Anomaly detection is the automatic flagging of a data point that falls outside the range you would expect from past patterns. In marketing it is the alert that tells you enquiries halved yesterday, ad spend doubled overnight or sessions from one country spiked, long before a monthly report would.
How anomaly detection works
Every method compares what happened with what was expected. They differ only in how the expectation is worked out:
- Fixed thresholds: you set a rule, such as an alert when daily key events fall below a level you choose. Simple, but blind to normal weekly and seasonal swings.
- Comparison with matching periods: today is compared with the same weekday over the past few weeks, which allows for the usual weekday pattern.
- Statistical models: the tool learns the trend and seasonal pattern from history and draws an expected range around each day. Anything outside the range is flagged, usually with a confidence level.
GA4’s Insights feature scans a property for unusual changes automatically, and custom insights let you set your own conditions and receive an email when they are met. Google Ads offers explanations for some significant swings in campaign performance. For more control, an analyst can build models on the raw GA4 data exported to BigQuery, which includes anomaly detection functions for time series.
Why it matters
The expensive problems in digital marketing are usually silent. A Tag Manager publish removes the conversion tag from the thank-you page, and Google Ads keeps spending while its bidding learns from the wrong signal. A plugin update breaks a contact form on mobile. A checkout change causes double firing, and revenue appears to double overnight. Each can run for weeks if the only check is a monthly report, and on a paid account those weeks are spent bidding on broken data.
Good detection also separates real change from noise. A UK retailer whose sales dip on a bank holiday Monday does not need an alert; one whose sales dip on an ordinary Tuesday does. Expected ranges that allow for seasonality make that distinction for you.
Common mistakes
- Alerting on too many metrics with tight thresholds. Within a month the alerts are ignored, including the one that mattered.
- Watching only for drops. Spikes are often bot traffic, spam or duplicated tags, and they distort conversion rates and automated bidding as badly as a fall.
- Ignoring the calendar, so Christmas, Easter and the summer holidays all set off false alarms.
- Treating the flag as the finding. An alert says something changed; working out what changed is still a human job.
- Sending alerts to a shared inbox that nobody owns.
How to act on it
Choose a handful of metrics that cost money when they break: key events or orders, revenue, paid spend, cost per conversion and sessions from your main channels. Set alerts on those, using same-weekday comparisons or GA4 custom insights, and give each alert a named owner.
Be careful with small numbers. A business that averages four enquiries a day will see days with none purely by chance, so a daily alert on that metric will cry wolf. Widen the range, or alert on a rolling seven-day total instead, and keep daily alerts for metrics with enough volume to be stable.
When an alert fires, check your change log and annotations first, then the tracking, then the market. Once you find the cause, record it so the next person does not repeat the investigation.
When I run paid channels in performance marketing, daily checks on conversion volume and spend are part of the routine, because a tracking fault left for a fortnight costs more than most optimisation can win back.