Claude AI Google Ads Audit: My Exact Workflow

About 45 minutes into a new client onboarding audit, I had a thought I couldn't shake: most of what I was doing at that moment did not require me specifically. I was reading search terms, checking impression share numbers, comparing spend to conversion volume. I have done that exact sequence hundreds of times. The parts that actually require judgment — structural decisions, bid strategy logic, which signals to trust — those took maybe 20 minutes. The other 70 minutes was just reading.

So I built a direct connection between Claude and Google Ads. No exporting. No formatting CSVs. No copy-paste. Claude queries the account live, in real time, and I spend my time on the decisions that matter rather than the data retrieval that doesn't.

This post walks through exactly how that workflow operates — the setup, the questions, and the one step most people skip that will prevent you from making a wrong call. Everything is based on a real audit of a fine jewelry account I managed for three years, using actual 2024 data.

A standard Google Ads audit for a mid-size ecommerce account takes most specialists 60 to 90 minutes. That time breaks down roughly like this: 30 to 40 minutes reading and organizing raw data, 15 minutes identifying patterns, 10 to 15 minutes forming recommendations, and 10 minutes checking your own logic before writing anything up.

The first chunk — reading and organizing data — is almost entirely mechanical. You are scanning search term reports for irrelevant queries. You are checking each campaign's spend versus conversion output. You are noting which asset groups in Performance Max are dragging the average down. This is pattern recognition at scale, and it is exactly what a language model connected to live data does better and faster than a human reading rows in a spreadsheet.

The last two chunks are not replaceable. Deciding whether a campaign needs a structural rebuild versus tighter match types requires judgment built from managing dozens of accounts. Knowing when "low conversion volume" means the campaign is failing versus when it means the product simply has a long consideration window — that requires context a model does not have automatically. You have to provide it.

The workflow I use now is designed around that split. Claude handles the reading and pattern recognition. You handle the decisions. The result is that the parts requiring your specific expertise get your full attention instead of whatever mental energy is left after an hour of data review.

The first thing that separates this from every "use AI for Google Ads" tutorial you have seen is the data layer. Most of those videos show someone exporting a report, pasting it into ChatGPT, and asking for analysis. That works, but it creates friction: you export, you format, you paste. The data is always a few steps behind where you actually are in the account.

My setup uses the Model Context Protocol (MCP) to give Claude a direct, live connection to Google Ads. When I ask Claude to pull search terms, it queries the account right now and returns actual current data. There is no intermediate file, no manual formatting step. The connection is persistent across the entire audit session.

At the start of every audit I pull exactly three things:

Three queries. About 20 seconds of actual data retrieval time. Then the audit begins.

If you do not have an MCP setup, you can replicate this by exporting those three reports as CSVs and uploading them at the start of a Claude session. You lose the live query capability but the rest of the workflow — the context sentences, the three questions, the cross-check — works exactly the same way. The MCP connection is a speed multiplier, not a requirement.

Before I ask Claude to analyze anything, I give it three sentences. This is the step most people skip, and skipping it is why they get generic answers that could apply to any account instead of specific answers that apply to this one.

The three sentences cover: what kind of business this is, what the customer value looks like, and what I am specifically trying to find in this audit.

For the jewelry account I used in the video, it sounded like this: "This is a fine jewelry ecommerce store targeting US women aged 25 to 45. Average order value is around 300 dollars. I want to find where budget is being wasted and whether the campaign structure makes sense for that AOV."

That is it. Three sentences.

Here is why it matters in practice. Without that context, if Claude sees a search term like "affordable ring under 50," it might flag it as a potential negative depending on what the campaign goal appears to be. With the context that this is a fine jewelry store at 300 dollar AOV, Claude immediately flags that term as an audience mismatch — a buyer at that price point is not the customer, and any spend there is waste. The model cannot make that call confidently without knowing the customer.

The same logic applies to cost per click thresholds, impression share targets, and ROAS expectations. What counts as "high CPC" for a 30 dollar product is completely different from what it means for a 600 dollar ring. Context transforms generic pattern matching into account-specific diagnosis.

Once the data is pulled and the context is set, I ask three questions. Always in this order. The sequence matters because each question builds on what the previous one revealed.

"Show me the top 20 search terms by cost and flag any that don't look like a fine jewelry buyer at this price point."

On the jewelry account I audited in the video, this came back with a clear cluster: variations of "promise ring" and "purity ring." Decent volume, decent spend. The problem is that the buyer searching for a promise ring is typically a teenager with a 40 dollar budget. That is not the customer. Those search terms had been running for months inside campaigns that were otherwise producing sales — which is exactly why nobody noticed. The account was profitable overall, but a slice of the budget was being consumed by traffic that could not convert at 300 dollar AOV.

Claude caught that cluster in about 8 seconds. Reading through 2,000 search terms manually, I might have caught it eventually. I might not have. The point is not that I am incapable of finding it — it is that the mechanical scanning required to find it is the part that should not cost me 40 minutes of concentrated attention.

"Find any campaign where spend is meaningful but conversions are near zero. Give me the most likely structural reason. Consider conversion tracking issues, audience targeting, broad match without negatives, and asset group setup as possible causes."

The last part of that prompt is the critical addition. Naming the failure modes you want Claude to consider changes the quality of the answer significantly. Without that list, you get a confident-sounding general response. With it, Claude looks for evidence of each specific failure mode in the actual account data and tells you which one the data supports.

In this account there was a catch-all campaign running on broad match with no negative keyword list attached. It was pulling in competitor branded terms and generic searches unrelated to fine jewelry. Spending a few hundred dollars a month. Almost no conversions. Claude's structural diagnosis was correct: that campaign needed a rebuild, not just a bid adjustment.

"Given everything you have seen, what would you touch first and why? Back your answer with two specific numbers from the data."

The "two specific numbers" requirement is the most important clause in the entire workflow. It forces Claude to ground its recommendation in the actual account data rather than generating a recommendation that sounds credible but is not anchored to anything real. If it cannot give you two real numbers from the data you pulled, that is your signal that it is improvising. Push back and ask again.

When Claude can back a recommendation with specific numbers — "the promise ring cluster spent $340 last month with 0 purchases" or "the broad match campaign has a 28 dollar CPC against an account average of 4 dollars" — those numbers give you something to verify and something to present to a client with confidence.

Before I touch anything in the account, two things happen. Both are non-negotiable.

First, I open Google Ads and verify the specific numbers Claude cited. Not because Claude is usually wrong — in my experience, the numbers are accurate the large majority of the time. But the one time in twenty that it is not is the time you make a change based on incorrect data and spend three weeks diagnosing why performance dropped. The verification step takes two minutes. That is a cheap insurance policy.

Second, and this is the one most people do not know to do: I ask Claude to argue against its own recommendation.

Literally: "What is the strongest case for not making this change right now?"

On the jewelry account audit, the first recommendation was to cut the broad match catch-all campaign immediately. It was spending budget with near-zero return and the diagnosis was clear. But when I asked for the counter-argument, Claude raised something worth considering: that campaign might be capturing some branded return traffic that does not show up in direct conversions because of attribution lag. It recommended checking the assisted conversion report before cutting anything.

It was right. When I checked, there were assisted conversions attributed to that campaign that would have disappeared if I had deleted it outright. The right call was to restructure the campaign — tighten match types, add a negative list, reduce the budget — not to eliminate it. If I had acted on the first answer without asking for the counter-argument, I would have made the wrong call.

The audit that used to take me 90 minutes now takes about 25. That is not because I am cutting corners on the analysis. It is because the 65 minutes I recovered were not the 65 minutes that required my judgment in the first place — they were the data reading phase that a live AI connection now handles in seconds.

The quality of the output is also better, not because Claude is more capable than I am at making the decisions, but because I am making those decisions with my full attention rather than whatever cognitive capacity is left after an hour of manual data scanning.

Specifically, this workflow consistently surfaces things I would not have caught or would have caught much later:

• Audience mismatches buried in high-volume search term data — terms that look fine in isolation but are wrong for the specific customer and price point
• Campaigns that are spending with no conversion output, where the structural cause is not obvious from the campaign view alone
• Counter-arguments to my own first instincts that prevent premature or wrong changes
• Specific numbers that make client-facing recommendations defensible rather than opinion-based

There are limits. Claude does not have visibility into your client relationship, seasonal business context, or the reasoning behind decisions made six months ago unless you tell it. It does not know that a campaign was intentionally left broad to capture competitor terms during a price-sensitive period. You know those things. The workflow is built on the assumption that you bring that judgment and Claude handles the mechanical reading — not the other way around.

The combination is genuinely faster and better than doing it manually. That is the only honest benchmark that matters.