Auditing Google Ads with AI: A Live Claude + MCP Walkthrough

A Google Ads audit usually means hours of pulling reports, eyeballing tabs, and copy-pasting numbers into a deck. In a recent live session we did it differently: we connected Claude to a Google Ads account through the official Google Ads MCP and let it pull, structure and interpret the data in natural language. This post walks through exactly what that looks like — what it caught, where it confidently got things wrong, and how to set it up yourself.

What the Google Ads MCP actually is

MCP (Model Context Protocol) is the bridge that lets an AI agent like Claude talk to an external system — in this case, your Google Ads account — without anyone hand-building a custom integration. Increasingly every ad and analytics platform ships an MCP, because routing through it is far easier than calling raw APIs (and many tools have no friendly API at all).

Practically: you ask a question in plain English, Claude turns it into the right Google Ads query, the platform returns the data, and Claude analyses and reports back. No SQL, no spreadsheets, no manual exports.

The setup difference: Meta is plug-and-play, Google is self-hosted

This trips people up, so it's worth being clear. The Meta Ads MCP is a single hosted URL you paste into Claude and connect with your Business Manager login — a two-minute job. Google has not shipped a hosted URL. Instead it publishes a GitHub repo, and you host it yourself (we run ours on Google Cloud Run) to generate a working connector URL. It's a bit more technical up front, but if you're comfortable with Claude Code it's a one-time setup.

What you need before you start

• An MCC (Manager) account — you can only generate the Google Ads API token from a manager account, not a standard one.
• API access tier. New accounts start on Explorer access (limited tokens), then graduate to Basic and Standard as Google sees clean usage. On Explorer, configure things so the connection only refreshes when you ask — so you don't burn your token quota.
• The hosted MCP URL from the step above, added as a connector in Claude.

An audit dashboard, built from a prompt

The whole dashboard in the session was generated by prompting — no code. We started deliberately generic to see what metrics Claude would suggest, and it asked the right scoping questions back: lead-gen or e-commerce? Search only? An operator view or a client-facing view? 50 metrics or the 12 essentials? Once the framework was set, it pulled cost, clicks, CTR, conversions, cost-per-result and impression share into a clean weekly view.

The budget leak it caught instantly

One of the first things it flagged: spend bleeding onto the Display network inside a Search campaign. As these platforms get more programmatic, they quietly switch on Display unless you turn it off — and you end up paying for low-intent placements. Because we'd told Claude that framework ("flag any Display spend on a Search campaign"), it surfaced the leak immediately instead of burying it in a tab nobody opens.

Impression share — where your real headroom is

Claude pulled impression share and split lost impression share into budget vs rank. That distinction matters: losing share to budget means you're capped and could scale; losing it to rank means your ad rank is too low and more budget won't help. It's exactly the read you want before deciding where to push spend.

The honest part: AI will confidently give you the wrong answer

This is the most important section. We asked Claude why cost-per-conversion had dropped sharply between two weeks. It produced plausible explanations — smart bidding exiting its learning phase, a new soft conversion action, a changed conversion goal. All wrong. It even mis-read a "page view" goal as a lead submission, which would have completely skewed the read.

The actual reason was a landing-page improvement that lifted conversion rate — something only a marketer with context would know to check. The lesson: an MCP is not a magic wand. Give it a definitive framework and rubric, and it's excellent at applying it fast. Hand it open-ended questions on a blank account and it will hallucinate a confident, wrong story. Always cross-reference its findings against the actual data, and tell it to save your corrected framework to memory so it doesn't repeat the mistake.

Quality Score: it found the landing-page problem

When we asked for a Quality Score breakdown, Claude correctly flagged that landing-page experience was red across the board — and identified why: the page lacked server-side rendering, so Google couldn't read it properly. Its remediation steps were standard but solid (audit Core Web Vitals, match ad headline to landing-page copy, trust signals above the fold, one clear CTA), plus pausing or rewriting the quality-score-1 keywords where the search term and ad didn't match. Useful, structured, fast.

MCP vs the Google Ads API — when to use which

For most accounts, the MCP is enough. If you're running at massive scale (tens of thousands of keywords) or want fully scripted, repeated jobs, building directly against the Google Ads API can be more token-efficient. But you can also engineer your prompts to stay lean, so token cost rarely needs to stop you from using the MCP for day-to-day audits.

MCP vs Google's built-in recommendations

Google's native suggestions are often generic — "switch on maximize conversions," "enable everything in PMax" — and tend to hand more control to the platform. An AI agent given your frameworks and account context can weigh those same suggestions the way an experienced marketer would: accept some, defer others, and explain the trade-off. That context is the difference.

Watch the full live walkthrough

The complete session — dashboard build, budget-leak catch, the cost-per-conversion back-and-forth, and the Quality Score audit — is on our YouTube. It's an honest, unedited look at what works and what doesn't.

The takeaway

AI won't replace your judgement — it removes the grunt work so your judgement goes further. The marketers who win with tools like the Google Ads MCP are the ones who bring a sharp framework and verify what the model says. If you want to learn to set this up and build your own audit skill, that's exactly what we teach in Claude Code for Marketers.