What is anomaly detection in marketing?
Anomaly detection in marketing uses statistical models to spot when a campaign or business metric moves outside its normal pattern, like a sudden drop in conversions, a spike in cost, or an unexpected surge in clicks. Instead of someone noticing the change in next week's report, the system flags it as soon as the data shows it.
The key word is "normal." Every metric has its own rhythm. Anomaly detection learns that rhythm and flags departures from it, rather than reacting to every movement.
How does anomaly detection work?
- Learn the baseline. A model studies each metric's history, including day-of-week patterns, trends, and seasonality.
- Predict the expected range. For each new data point, the model estimates where the metric should fall.
- Flag the outliers. Values well outside that range are flagged as anomalies, in either direction.
- Route to people. Alerts go to the team in the tools they already use, with enough context to triage.
What is an example of anomaly detection?
A weekend drop nobody was watching
A campaign stops converting on Friday night after a landing page change. Weekends are often quieter, so a fixed "down 20%" alert would fire every Saturday and get ignored.
An anomaly model knows what a normal Saturday looks like for this campaign, flags that conversions are far below even that, and the team fixes the page before Monday.
Anomaly detection vs. threshold alerts
| Fixed threshold alerts | Model-based anomaly detection | |
|---|---|---|
| How it decides | A fixed rule, such as "spend fell 20%" | Compared with the metric's learned normal pattern |
| Handles weekday swings and seasonality | No | Yes |
| Noise | Often high, so alerts get ignored | Lower, focused on what is unusual |
| Best for | Hard limits like a budget cap | Spotting unexpected shifts in performance |
Many teams use both: rules for hard limits, models for everything that's hard to predict.
Why does anomaly detection matter?
The window to fix a campaign problem, or to capitalize on a breakout, is usually short. A feed failure or broken page can affect results within hours, and nobody can watch every account all day.
- Protect spend. Catch broken campaigns before they waste budget.
- Catch opportunities. Unusual gains are flagged too, so teams can lean in.
- Less manual monitoring. People respond to signals instead of hunting for them.
- Check the data first. Many anomalies are data issues, such as a stale feed or a renamed field, so context matters.
How does Alli do this?
Alli Alerts is marketing anomaly detection and campaign monitoring for every connected channel. Alli Anomaly Detection uses models built by PMG's data science team that judge each metric against its own historical pattern, and flags both unusual drops and unusual gains.
It runs alongside Custom Alerts for budget caps, pacing, feed health, and landing page checks, and delivers alerts into Slack. Teams can check feed health in Datasources, investigate in Data Explorer, or start a follow-up through Generative Workflows. The models read the governed data but don't change it.
Explore Alli AlertsAnomaly detection FAQ
What is an anomaly in marketing data?
A value that falls well outside a metric's normal pattern, such as a sudden drop in conversions, a spike in cost, or an unexpected surge in clicks.
How is anomaly detection different from alerts?
Threshold alerts fire on fixed rules. Anomaly detection compares each value with the metric's learned pattern, so it accounts for normal swings and produces less noise.
Does anomaly detection only find problems?
No. It flags unusual gains as well as drops, so teams can spot breakouts and opportunities early.
What should you check first when an anomaly fires?
Check the data: is the feed current, and did a platform change a field? If the data is healthy, look at the campaign, landing pages, and tracking.
Last updated ‹ Back to the glossary