Most Google Ads audits are run once, produce a list nobody keeps, and are argued from scratch the next time. The part Claude changes is not the analysis — it is the repeatability.
Through a connector, Claude reads the spend, search terms, conversions and pacing in your live account itself, ranks what is losing money, and proposes fixes rather than applying them. The audit takes about thirty minutes; the same checks by hand take roughly three hours.
The checkpoints themselves are assistant-agnostic. The seven-checkpoint AI audit has them in order, each with the prompt to run and the number that signals a problem: tracking, search terms, pacing, bidding, keyword sprawl, creative, geo. Read that if you have never run one. This page is about the three things that are Claude's own — a project that remembers your definitions between audits, a context window that takes a 90-day export in one paste, and artifacts, which turn findings into a document you can actually send.
Claude reaches your account through LoomaScale's listing in Claude's connector directory: one click plus a sign-in, nothing to paste, and it works on free and paid Claude plans alike.
First, what connecting actually grants
Worth being straight about this, because a claim you will see on competing pages is wrong.
Google Ads has no read-only OAuth scope. Google's own OAuth documentation lists exactly one scope for the Google Ads API, and it covers reading and writing both. Anyone offering you a read-only Google Ads connection is describing something Google does not sell. We were repeating that line ourselves until we checked.
So the guardrails have to live in the tools, not in the consent screen:
- Nothing that can spend gets created. Campaigns, ad groups and ads are created paused, and the create tools have no status field to set otherwise — it is not a default you can talk the model out of.
- Enabling something re-checks your daily cap first. The budget tool refuses budgets shared across several campaigns, and the account-budget tool will raise a spending limit but will not set it to unlimited.
- An audit only touches read tools anyway. Listing ad accounts, campaigns, keywords, search terms, ads, conversion actions, sitelinks and asset groups; insights overall, by geography and by audience; keyword ideas, location lookup, and the billing and pacing checks. Fifteen of the forty-three Google Ads tools, none of which write anything.
Then there is the rule you write yourself, which is the one that governs day to day.
Put the audit's definitions in a project
This is the genuinely Claude-shaped move, and it is what makes the second audit cheaper than the first.
Create a project for the account and paste this into its instructions, edited to fit. Every conversation you open inside that project inherits it.
You audit Google Ads for [business, one sentence]. Market: [countries, language]. A conversion is [action], worth roughly [value]. Count a search term as waste at more than [20] in spend with zero conversions. Compare every period against the same period a year earlier, not the one before it. Standing rules for every conversation in this project: - state the date range you used before any number - show the comparison period next to the current one - never estimate search volume or CPC; say where to get them - name the finding you are least confident about - propose changes, never make them without my explicit yes
Two of those lines do most of the work. The waste threshold stops the argument about what counts, and it belongs to the account: twenty dollars is generous at an $8 cost per acquisition and meaningless at $300. The least-confident line is how you tell an audit from a fluent summary — a total with no stated uncertainty is a headline.
The payoff arrives next quarter. Open a new conversation in the same project and the definitions are already there, so the two audits are comparable. Without a project you re-specify all of it, slightly differently, and the numbers quietly stop lining up.
Run it as one conversation, not seven
Open the audit with the connection check, because everything after it is worthless if Claude is answering from general knowledge:
List the Google Ads accounts you can currently see, and for the main one tell me the most recent date you have complete data for. Names and dates only, no analysis yet.
A good answer is boringly specific: account names, customer ids, a date. Generalities, or a request that you paste something, mean the connector is not attached to this conversation — the setup guide's troubleshooting section covers why.
Then work the seven checkpoints in order, one message each, in that same thread. Two reasons to keep it in one conversation rather than starting fresh each time:
- Checkpoint 1 constrains the rest. If conversion tracking is wrong, every efficiency number downstream is confidently wrong, and Claude will only carry that caveat forward if it can still see it.
- Follow-up questions are free. Asking where the account wastes money makes the agent list campaigns, pull insights, pull the search terms report and then propose negatives — four calls, and none of them is metered. Plans are priced by the ad account, which counts once a month however much you ask of it. That is the whole reason to ask "now exclude the branded ones" instead of accepting the first table.
Keep the follow-ups adversarial. "Which of those would you drop if you had to defend the list to the person who owns the budget?" gets you a shorter and better list than asking for more findings.
Turn the findings into an artifact
The parent playbook stops at a prioritized list. This is where Claude goes one step further, and it is the step that makes the audit presentable to somebody who was not in the conversation:
Turn this audit into a report as an artifact. Structure it: the three findings costing the most money, with the number attached to each; what I should do about each one and in what order; what you could not check and why; and one paragraph a non-marketer can read. Use only figures that appeared in this conversation. Do not round anything up.
"Only figures that appeared in this conversation" is not decoration. It is the instruction that keeps a summarizing step from inventing a plausible number to fill a gap, and it is easy to verify — every figure in the artifact should be findable above it in the thread.
Ask for the artifact at the end, once. Asking after every checkpoint produces seven documents and no report.
No connector yet? The paste version
You can run a reduced audit without connecting anything, and Claude is unusually good at it because a full export fits in one message.
Export the search terms report and the campaign report for the last 90 days, paste both, and ask for checkpoints 2, 3 and 7 — waste, pacing, and geo leakage. Those three are the ones a flat export can genuinely answer.
The limit is worth stating plainly, because it is where pasted audits go wrong: Claude can only see what you pasted. Export campaign-level data and ask an ad-group question and you will get a guess or a refusal, and the guess reads exactly like the refusal did not happen. A connected account has no such edge, which is why the paste version is a trial rather than a workflow.
How often to re-run it
A full seven-checkpoint audit is quarterly work. Checkpoints 2, 3 and 7 are worth a monthly pass — about twelve minutes inside an existing project.
On automating it: Claude offers recurring tasks on its paid plans, but whether a scheduled run can reach your connector depends on where you schedule it from, so test one before you depend on it. For a quarterly audit, a calendar reminder is also sufficient.
What does deserve automation is the gap between audits, where a dead conversion tag or a capped budget can burn a month unnoticed. The six-check weekly review is built to run unattended as a scheduled ChatGPT Task, and there is no reason both assistants cannot be pointed at the same account.
If you would rather see the output before doing any of this, the free Google Ads waste audit runs the money-losing half of the list from a single prompt and proposes fixes for your approval. And once the audit is a habit, fifteen Claude prompts for Google Ads covers the work between them.
