A Google Ads grade of 72.4 does not tell me which dollar to stop spending. For the marketer defending a budget change, I want seven findings with evidence: what is wrong, how much spend it touches, what remains uncertain, and what deserves approval. Google’s Optimization Score is not a measure of profit. Neither is an audit grade. These are the findings I ask Claude connected through groas to deliver instead, with a prompt for each and a clear standard for what comes back.

1. Prove your conversion tracking deserves trust

A duplicate lead can make an expensive campaign look efficient. If the same form is counted by both the Google Ads tag and an imported GA4 key event, with both set to Primary, one submission can become two reported conversions. CPA looks lower, and the account starts telling a flattering story. The first finding should establish whether that story survives a form-fill test.

Before asking Claude to investigate, connect the ad account through groas and set the boundary: analysis only, no changes yet. Reading comes before editing. If tracking is wrong, the rest of the audit rests on bad evidence.

Paste this into Claude:

Read our conversion actions and flag missing, duplicate, or questionable signals. List each action, whether it is Primary or Secondary, whether it counts One or Every, and what evidence you used. Separate confirmed problems from checks that need a test outside the account. Do not make changes.

I want a table, not a paragraph about tracking best practices. Each row should identify the action, the suspected problem, and the check needed to confirm it. A suspicious pair of conversion actions is a lead. A test submission counted twice is evidence.

An advertiser reporting 490 clicks with zero recorded conversions has a problem worth investigating, but the number alone does not diagnose it. Check the status in Goals > Conversions before rewriting the campaign. Unverified, Tag inactive, or Needs attention belongs on the investigation list.

Clear tracking problems before using conversion-based findings to approve changes. Otherwise, Claude is doing precise arithmetic on unreliable inputs.

2. Isolate search-term spend with no recorded return

“Improve negative keyword coverage” is not a finding. A query that spent $400 last quarter without a recorded conversion is something I can investigate.

WordStream’s grader scores wasted spend, Quality Score, impression share, and account activity against its benchmarks. I want the terms and reporting dates underneath the assessment. A grade cannot tell me whether a particular query deserves another dollar.

Timing matters. Google credits conversions to the click date, so a click that converts next week can still look unproductive today. A 90-day window ending 7 to 14 days ago, filtered for cost above target CPA and zero conversions, gives me a useful starting point. It is not a universal verdict on waste.

If target CPA is $60, I use $60 as the initial cost filter. That prioritizes the review; it does not prove every term above it should be blocked or every term below it deserves a pass.

Paste this:

List search terms over the last 90 days ending 14 days ago where cost exceeded our target CPA and recorded conversions equal zero. Include term, match type, cost, clicks, conversion action measured, and reporting dates. Add the available evidence on conversion lag for each campaign, and flag anything you cannot establish.

Call the total exposed spend, not savings. Before recommending an exclusion, Claude needs to account for tracking and conversion lag. If it cannot clear either, that uncertainty goes beside the number, not in a footnote nobody reads.

3. Pair irrelevant intent with the narrowest useful negative

Zero conversions and irrelevant intent are different findings. The first describes a report. The second explains why a query conflicts with the offer.

Take “free invoice template” for paid invoicing software. The query reveals a possible conflict: someone wants a free template, while the advertiser wants to sell software. Claude should state that conflict plainly rather than label the term “low quality” and move on.

The proposed negative needs equal scrutiny. Broad negatives block searches containing all the terms in any order; phrase negatives require the same order; exact negatives block that exact query without extra words. The match type determines how much else disappears with the unwanted traffic.

Paste this:

From the same 90-day window, list the five highest-cost terms where query intent conflicts with our offer. For each, quote the query, state the conflict in one line, propose one negative match type, and flag likely collateral exclusions. Separate clear conflicts from uncertain interpretations.

I want five rows I can read aloud in a meeting, not a bulk upload dressed as analysis. Each row needs the query, its cost, the conflict, and the proposed exclusion.

Bring the query, the conflict, and the narrowest negative that stops it. If Claude proposes a broad negative for “jobs,” make it show what else that would block before approving it. Surgery first. Gardening with a chainsaw can wait.

4. Show which conversion goals guide bidding

Finding 1 checks whether the signals are recorded correctly. This finding checks whether they represent the outcome the business actually wants.

A newsletter signup can be tracked perfectly and still be the wrong goal for a campaign meant to generate qualified calls. Leaving that signup as a Primary action can teach bidding to pursue more signups. Clean measurement does not rescue the wrong objective.

I want Claude to distinguish the actions that guide campaign bidding from those that merely appear in reporting. A list of every conversion action is not enough. The finding needs to connect the settings to the campaigns exposed to them.

Paste this:

Show every conversion action, its Primary or Secondary setting, and its 30-day count. Identify which actions are included in each campaign’s bidding goals and show those campaigns’ spend. State plainly where downstream quality, such as SQLs or revenue, is missing from the account. Do not present campaign spend as an exact allocation to individual conversion actions.

Say the account spends $20k a month and 60% of reported conversions are soft events. That is a reason to inspect the goals, not proof that 60% of the budget bought soft events. Those are different claims. The second requires evidence the first number does not supply.

With groas connected, Claude can read the settings and spend together. The useful output is a map: these campaigns use these signals, this much spend is exposed, and this downstream information remains unknown.

Missing revenue or lead-quality data is unknown, not zero. Take that gap to the meeting alongside the settings, rather than letting a tidy CPA stand in for pipeline.

Illustration of Smart Bidding learning from soft signup signals instead of revenue

5. Put the ad promise beside the landing-page evidence

The ad promises a free trial; the page asks for a credit card above the fold. The description says prices start at $29; the pricing page starts at $79. The ad offers local same-day service; the contact form says replies take 48 hours.

Those are promise gaps. They give the visitor a reason to stop trusting the ad after the click has already cost money.

I do not want Claude to begin with a rewrite. I want it to quote both sides. Otherwise, “improve message match” becomes another recommendation that sounds useful while leaving the actual problem unnamed.

Paste this:

Compare our top-spend ads with the final URLs they point to. Quote the ad claim and the landing-page line that supports or contradicts it. Attach source URLs and identify the spend associated with the ads reviewed. Propose no rewrite unless you can cite the mismatch.

The output should be pairs: claim on the left, page evidence on the right. If Claude cannot access the page or find the relevant wording, it should say so. An unread page is not evidence of a contradiction.

If groas can pull competitor landing pages and funnels alongside yours, ask for one rival page that handles the same promise cleanly. That provides a concrete reference for the decision rather than another lecture about relevance.

Approve a correction to a documented mismatch, not a speculative copy refresh. The quote earns the rewrite.

Paper-cut ad promising a free trial leading to a checkout page requiring a credit card

6. Treat recent changes as leads, not verdicts

Someone changes the budget on March 4. CPA rises on March 9. The timeline looks persuasive, especially when everyone already suspects the person who changed the budget.

It still does not establish the cause.

Change History lists account, campaign, and ad group changes for the past two years alongside performance metrics. That makes it useful for investigation and dangerously convenient for drawing arrows between unrelated events. The same documentation notes limits: API or system actors appear in the User column, and some account-level and Google-side changes do not appear.

A landing-page swap or competitor sale also sits outside that account history. A neat timeline is not a complete one.

Paste this:

List dated account, campaign, and ad group changes from the last 60 days beside weekly spend, CPA, and conversion volume. For each shift, state what changed, when performance moved, how many conversions the comparison rests on, and what outside cause you cannot rule out. Distinguish sequence from evidence of causation.

If a bid strategy switch on April 2 lines up with a CPA increase starting April 7 on 11 conversions, Claude should show those details together. It should not turn them into “the strategy switch caused the increase” because the sentence fits neatly in a summary.

Use the timeline to decide what to investigate, not whom to blame. The 2026 audit checklist puts change review after tracking and intent for this reason. Order protects me from fixing the wrong thing loudly.

7. Request one bounded change with a review date

A grade ends with twenty recommendations and no owner. That helps explain why free account reviews produce surface findings instead of decisions. The document gets longer. The approval gets harder.

One bounded edit gives me a clearer before and after. Ten simultaneous edits leave ten possible explanations for whatever happens next. Scope control does not prove causation, but it makes the result easier to interpret.

Paste this:

Based on findings 1 through 6, propose one bounded action. State the evidence, what remains uncertain, what needs approval, exactly what you will change, and when we review it. Do not apply anything until I approve.

I want something a founder can sign: change this, because of this evidence, then review it on this date or after this much conversion volume. “Review in 21 days or after 30 conversions” is a concrete proposal, not a default that belongs on every account.

A candidate might be demoting a soft Primary action to Secondary or adding two exact negatives from finding 3. The evidence should choose the action, not Claude’s enthusiasm for producing a longer list.

Before asking for approval, turn the finding into a five-line brief:

  • Observation: What Claude found, without interpretation pretending to be fact.
  • Source: The report or page, with the relevant dates.
  • Exposed spend: How much budget touched the issue during the review window.
  • Uncertainty: The tracking, lag, or sample-size question still unresolved.
  • Proposed action: The bounded change and its review date.

Exposed spend is not promised savings. It is the amount I can defend examining. That distinction should survive the trip from the account report to the budget slide.

Through groas, approval matters: autopilot stays off until switched on, and every change is logged with its reason and can be undone in one step. Inspect the reason log after the edit. The record should match what was approved, not merely say that something was optimized.

If I had to start cold on a $20k account tomorrow, I would start with finding 1. Clean tracking determines whether the numbers deserve belief; the bidding-goal review then tells me whether the account is learning from the right outcomes. Get that first table right, then earn the right to argue about search terms.