Why Claude should not run your Google Ads

The short version

Claude can be extremely useful around a Google Ads account. It can review structure, query data, summarise changes, inspect landing pages, write hypotheses and help build the tracking layer. That is different from letting it run the campaign.

The core bidding engine in Google Ads is already a predictive AI system. It is not guessing from a CSV export. It is making auction-time decisions with data that no external agent receives.

Google is already the agent in the auction

Smart Bidding sees the live auction. Claude sees the text you paste into it, reports you export, screenshots you upload, API fields it can access and whatever warehouse tables you connect.

Google sees far more at decision time: query context, device, location, time, browser and app context, previous conversion likelihood, modelled conversion probability, competition, user behaviour patterns and auction signals that never appear in your interface.

That is the real asymmetry. Claude can reason about the account. Google can predict the auction.

Campaign setup is not just clicking Next or calling an API

There is also a practical reason not to hand campaign creation to an agent or connector. The Google Ads interface is not just a slow way to set fields. It is a control surface.

When you create a campaign manually, you see each choice in context. You check the conversion goals. You inspect location settings. You notice whether Google is pushing search partners, broad targeting, auto-created assets, final URL expansion, brand exclusions or asset sources. You see the warnings, recommendations and defaults as they appear.

If Claude changes the setup for you, the danger is not that it cannot press buttons. The same risk exists if it uses the Google Ads API, an MCP-style connector or another automation layer. It may make a setting choice that looks minor, saves correctly and is difficult to spot later. A campaign can be technically created and still be commercially wrong.

The interface is better than a hidden agent

For setup work, the best use of Claude is to prepare the plan before you go into the interface: campaign objective, budget, conversion goals, naming, assets, audiences, exclusions, URL strategy and checks. Then you go through Google Ads deliberately and verify each step.

That is slower than blind automation, but it is safer. You still have to check the machine. If you have to check it anyway, checking each step while the interface shows the actual account state is better than reviewing a finished campaign after an agent, API call or connector has made invisible decisions.

Google Ads already has AI in the workflow

The choice is not manual drudgery versus Claude doing everything. Google Ads already has Google AI inside the campaign workflow. It can help draft ad groups, keywords, headlines, descriptions, sitelinks and images. Asset tools inside Google Ads can also help generate or edit images and create video assets for campaigns.

That means the interface is not necessarily much slower. You can let Google's own AI produce candidates, then approve, reject or edit them while the account settings are visible. That gives you automation and control at the same time.

The wrong trade is letting an external agent make silent account changes, then trying to audit what happened afterwards. Native automation inside the interface is easier to supervise because it sits next to the real campaign state.

The "TV guide" problem

Search intent is contextual. The phrase "TV guide" could mean television listings, a magazine brand, a streaming help page, a setup manual, a channel list, a navigational query, an informational query, or even someone researching what TV to buy.

An external AI sees the words and maybe a row in a search terms report. Google sees the auction context. It can infer intent from signals around the search, not just the query string itself.

This is why broad match and Smart Bidding can work when the conversion signal is clean. The work is not to bolt another agent on top of Google and ask it to outguess the auction. The work is to make sure Google is optimising toward real business outcomes.

Negative keywords need context

This is where AI can be useful, but also dangerous. Claude can review search terms and find possible negative keyword candidates. That is a good use case. It can group themes, flag obvious junk, surface repeated low-quality patterns and prepare a review list.

What it should not do is automatically negate terms without context. A search term can look irrelevant in isolation but still be part of a profitable path once Google has auction, audience and conversion signals around it. With Smart Bidding, it is usually better to err on the side of caution.

Before a term becomes a definite negative, you need to understand the offer, margin, geography, intent, landing page, match type, conversion quality and whether Google is already filtering through bid pressure. The safest workflow is candidate discovery first, human review second, account change last.

Where Claude belongs

Claude is useful where the task is reasoning, review or infrastructure rather than auction-time prediction.

• Audit campaign structure and conversion actions.
• Stress test platform ROAS against warehouse, CRM and order data.
• Find contradictions between GA4, Google Ads, backend revenue and CRM status.
• Draft landing page variants, ad copy angles and testing plans.
• Write SQL, Python, n8n logic and reporting documentation.
• Summarise search term themes for human review.
• Find negative keyword candidates that still need business-context review.
• Prepare campaign setup checklists before changes are made.
• QA tracking plans, enhanced conversions and offline conversion feedback.

Where Claude does not belong

Do not let an external agent own the parts where hidden context and silent settings matter most.

• Do not ask it to outbid Smart Bidding from the outside.
• Do not let it blindly create campaigns through the interface, API or connector without human verification at each step.
• Do not let it automatically add negative keywords without account context.
• Do not let it change conversion goals, bidding, targeting or account settings without a visible review path.
• Do not treat a campaign that was created successfully as a campaign that was set up correctly.

I do not use AI to second-guess Google's auction system. I use it to build the measurement layer Google needs to make better decisions.

The real work

Clean conversion tracking. Enhanced conversions. Offline uploads. Qualified lead feedback. Product-level LTV analysis. Landing pages built for intent. Campaign structure simple enough for Smart Bidding to learn.

That is the useful version of AI-powered PPC. Claude helps build the operating system around the account. Google handles the auction.

Check your PPC data layer