An autopilot does not decide where the plane goes. It holds the course somebody set, through the long boring middle of the flight, and the pilot stays in the seat. That is the right picture for Google Ads in Claude or ChatGPT: hand over the weekly routine, keep the hand on budget and on the switch that makes ads serve.

What a Google Ads week consists of

Most of account management is not strategy. It is the same five checks, every week, in an interface that needs a dozen clicks for each:

  • Which search terms did we pay for that we never wanted?
  • Is each campaign spending at the pace its budget allows?
  • Which keywords cost money and produced nothing in 30 days?
  • Which ads lose against their siblings in the same ad group?
  • What changed compared with last week, and is it a trend or a weekend?

All five are reading, comparing and writing a short summary. That is exactly the work a language model with access to the account does well and without getting bored.

Level 1: the report that writes itself

Start read-only. Connect the account, and give the model one standing prompt:

For the last 7 days against the 7 days before: spend, clicks, conversions and cost per conversion by campaign. Then the 20 search terms with the highest cost and no conversion. Then every campaign that reached its daily budget on more than three days. Tables first, then five sentences on what I should look at.

If your client can run a prompt on a schedule, this arrives on Monday morning without anybody asking. Nothing in the account can change at this level, because the connection has no write tools until somebody ticks Allow changes. For many teams this level is already most of the value: the hour of clicking through reports is gone, and the meeting starts with the findings.

Level 2: the model prepares, you approve

The search terms report produces the most reliable chore in Google Ads: negative keywords. The model reads the terms, groups the irrelevant ones (job seekers, free, login, a competitor's support hotline) and proposes a list.

From those search terms, propose negative keywords with match type. Show me the list, the spend each one would have saved in the last 30 days, and anything you are unsure about. Do not change anything yet.

You read the list, strike two entries, and say go. With changes allowed, the model creates a negative keyword list, fills it and attaches it to the campaigns. The same pattern works for pausing keywords without conversions, for drafting new ad variants into an existing ad group, and for building a complete new campaign that waits, paused, for your review.

The sentence "do not change anything yet" is a habit, not a safeguard. The safeguard has to sit somewhere the model cannot talk its way past.

What the connector lets through

We wrote about the differences between vendors in Letting Claude change Google Ads. For Loopthink's connector the rules are these, and they are enforced on the server, not in the prompt:

  • A campaign is never switched on. New campaigns are created paused, and an existing campaign cannot be set to enabled. Whether money starts flowing is decided by a person in Google Ads.
  • Nothing is removed. Pausing is possible, deleting is not. A paused keyword can be switched back on, a removed one is gone together with its history.
  • Inside a campaign the model can work. Ad groups, ads and keywords can be created, changed, paused and enabled. Budgets can be created and changed.
  • Negative keyword lists grow, they do not shrink. The connector can create a list, add keywords and attach it to campaigns. Taking a keyword out of a list or detaching a list is a click in Google Ads.
  • Every change is validated by Google first and then written, and the audit log holds who asked, what changed, and the values before and after.

Who may use the change tools at all is a role. The working student gets the reports, the account manager gets the changes.

What stays with the pilot

Three things we would not hand over, whatever the tool allows.

Budget decisions. The connector can change a budget, and it is logged. But "raise the budget because cost per conversion looks good" is a decision about the company's money made on seven days of data. Let the model make the case and a person make the call.

Bidding strategy. A change of strategy sends a campaign back into its learning phase. The model can tell you that target CPA would fit. It cannot know that sales is on holiday next week and cannot take the leads.

Anything the data does not contain. Margin per product, which leads became customers, the campaign the CEO likes. An autopilot flies by instruments. It does not see what is not on them.

A setup that works

  1. Connect the account read-only and run the weekly report for two or three weeks. You learn where the model's reading of your account is reliable.
  2. Give one person the change tools. Start with negative keywords, the change with the smallest downside.
  3. Keep the rhythm: report, proposal, approval, change. Look at the audit log once a month, the way you would look at Google's change history.

That is not a self-driving account. It is an account where the routine takes fifteen minutes instead of half a day, and the half day goes into the questions no report answers.

Google Ads in Claude →
Google Ads in ChatGPT →