A first AI audit of a Google Ads account is six reads in a fixed order: tracking, wasted spend, search terms, ads, structure, verdict. Each step below gives you the exact prompt, what the answer means, and what to do about it. Total time: thirty minutes. Total risk: none — reads cannot change an account.
This is the doing companion to
what an AI Google Ads audit checks, which covers
the reasoning behind each check. You need a connected
MCP client and your customer ID. Start every prompt with
the account — In account <account>, ... — and keep the whole audit in
one conversation so later steps can reference earlier findings.
Minutes 0–5: is the conversion data believable?
Every later judgement divides by conversions, so start by checking the denominator.
In account <account>, list the conversion actions: name, category, status,
counting method, and how many conversions each recorded in the last 30
days. Anything that looks like double counting or a dead action?
What the answer means. Healthy accounts show a small set of primary actions with steady volume. Trouble looks like: zero conversions anywhere (tracking broken or never set up), two actions counting the same purchase (inflated results), or a "page view" style action treated as primary (everything downstream is optimistic).
What to do. If tracking is broken or double-counted, stop the audit here and fix tracking first. The connector can do much of that fix with you. As of v2.16.0 it has a read-only conversion setup audit that says whether tracking can be trusted, and — each step proposed first, applied only on your yes — it can create a new conversion action, stage its tag and trigger in Tag Manager for you to publish, or mark an Analytics event as a key event and link Analytics to Google Ads. Demoting or switching off an action that double-counts stays in the Google Ads UI: the connector never edits an existing conversion action. Every number in steps two to six is only as honest as this one.
Minutes 5–10: where is money going with nothing back?
In account <account>, for the last 30 days: which campaigns, ad groups and
keywords spent the most with zero conversions? Three short tables, ranked
by cost.
What the answer means. Some zero-conversion spend is normal — low-volume keywords need time. The findings that matter are concentration: one keyword quietly eating a meaningful share of a campaign's budget, or a whole ad group that never converted in a month of steady spend.
What to do. Mark the top three rows. They become pause candidates on your work list — not pauses now. The discipline of this audit is that it reads first and acts later, with each action getting its own approval.
Minutes 10–15: what are people actually searching?
In account <account>, pull search terms from the last 30 days with cost
but no conversions, ranked by cost, top 25. Classify each: irrelevant,
competitor, research intent, or plausibly fine. Note which keyword matched
it.
What the answer means. This is usually where the audit pays for itself. Irrelevant terms with real cost are pure waste. A cluster of research-intent terms ("how to", "free", "what is") matching a broad keyword tells you a match type is too loose. Competitor names are a strategic choice you should be making deliberately, not by default.
What to do. The irrelevant cluster becomes tomorrow's negative
keyword list — the mechanics are in
the negative keyword sweep,
and the match-type reasoning in
negative keyword automation. If one
keyword matched most of the junk, add "review match type of
Minutes 15–20: are the ads actually running, and are they tested?
In account <account>: first, list ads that are disapproved or limited by
policy, with the policy topic. Second, list ad groups with zero or one
enabled responsive search ad.
What the answer means. Disapprovals are silent — the campaign looks enabled while parts of it cannot serve. Policy topics tell you whether the fix is copy (easy) or landing page (a task). Ad groups with a single RSA have no creative experiment running: whatever that ad does is the ceiling.
What to do. Disapprovals go to the top of the work list; they are blocked distribution, not optimisation. Single-RSA ad groups become a drafting session later — the agent can write candidate RSAs and create them on approval, as in the campaign launch walkthrough.
Minutes 20–25: is the structure fighting itself?
In account <account>, three structural checks for the last 30 days:
1) campaigns limited by budget, with lost impression share;
2) negative keywords that conflict with keywords I'm bidding on;
3) geo targets or ad schedules that look unintentional — spend in
locations or at hours that don't fit a business like mine.
What the answer means. Budget-limited winners are your cheapest growth: demand exists that the cap is refusing. Negative conflicts mean one hand is paying for clicks the other hand is blocking. Off-pattern locations and always-on scheduling are defaults nobody ever revisited.
What to do. Budget-limited converting campaigns become a budget proposal for later — with evidence attached, as in the budget prompts. Conflicts are same-day fixes. Odd geo and scheduling go on the list as questions, because sometimes the odd setting is deliberate and undocumented.
Minutes 25–30: the verdict
Close by making the agent commit to priorities:
Based on everything above, give me the audit verdict for account
<account>: the five most valuable actions, ranked by expected impact, each
with the evidence from this conversation and whether it's a read, a
zero-spend write, or a spend change. Then note anything you could NOT see
that I should check in the UI.
What the answer means. The ranking forces trade-offs, the evidence column keeps it honest, and the read/zero-spend/spend-change label maps each action to how much approval friction it deserves. The "could not see" note matters too: parts of Performance Max and anything about landing pages are outside any API's view, and a good audit says so.
What to do. Save the verdict — it is your baseline. Rerun this script in a month and diff the two.
Three rules for reading whatever it finds
The prompts are fixed; the judgement is yours, and three rules keep it honest. Volume before percentages — a 40% conversion-rate drop on nine clicks is noise wearing a costume; ask for the absolute numbers whenever a percentage looks dramatic. Concentration beats totals — $250 of zero-conversion spend spread across two hundred keywords is the cost of doing search; the same amount in three search terms is a finding. Recency cuts both ways — anything that changed inside the last two weeks (new campaign, new bidding, a sale ending) contaminates its own data, so check the change history from step five before treating a recent mover as a trend. When an answer survives all three rules, it goes on the work list; when it fails one, ask a sharper follow-up.
From audit to action
You now hold a ranked work list, produced in half an hour, without granting a single approval. The natural next steps, in order: add the negatives (zero-spend write, same day), fix disapprovals, draft the missing RSAs, then take the budget proposal with evidence. That sequence — reads, then harmless writes, then spend changes — is the whole first week playbook in miniature.
If you want to run this script and nothing is connected yet: Start free — add the connector URL to your client and sign in with Google, and the first prompt works immediately. Seven days on the full Pro plan for a new workspace, no card, and this audit is the best first use of minute one.