A useful Google Ads audit gives you a short list of fixes someone else can verify. Most people start with a grade instead, then discover that a B+ cannot answer the questions finance asks: what should we change, what proves it, and what happens if we do nothing?

I want a decision document, not an account report card. With Claude connected to groas over MCP, that means pulling live account data, separating confirmed problems from plausible explanations, and proposing changes only when the evidence supports them.

Here is the sequence: confirm the account, check measurement, investigate spend, rank the findings, and ask for narrow approvals. Skip the score. Keep the evidence.

Why won’t a free grade get a change approved?

A grader compares your account with a benchmark. WordStream’s version offers an instant A–F grade across 10 metrics, including wasted spend and Quality Score. PPC Grader analyzes more than 40 signals through live API access. Those comparisons can flag areas worth investigating. They do not establish whether spend caused a business loss, whether tracking recorded the outcome, or which change your approver should sign.

A claim like $1,240 per month in recoverable waste sounds like money found. Then someone asks when the conversions were supposed to arrive, which conversion action counted, and whether the tag fired at all. Now you have an investigation, not a saving.

If you cannot point to the period, the entity, and the measurement behind a number, you do not have a defensible finding. You have a slide.

Use a grader for a quick benchmark if you want one. Use Claude with groas to build the account-specific evidence and approval trail the grade leaves behind.

Clay auditor studying a giant report card beside an unplugged account-data cable

1. Am I auditing the right account and period?

Before asking for recommendations, connect Claude to groas over MCP and make it confirm the basics:

  • Account ID.
  • Reporting currency.
  • Account time zone.
  • Audit period and comparison period.

This follows the practical pattern of confirming account IDs and targets before pulling data. With groas, the connection runs through one sign-in for your connected accounts. groas Unlimited costs $19 per month, flat, for the whole team. The account you want to audit must actually be connected. No connected account, no audit.

Start with this prompt:

Read this Google Ads account for the last 30 complete days, compared with the prior 30 complete days. Confirm the account ID, currency, time zone, and exact dates before analyzing performance. Do not change anything.

That gives you a written scope line for the final document and keeps an incomplete day out of the comparison. It also states your instruction not to make changes.

A read-only instruction is not a permissions control. Before proceeding, inspect what the connected tools can do and how approvals work. This guide to reading an MCP tool list for scope, approvals, logging, and undo covers that boundary.

I do not want recommendations from the wrong account delivered with perfect confidence. If Claude cannot restate the account and dates, stop there.

2. Can I trust the conversion numbers?

Smart Bidding learns from the conversion signal you feed it. If that signal is broken, a tidy CPA comparison can describe the tracking failure rather than the advertising performance.

So I check measurement before pricing any supposed waste. Use this prompt:

List every conversion action, its status, primary or secondary setting, source, and last recorded conversion date. Flag spend with zero conversions, any action with a greater than 50% week-over-week conversion drop without a corresponding click drop, and any action that stopped recording. For each flag, separate the observed fact from possible explanations. State what evidence is needed to distinguish a tracking problem from a genuine performance problem. Do not change anything.

The structure follows the conversion sanity-check workflow used with live account data. The important part is the distinction between an observation and its explanation. “Conversions stopped recording” is something to investigate. It does not, by itself, prove why they stopped.

Claude can inspect what the connected account exposes. Account data alone cannot prove that a website tag fires correctly or that reported conversions reconcile with backend orders.

For actions marked Unverified, Tag inactive, or Needs attention, Google’s Tag Assistant troubleshooting process involves launching the test from Google Ads, connecting a landing URL, triggering the conversion, and checking the result. That is a browser test, not proof that the money reconciles.

Bring separate evidence for the failures an account report cannot settle:

  • Missing or duplicated tags.
  • Thank-you page refreshes counting the same conversion repeatedly.
  • Scrolls or video views treated as primary conversions.
  • GA4 versus Google Ads attribution differences.
  • Stripped GCLIDs.
  • Consent-mode gaps.
  • Delayed offline imports.

Those are the issues covered in this seven-failure tracking checklist. Assign the relevant checks rather than asking Claude to guess their results.

For the longer explanation, read why signal quality matters more than bidding strategy. For this audit, the rule is shorter: if measurement is unreliable, make fixing it the first decision. Do not present budget cuts as proven savings.

3. Which spend actually needs an explanation?

Once measurement holds up, ask where spend and outcomes diverge. I use three cuts so Claude cannot pick the ugliest campaign and call the audit finished:

For the last 30 complete days versus the prior 30 complete days, show:

  1. Campaigns with spend and zero conversions.
  2. Campaigns where CPA rose at least 20%, or conversions fell while spend held steady.
  3. Search terms with at least 20 EUR-equivalent spend and zero conversions. State the spend threshold used in the account currency.

Include entity IDs, spend, conversions, CPA where applicable, and period-over-period changes. Group search terms into themes such as jobs, DIY, or wrong category. Show the spend behind each theme. Treat these as investigation candidates, not proven waste.

This combines the 30-day zero-conversion spend review with a search-term workflow using an explicit spend floor. The threshold starts the review. It does not decide the verdict.

Ask for the themes and the math, not just a long list of terms. Then check three things before calling anything wasted.

Has enough time passed for conversions to arrive?

Google credits conversions back to the click date, so recent performance can change as conversions arrive. Compare equally mature periods rather than treating fresh clicks as finished results. This conversion-lag breakdown explains the problem.

Say you spend $20k per month with a 14-day sales cycle. Last week’s zero-conversion spend may include clicks whose conversions have not arrived yet. Tell Claude which recent results remain provisional; do not let it turn them into a negative-keyword list by default.

Does the business context change the decision?

Branded research clicks, traffic from locations you do not serve, and a sole ad stuck in rejected status need different responses. One may need patience, another an exclusion, another an ad fix.

The account checklist covering location targeting, budget-limited winners, and rejected ads is useful here. Give each issue to the person who can establish the business rule. A spend table cannot tell Claude which service areas you actually cover.

How much spend could Claude inspect?

Search-term reporting does not expose every query, and Performance Max gives you less query detail. The discussion of hidden search terms in Claude audits makes that limitation worth stating explicitly.

Ask Claude to report the share of spend represented in the visible search-term data. Unseen spend is an evidence gap, not permission to invent a waste estimate.

4. How do I turn observations into a ranked fix list?

Claude will produce ten suggestions unless you define what earns a place on the list. I require six fields:

  1. Entity ID: the campaign, ad group, or other affected entity.
  2. Source period: the exact dates behind the finding.
  3. Supporting numbers: spend, conversions, and the relevant comparison.
  4. Uncertainty: what the account data cannot establish.
  5. Proposed action: one specific change or evidence-gathering task.
  6. Consequence of doing nothing: what remains at risk, without invented revenue forecasts.

Use this prompt:

Turn the findings into a ranked fix list using those six fields. Label each item Confirmed, Plausible, or Missing evidence. Rank verified problems requiring action first, then unresolved issues by spend at risk. Do not call spend recoverable unless the evidence supports that claim. Do not estimate recoverable revenue.

The labels need discipline:

  • Confirmed means the evidence establishes the stated fact. A conversion action showing Tag inactive is a confirmed status, not proof that every associated click was wasted.
  • Plausible means the account suggests an explanation, but lag or business rules could still account for it. Last week’s zero-conversion spend inside a 14-day sales cycle belongs here.
  • Missing evidence means the next step sits outside the available data: a Tag Assistant test, backend call logs, or confirmation of a service area.

A confirmed observation is not automatically a confirmed cause. Keep that distinction in the document, not buried in the chat.

Where a finding needs more context, pull the relevant material: 90 days of search terms, account settings, six months of change history, or a conversion-action export showing primary and secondary status and source. Use those records to investigate the issue, not to make the report look heavier.

If a proposed fix has no entity ID, no period, or no supporting numbers, send it back for evidence.

A short fix list backed by evidence outweighing a tall pile of account grades

5. What does the approver actually sign?

Nobody should have to approve a 40-message transcript. Compress the audit into one page with a final prompt:

Summarize the audit in one page with four sections:

  1. Confirmed issues, including entity IDs, periods, spend, and conversions.
  2. Unresolved questions, each with an owner and the evidence needed.
  3. Proposed changes, showing current value versus proposed value.
  4. Explicit approval requests.

Preserve the distinction between confirmed facts, plausible explanations, and missing evidence. Add no new findings.

One page, four sections, no new findings. What the deck calls full audit coverage, I call a transcript nobody can verify.

Skip recovered-revenue math, benchmark grades, and fixes that depend on evidence you have not supplied. If Claude says pausing three ad groups will save $4,800 per month, remove the claimed saving until measurement is trusted and lag is accounted for. Spend that might stop is not automatically waste recovered.

Give every unresolved question an owner:

  • The tracking owner runs Tag Assistant.
  • The backend owner supplies orders or call logs.
  • The account owner confirms service areas, exclusions, or out-of-stock SKUs.

An audit with three owned follow-ups beats one with twelve unowned guesses. Keep the one-pager. Everything before it is working material.

6. How do I approve changes without approving everything?

Approve only the lines that show the current value and proposed value. Apply just those changes through groas.

Say you spend $20k per month and the one-pager proposes reducing one campaign budget from $150 per day to $80 per day and adding three exact-match negatives. Approve those specific changes, not a vague instruction to “optimize the account.”

The reason to use groas here is practical: every change is logged with its reasoning, and any change can be undone in one step. After the run:

  1. Open the action log.
  2. Match each entity ID and change against the approved one-pager.
  3. Review the outcome over the next complete buying cycle.
  4. If the approved change needs reversing, undo the setting.

Undo reverses the setting. It does not recover the spend. Put that distinction in the decision document so nobody mistakes reversibility for a refund.

The audit is not permission for an open-ended cleanup. It is a request to make named changes for stated reasons, then check what happened.

7. What should I do first today?

Connect Claude to groas, paste the scope prompt, and get the account ID and dates back in writing. Confirm tool permissions and approvals. Then run the conversion check before asking Claude to price a single “wasted” click.

If tracking is broken, your first fix is measurement. If tracking holds, build a short ranked list with entity IDs, periods, evidence, and owners. Keep plausible explanations out of the confirmed column. Keep missing evidence attached to a person who can supply it.

That is the job: turn account data into a decision someone else can inspect and approve.

A grade tells you how you look. A fix list tells someone else what to sign.