How to audit an ad account with an AI agent
A structured pass over one ad account: settings, structure, spend concentration, conversion wiring and recent changes, ending in a ranked list of findings.
IntentTurn an unfamiliar ad account into a ranked list of problems and confirmations, each backed by numbers.
- 8
- Steps
- 12
- Tools
- 5
- Rules
- 4
- Failure modes
When to use it
Use it when
- Taking over an account from another team or agency.
- A periodic health check, quarterly or before a budget increase.
- Before proposing any restructure: the audit is the evidence.
Do not use it when
- Fewer than 30 days of history in the account: most checks below need a baseline.
- During a declared incident. Use campaign-anomaly-investigation first; an audit is a survey, not a search party.
What it needs first
Required
Optional
Preconditions
- The account is ticked in the dashboard, confirmed via list_accounts.
- At least 30 days of spend history.
- The account currency and timezone are read before any figure is compared.
Inputs
accountrequired- The ad account to audit.
window- Analysis window.Default: last 90 days.
target_cpa_or_roas- The efficiency target the account is judged against. Ask rather than assume.
Procedure
In order. Every tool named is one the gateway ships, and links to its reference.
Read account settings, currency, timezone and the account tree.
Map structure: campaigns, their types, statuses and budgets.
Read performance by campaign over the window, then rank by spend.
Spend concentration first: findings on 2% of spend are footnotes, not findings.
Check conversion wiring: actions, their counting and recent volumes.
Read search terms and keyword performance for waste and gaps.
Read the change history for the window: who changed what.
Google keeps 30 days. Say so when the window is longer.
Collect the platform's own recommendations, as claims to verify rather than accept.
If GA4 is connected, compare platform conversions with GA4 key events over the same window.
Decision rules
The observable condition, what it lets you conclude, and what takes the conclusion back.
When
One campaign holds more than 60% of account spend
Conclude
Flag concentration: the account's result is that campaign's result. Audit it at ad group depth.
Unless
The account is deliberately single-product.
When
A conversion action counted in bids recorded zero conversions in 30 days while spend ran
Conclude
Flag as probable broken wiring or a dead action still steering bids.
When
Search terms with more than 20 clicks and zero conversions sum to over 5% of spend
Conclude
Flag as waste with the exact term list. Hand to google-search-waste for the full pass.
When
GA4 key events differ from platform conversions by more than 25% over the same window
Conclude
Flag a measurement gap, not a performance change. Hand to tracking-audit.
Unless
Attribution windows or counting methods visibly differ; name them instead.
When
Fewer than 100 clicks in the window on the audited scope
Conclude
Do not grade efficiency. Report volumes and structure only, and say the sample is too small.
Evidence every conclusion must carry
- Every finding names the campaign or term, the metric, the window and the account it came from.
- Every percentage names its denominator.
- Recommendations from the platform are marked as the platform's, not the audit's.
Where the agent stops
- The audit changes nothing. Every fix it proposes is a proposal.
- Pausing anything found here goes through Safe Writes: preview, human confirmation, then apply.
How it goes wrong quietly
The cases where the analysis is wrong and still looks right. Read them before trusting a number.
- The last 72 hours under-report conversions on every platform: never grade recent days as decline.
- Currency mixing across accounts silently corrupts every ranking; read the currency first.
- Change history beyond 30 days is not absence of change, it is absence of records.
- A restructure inside the window resets learning phases; performance dips there are expected, not findings.
What the answer contains
- 01Account snapshot
- Spend, structure, currency, timezone, window.
- 02Findings, ranked
- Each with metric, evidence, affected spend share.
- 03Confirmations
- What was checked and found healthy: absence of findings is a result.
- 04Proposed next actions
- Each mapped to a skill or a Safe Write, none applied.