How to investigate a sudden ad performance change with an AI agent
Spend stopped, CPA doubled, conversions vanished: a fixed investigation order that finds the cause instead of the first plausible story.
IntentName the cause of a sudden change, with evidence, in minutes, checking causes in the order of their base rates.
- 6
- Steps
- 9
- Tools
- 4
- Rules
- 3
- Failure modes
When to use it
Use it when
- Any 'why did X suddenly change' question.
- An alert fired and the dashboard offers no reason.
Do not use it when
- Slow drifts over weeks: use account-audit or creative-fatigue; anomaly hunting on a trend finds coincidences.
What it needs first
Required
Preconditions
- The anomaly is defined: metric, direction, date, scope, before any tool call.
Inputs
anomalyrequired- What changed, when, on what scope.
platformrequired- Where it happened.
Procedure
In order. Every tool named is one the gateway ships, and links to its reference.
Confirm the anomaly in the data: exact date, magnitude, scope. Rule out a reporting artefact first.
Check humans first: the change log around the date.
Most anomalies are edits. Attribute and date them before theorising.
Check the platform: delivery status, disapprovals, learning resets.
Check measurement: did tracking move rather than performance.
Check the market: CPM and auction pressure on the same window.
Stop at the first cause that explains magnitude and date, and say which checks were not needed.
Decision rules
The observable condition, what it lets you conclude, and what takes the conclusion back.
When
A logged change lands within 24h of the anomaly and touches the affected scope
Conclude
Primary suspect; verify magnitude fits before closing.
When
Conversions fell while clicks and spend held
Conclude
Investigate measurement before performance: the funnel rarely breaks that cleanly on its own.
When
CPM rose sharply across campaigns at once
Conclude
Market pressure, not account error; name the affected inventory.
When
The anomaly date is within the last 72 hours and the metric involves conversions
Conclude
Attribution lag is a candidate cause; re-check after the lag before escalating.
Evidence every conclusion must carry
- The confirmed anomaly: metric, dates, magnitude, scope, before any cause is discussed.
- The cause names its evidence: a change event, a delivery status, a diagnostic, never only a correlation.
Where the agent stops
- Reverting someone's change is a human decision, then a Safe Write.
- Policy appeals and billing issues route to the platform; the skill identifies, humans escalate.
How it goes wrong quietly
The cases where the analysis is wrong and still looks right. Read them before trusting a number.
- The first plausible story ends the search while the real cause survives: the order exists to prevent exactly this.
- Timezone mismatch between platforms makes 'the same day' off by one; align before matching events to effects.
- Google's change history stops at 30 days; older anomalies lose their best evidence source.
What the answer contains
- 01Anomaly confirmed
- Metric, date, magnitude, scope.
- 02Cause
- The finding with its evidence, and the checks that ruled alternatives out.
- 03Remedy
- Proposed fix, its owner, and the Safe Write if one applies.