You have two hours, one Google Ads account, and no permission to change anything. I want Claude to leave that session with a defensible spending decision, not an 87 out of 100 and 40 flags pretending to matter equally.

That is the constraint: inspect, challenge, recommend. Do not execute. Through groas, Claude can inspect the connected account, but this audit gives it no permission to touch a bid, budget, keyword, or setting. The clock expiring does not grant permission either.

Most free audits are built to sell the next conversation. This one has to survive it. Every finding must identify the affected spend, supporting account evidence, uncertainty, and proposed next action. If a founder asks why you want to cut something, “the grader flagged it” is not an answer.

Here is how I spend the two hours, including what I leave out.

1. Minutes 0–10: verify the account and set the no-edit rule

Cost: 10 minutes. Buys: the right account, a defined reporting window, and an explicit boundary.

Connect Google Ads in groas, then add groas to Claude as a custom connector. Before asking for analysis, make Claude repeat back:

  • The customer ID and account currency.
  • The reporting date range.
  • The conversion window used by the actions it will assess.
  • The recent dates excluded from spending decisions because conversions may still arrive.

I use the last 30 to 90 days and initially exclude the last 7 days from performance judgments. Google attributes conversions to the click date, so recent totals can keep rising. Its clicks and cost refresh hourly while conversion data lags by hours. Seven days is a starting buffer, not proof that every conversion has arrived. If the account’s sales cycle runs longer, the audit has to acknowledge that.

Paste this instruction:

Audit this account. Do not make changes. Separate observed facts, hypotheses, and missing evidence. For every finding, report affected spend, supporting account evidence, uncertainty, and a proposed next action. Recommendations are not permission to execute.

That is an operating instruction, not a technical access control. Do not mistake Claude agreeing to the rule for its connector losing write access. Keep the session focused on inspection, and keep autopilot off.

What gets cut: the score. The classic Performance Grader grades a connected account on 10 metrics and returns an overall number. I do not need another number to explain. I need to know which spending decision the evidence supports.

If Claude cannot identify the account and reporting window from connected data, stop. Spending the remaining 110 minutes analyzing the wrong scope does not improve the audit.

2. Minutes 10–30: decide whether the conversion column deserves belief

Cost: 20 minutes. Buys: a decision on whether CPA and ROAS are usable evidence.

Ask Claude to list every conversion action under Goals, including its Primary or Secondary status, count setting, and last 30 days of recorded results. Smart Bidding steers toward what the account counts. Count the wrong outcome and the bidding follows it faithfully. That is not an algorithm mystery.

This is why a serious audit checklist starts with tracking before structure or bidding. Inflated or duplicated conversions can make the rest of the analysis fiction.

Have Claude inspect four things:

  1. Primary outcomes: identify page views, scrolls, time on site, or other actions that do not represent the business outcome being purchased.
  2. Duplicate counting: check whether the Ads tag and a GA4 import both count the same event as Primary.
  3. Count settings: flag lead-generation actions that need review against One per click, and purchases against Every.
  4. Lead quality evidence: for B2B, state plainly whether CRM outcomes come back through an offline import or whether the account stops at the form fill.

Ask for proposed corrections only. “Move this action to Secondary” belongs in the recommendation column, not in the change history.

If Primary includes junk, duplicates, or untested actions, stop ranking apparent waste by CPA. Request two pieces of missing evidence: a live test conversion that fires end to end, and a backend reconciliation against real orders or qualified pipeline. Do not pretend a list of conversion settings proves either one happened.

What gets cut: explanations about expensive keywords before the outcome measurement earns trust. Google can optimize toward cheap form fills while sales starves. A tidy CPA does not settle that argument.

The remaining audit time can still document affected spend, obvious intent mismatches, and evidence gaps. It cannot turn unreliable conversion counts into confident savings estimates. No tracking proof, no conversion-based waste claim.

3. Minutes 30–70: attach money to the strongest findings

Cost: 40 minutes. Buys: a spend-ranked shortlist instead of a settings inventory.

Ask Claude for search terms over the agreed reporting window, sorted by cost, with spend and recorded conversions on each row. Start with the most expensive rows showing zero conversions. That puts the largest potential exposure first, not the easiest setting to criticize.

Split brand and non-brand. Cheap brand clicks can flatter the account average and hide an expensive non-brand problem. Then inspect intent: jobs, free, DIY, courses, and location mismatches. These are candidates to investigate, not a universal negative-keyword list. The query still has to conflict with what the business sells.

For every candidate finding, make Claude return:

  • Affected spend: the money attached to the terms or campaign under review, over a named period.
  • Evidence: the actual terms, conversion records, or settings supporting the concern.
  • Uncertainty: what the account data cannot establish.
  • Proposed action: the smallest defensible next move, still unexecuted.

Before calling zero-conversion spend waste, clear two traps.

First, conversion lag. A click that converts next week looks unproductive today. The initial seven-day exclusion helps, but a longer sales cycle can still defeat it. Claude must explain whether the chosen window is mature enough for the claim.

Second, missing queries. Google omits some low-volume searches for privacy, so campaign spend can exceed the sum of visible search-term rows. Report that coverage gap. Do not label the invisible portion waste because the visible report cannot explain it.

Thin conversion data also limits bid-strategy judgments. Do not turn a sparse report into “tCPA is broken” or “tROAS needs replacing.” If the evidence supports only a containment proposal or a later test, say that. Diagnosis does not improve when Claude sounds more certain.

What gets cut: the exhaustive tour of every setting. Forty minutes spent finding all possible imperfections is forty minutes not spent examining the expensive ones. For the hands-on version of this pass, use the search-terms prompt against the live account.

Every retained finding needs a spend figure and a limit on what that figure proves.

4. Minutes 70–100: make Claude argue against its own recommendations

Cost: 30 minutes. Buys: fewer recommendations, with clearer failure conditions.

Paste this:

For each recommendation, explain what evidence would make it wrong. Separate what the account shows from what you infer. Do not execute any recommendation.

Claude can hand back a clean fix that matches the premise of the question. I want it to challenge that premise before anyone pays for the answer.

A negative keyword looks certain until a converting query in the same cluster undermines it. A budget shift looks obvious until the supposedly losing campaign has a 9-day sales cycle and the analysis judged it on 7 days. These are not footnotes. They change the decision.

Sort the survivors into three buckets:

  • Proposed immediate containment: clear irrelevant intent, no plausible conversion path, and a narrow proposed exclusion. Name the scope and the risk of blocking useful traffic. “Immediate” describes priority after approval, not permission during this audit.
  • Controlled test: a structure or bidding hypothesis where the evidence points one way but lag or volume leaves doubt. Propose one campaign, one change, and a review date.
  • Insufficient evidence: an interesting pattern that cannot yet carry a decision. Record the missing evidence and leave the account alone.

That third bucket matters. Without it, every observation eventually becomes an action because nobody wants an empty recommendation column.

Aim for three to five fixes ranked by money and evidence, not forty equal flags. If only two survive, keep two. The audit does not owe anyone a longer list.

What gets cut: recommendations that cannot explain their own failure condition. “Best practice” is not a substitute for an account-specific reason.

A fix without a falsifier is a guess. Put guesses in the evidence bucket, not the execution queue.

5. Minutes 100–120: write the memo the budget owner can approve

Cost: 20 minutes. Buys: a decision document with explicit limits, not a sales presentation.

Here is where audits quietly inflate their value: they add up every dollar touching a suspected problem and call it savings.

Affected spend is the money passing through the issue. Estimated recoverable savings is the portion judged wasted. Those are different numbers, and the second remains an estimate.

Say the account spends $20k a month. The audit flags $2k in Display expansion that appears mostly irrelevant, $4k in campaigns with no real negatives, and $3k in brand-heavy Performance Max. That is not automatically $9k in savings.

The proposed judgment might be most of the $2k, about a third of the $4k, and about half of the $3k. Those shares need evidence and uncertainty attached. The calculation is affected spend times judged-wasted share, not affected spend times enthusiasm. Do not total overlapping findings as though the same dollar can be saved twice.

Have Claude produce a short tracking verdict followed by this table, ranked by affected spend. Use the actual audited period in the column heading. The row below illustrates the format, not an account result.

FindingAffected spendAccount evidenceJudged-wasted share + estimated savingsUncertainty / risk if wrongProposed actionOwner + review date
Example: jobs intent on non-brand$1,840 over 90 daysTop 12 terms; 0 recorded conversionsState a supported estimate, or mark unestimatedCould block a converting query; confirm lag and trackingPropose narrowly scoped negativesYou; 14 days after approval

The required fields do useful work. Affected spend establishes scale. Evidence makes the claim inspectable. The uncertainty column prevents confidence from becoming decoration. The action and review date turn the finding into a decision someone can own.

If the wasted share cannot be defended, leave savings unestimated. The row can still justify gathering evidence. Affected spend is not a savings promise.

What gets cut: the slide deck, account grade, and speculative total at the top. The memo should be short enough that the budget owner reads the caveats before approving the action.

That is the distinction behind searches for a Google Ads performance grader. A grade summarizes checks. This workflow uses groas to connect Claude to the actual account so it can show terms, spend, and settings behind a recommendation. groas Unlimited costs $19 a month, flat, for the whole team. The subscription buys access; the memo still has to earn approval.

At minute 120, stop. Deliver the memo with the account unchanged. An unsigned decision document is a completed audit, not permission to keep working until someone says yes.

6. When the constraint lifts: execute approved rows, not the whole audit

Audit time cost: none. Execution starts separately, after explicit approval. Buys: a boundary between diagnosis and changes.

Send the memo to whoever owns the budget. Get approval on specific rows, not on the general idea that the account needs improvement. The two-hour limit ends the audit; only authorization lifts the no-edit constraint.

Then give Claude a narrower instruction: apply the approved containment rows only, and list each proposed change before making it. Do not bundle the test bucket into the same request because everyone is already on the call.

After execution, open the groas action log. Check what changed and the reason attached to each line. The log is the execution record, not proof that the change worked. Compare results at the agreed review date. If they move the wrong way, use the record to identify and undo the relevant change rather than guessing what happened.

Keep autopilot off. In groas, it stays off until switched on, and approving an audit recommendation is not permission to enable it.

What stays cut from the first execution pass: bid-strategy moves, budget reallocations, and a Performance Max rebuild. Where tracking needs repair, those wait for clean evidence rather than joining the containment batch. Moving faster on dirty conversion data only gives the account more opportunity to pursue the wrong outcome.

This is the structural problem with automated audits that fail at diagnosis: checking a setting is easier than defending what to do about its spend.

If the approved containment holds up at its review date, book the next two hours for the test bucket. One campaign, one change, same memo. No score, no 40-flag PDF, no edits nobody can defend.

Two hours make Claude show its work. Approval determines what runs next. Defensible spend beats flattering scores every quarter.