigate, not a reason to cut spend.

Treat the report as a review queue, not an execution list. The threshold finds candidates. It does not make the decision for you.

4. Check Meta fatigue without pausing ads

Switch to your Meta project folder. Start with one account so the first output is easy to inspect:

Check Meta creative fatigue for this account: [ACCOUNT NAME OR ID].

Flag:
- Frequency over 3 in the last 7 days.
- CTR down more than 20% versus the prior 14 days.
- The same creative live for more than 21 days.

Show which signals each ad meets. List ad IDs to review for pausing
and replacement creative to draft. Check the post-click path before
recommending a pause, and state anything you cannot verify.

Save to meta-fatigue.md. Include today's date and the account identifier.
Do not pause, edit, or publish anything.

Expected result: meta-fatigue.md contains an ad-by-ad review list, the signals behind each recommendation, and replacement ideas.

Common mistake: treating every performance decline as tired creative. Check the post-click path before killing the ad. A broken landing page does not become a creative problem because the frequency column looks untidy.

This is why I use groas rather than a Meta-only connection. The review can draw on Meta, Google, and the landing page through the same infrastructure instead of stopping at a platform report. The ChatGPT walkthrough shows the same review-and-approval approach in a browser workflow.

Once the single-account output is useful, expand the prompt to all intended Meta accounts. Keep the account identifiers in the report so the pause list never becomes a guessing game.

Use fatigue signals to choose what to inspect, not what to switch off automatically.

5. Find striking-distance queries and draft three briefs

In your site folder, paste:

Review this site's Search Console data: [SITE].

Find queries with average positions 8 to 20 in the last 28 days
and more than 200 impressions. Return the top 10 matching queries
by impressions, then group them by intent.

Draft three article briefs with:
- Headline.
- Angle.
- Target query group.
- Two relevant competitor URLs each, where available.

State any unavailable data. Save to briefs.md with today's date
and the site identifier. Do not publish anything.

Expected result: briefs.md contains the query shortlist and three briefs you can hand to a writer. If the data does not support three useful briefs, the report should say so rather than fill the gap with confident nonsense.

Common mistake: asking for everything. An unlimited-looking prompt produces an unreadable brief long before it produces a useful editorial plan.

I use a row limit because the job is to pick the next pieces of work, not export the internet. Ten queries are enough to inspect whether the grouping makes sense before widening the review.

Read the angle in each brief, not just the headline. Three titles aimed at the same intent are not three distinct assignments.

Keep the shortlist small enough to make a decision. Expand it after the first set produces useful work.

Terminal printing a PPC review to audit.md beside ad performance charts

6. Apply one approved fix, then rehearse undo

Pick one small, reviewed change from audit.md. Do not choose it solely because it crossed the $50 threshold.

Before asking for execution, request a preview:

From audit.md, prepare one proposed exact-match negative keyword
for my review. Show the account, campaign or ad group, search term,
match type, and reason.

Do not apply it. Wait for my explicit approval.

Inspect the proposal. If it is correct, approve that specific change:

Apply only the exact negative keyword change I just approved.
Do not make any other changes. Show the resulting change log and reason.

Expected result: the approved negative is added, and groas records the change with its reason. Check the account, term, match type, and destination against the proposal.

Then rehearse the reversal:

Undo only the change you just made. Show the reversal and confirm
that the negative keyword from that change is no longer applied.

Common mistake: assuming Claude Code’s permission mode guarantees a separate approval for every advertising action. Review the permissions documentation, keep approvals enabled, and make the stop-before-write instruction explicit. Do not treat a conversational instruction as a substitute for checking tool permissions.

Keep budget changes and publishing under manual approval. Do not bypass permissions just to save clicks. I use groas autopilot for unattended work only after setting its rules; that is a separate decision from running this review.

Prove the undo path on one small change before approving a batch. A log is useful. A reversal you have actually checked is better.

7. Save the review as a reusable /audit command

Create this file in the project root:

.claude/commands/audit.md

Put this prompt inside it:

Audit only this Google Ads account: $ARGUMENTS.
If no account is supplied, ask for one before proceeding.

Review the last 30 days for wasted spend, search terms over $50
with zero conversions, signs of broken conversion tracking,
and campaigns capped by budget.

Rank findings by wasted spend where established. Keep tracking
concerns and budget constraints separate from confirmed waste.
State anything you cannot verify.

Save to audit.md with today's date, the account identifier,
and the reporting period. Do not change anything.

Run it inside Claude Code, replacing the placeholder with the intended account:

/audit [ACCOUNT NAME OR ID]

Expected result: the same review runs without pasting the full prompt, and audit.md updates in the project folder.

Common mistake: turning the review shortcut into a money-moving shortcut. Keep this command read-only. Review the file, then approve changes separately.

The file location controls where the command is available. Project commands live in .claude/commands/; personal commands can live in ~/.claude/commands/. Claude Code also supports the newer .claude/skills/<name>/SKILL.md form; the slash-command guide explains the distinction.

Share the prompt once it is stable. Keep credentials out of it, and pass the account through $ARGUMENTS rather than burying a client name in a reusable instruction.

Make the repeatable part the review, not the approval. That is the useful division of labor.

Paper-cut Ads, Meta, and SEO folders feeding into one terminal command file

Know when the browser is the better tool

Use the terminal for repeated work with saved outputs. Use browser chat when the job happens once and a file adds nothing.

Say you run one local service account spending $5,000 a month and check performance twice a week. Skip this setup if it creates more administration than it removes. Claude in the browser with groas connected can handle the conversation without a project folder.

The terminal earns its place when the work repeats: weekly audits across five accounts, monthly fatigue checks, briefs every Friday. A stable prompt and a saved file matter more on the tenth run than the first.

Claude Code is not a dashboard, either. Ask for the output you need, and read it before acting. Scriptability does not remove judgment. It just makes a bad instruction easier to repeat.

Questions before the first run

Do I need to know how to code?

No. You need to copy commands, edit account values, and inspect the results. The important skill here is recognizing whether the audit makes sense. Pasting a command is the easy part.

Why use groas rather than separate marketing connectors?

I use groas to keep the review and the approved fixes in one workflow. Separate SEO, Meta, and CRM connections leave the operator stitching the picture together; that fragmented approach shows up in this marketing connector roundup.

With groas, the same connection brings together ad accounts, SEO data, rival ads, and site context, with logged, reversible changes. Reporting-only connectors leave the fixing to you. That is the part I want out of the spreadsheet.

Does Claude Code replace groas autopilot?

No. This workflow runs when you invoke it; autopilot handles ongoing work under rules you set. I keep deliberate reviews separate from unattended changes. A saved audit is not permission to manage an account indefinitely.

Verify the setup, then change the threshold

Close the terminal, reopen it in the same project folder, and run /audit with one account identifier. Check:

  1. The file: audit.md updates with today’s date, the correct account, and the reporting period.
  2. The evidence: search terms have spend attached, and unavailable data is labeled rather than invented.
  3. The connection: claude mcp list still shows groas as connected.
  4. The boundary: the review makes no account changes.

If the connection fails, fix it before rewriting prompts. If the wrong account appears, fix the scope of the request before approving anything.

The first thing to change after verification is the dollar threshold. The example uses $50 for a $20,000-a-month account. At $5,000 a month, try $20. At $200,000, try $200 so the report does not drown you in rows. Treat those as starting points, not laws.

That is the shift: chat gives answers; Claude Code with groas gives saved reviews and approved fixes. Keep approvals on for money moves, tie thresholds to spend, and let the machine keep the files.