Automating Google Ads with Claude: 4 levels

In April I gave a live demo with a former colleague: setting up Google Ads campaigns with Claude at the core of the process. In about 40 minutes there were four campaigns, with ad groups, keywords with volumes and CPCs, complete responsive search ads with variants to test, sitelinks, price assets and negative keywords. By my own estimate, the same work by hand takes half a day.

The speed interested me less than the question of which parts the machine did and where a person still had to step in. One of the ad texts stated the wrong duration for a training course. Claude wrote it convincingly, and only the review caught the error.

Can you automate Google Ads with AI?

Yes, the execution work: keyword research, campaign structure, ad copy, negative keywords and reporting. That work is repeatable and you can measure it. The choices above it, such as budget and priorities, stay with the marketer.

Most AI marketing tools share one weakness: they work from training data, not from your search volumes and your results. You get generic ad copy with no knowledge of your customers or your market. The fix in the demo was simple: connect real data sources to Claude, so the model works with current numbers instead of assumptions.

How does Claude work with Google Ads?

Claude works with Google Ads through a four-layer stack, where each layer has one job:

1. Keyword data. Real search volumes per market and CPC estimates come from the DataForSEO API, so Claude does not have to guess which keywords are good.
2. Structure. Claude groups the keywords by intent, proposes the campaign structure and writes several RSA variants per ad group. It takes negative keywords from the search terms reports.
3. Execution. An MCP server connects Claude to the Google Ads API. Claude creates the campaigns, uploads the ads and attaches the keywords, with no export files or copying and pasting between tools.
4. Monitoring. Through a second MCP server Claude reads GA4 data. Every week a report with optimisation proposals lands in a Slack channel.

Which level of automation fits your team?

You do not have to build that stack in one go. The demo distinguished four levels, and each one changes something different about your work:

• L1, basic. Claude writes a campaign plan from a briefing or a web search. You build it in Google Ads yourself. The writing work goes away, but the keywords are still the model’s guess.
• L2, data-driven. Claude gets real keyword data, through DataForSEO or a CSV export. The plan then rests on volumes and CPCs.
• L3, automatic. Through the MCP connection Claude puts the campaigns straight into your account. The manual data entry goes away, and the model can now do things that cost money.
• L4, agentic. The weekly optimisation runs by itself: fetch the data, send the report to Slack, attach proposals. A person reviews. The cycle starts without anyone pressing a button.

My rule of thumb: L2 is the best starting point for most teams. It costs little, you learn how well Claude structures campaigns on your data, and nothing goes live without a person. Move to L3 only once the L2 plans have needed hardly any corrections for a few weeks.

What does the marketer still decide?

The marketer decides on budget allocation, geographic targeting, brand-sensitive keywords and which products or services get priority. Those are judgements that need knowledge of the business: what margin a product has, which regions you can deliver to, which competitor brand name you do not want to bid on.

In the demo the geographic targeting had to be adjusted afterwards, and the wrong course duration would have gone live without a review. Claude writes an ad that matches the page it reads. Whether that page is current, or whether the offer is different this month, it does not know. Someone who knows the business has to read every text that makes a claim before it goes live.

So the tool replaces the execution work and leaves the marketer in place. That is also why a 40-minute demo is not yet a system that runs for months. Why that gap is so large is the subject of Speed isn’t quality.

Who owns it when L4 optimises by itself?

At L4 you need one person who approves the proposals and watches the budget. Without that owner, after a month you have a Slack channel full of recommendations nobody opens, or worse, an account making changes nobody has seen.

In the demo setup the MCP server ran centrally for the whole team, with read-only access for most team members. That is a good start. Before you move to L3 or L4, go through this list:

• Owner. One name, not a department. That person approves proposals and is the one to call when something goes wrong.
• Write access. Only the owner and at most one stand-in may make changes to the account through Claude. Everyone else reads.
• Budget cap. A daily or monthly budget that Claude may not raise by itself.
• Proposal before change. Recommendations in Slack stay proposals until the owner approves them. Only after a few months without problems do you let small changes, such as a new negative keyword, through automatically.
• Fixed review slot. Once a week, in the calendar, with the report open.
• Claim check. A person always checks prices, lead times and product specifications in ad copy.

Who is this for?

This kind of AI marketing automation works best at companies whose revenue runs through a digital funnel: e-commerce, marketing agencies and marketing teams with a serious ad budget. There, campaign work is repeatable and you see within days whether a change does anything.

I do not make campaigns or content myself. I build the AI and the integrations underneath the marketing work, around campaign data, analytics and customer data in your CRM. That background comes from Free Now in Hamburg, where I grew the marketing data team from three people to sixteen. How I work with marketing teams and agencies is on my AI agency in Rotterdam page.