| name | anomaly-detector |
| description | Catch spend spikes, CTR or conversion-rate drops, budget runaway, and metric outliers across campaigns, and set sane alert thresholds. Use when a user asks why spend or performance changed suddenly, wants monitoring or alerts, or something looks off in the numbers. |
Anomaly Detector
Find the sudden change before it burns a budget. Reads are free; setting alerts or pausing is a change and waits for approval.
Confirm the account: list_connected_accounts.
1. Pull a baseline and today
Read a stable window and the recent period per platform: pull_<platform>_ads_performance(days=...), or run_gaql_query for Google detail. You need enough history to know what normal looks like before you can call something abnormal.
2. Find real outliers, not noise
- Spend spike / budget runaway: daily spend jumping well above the recent average, or pacing to exhaust the budget far too early. This is the one to catch fastest.
- CTR or conversion-rate drop: a sharp fall against the rolling average signals a broken creative, landing page, or tracking tag, not just a bad day.
- CPA / CPM jump: rising cost per result or per impression points to auction pressure, audience saturation, or fatigue.
- Judge against variation, not a single day. A metric one bad day inside normal swing is not an anomaly; a sustained multi-day move outside the usual range is. State the window and the size of the move.
3. Diagnose before you alarm
Rule out the mundane first: a tracking outage reads as a conversion cliff; a new campaign launch reads as a spend spike; a weekend reads as a volume drop. Confirm the cause before recommending action.
4. Act
- Set forward-looking guardrails with
set_spend_alert so the next runaway trips an alarm early.
- For a confirmed problem, recommend the fix (pause, budget cut, creative swap) and take it through the relevant skill on approval. Do not pause on one noisy data point; see pre-pause-analysis.
Rules
- Distinguish signal from noise; every flag names the window and the magnitude.
- Rule out tracking, launches, and seasonality before calling a metric broken.
- Report anomalies ranked by budget impact, with the recommended action and who approves it.
