How anomaly detection works
An anomaly detector learns what a measure normally does, then raises a flag when the latest value sits too far from that pattern. The history it learns from includes the rhythms of the business: quieter weekends, busier evenings, a festival period that lifts everything. A value is only unusual relative to what that day and that hour usually look like.
Detection ranges from the very simple to the statistical. A threshold rule fires when spend passes a fixed ceiling or when a conversion count drops to nothing. A trained model instead estimates a plausible range for each period and flags anything outside it. GA4 and Google Ads include their own detection; Looker Studio, spreadsheets and simple scheduled scripts can all carry the same logic for a small business.
What gets watched matters more than the sophistication of the method. Conversions per day, spend per campaign, form submissions, page load failures and the ratio between clicks and sessions all break in ways that a person will not notice quickly by eye.
Why anomaly detection matters
Most measurement faults are silent. A tag removed during a website release, a form that stops sending, a payment step that fails on one browser — none of these announce themselves. They show up as a gentle decline that gets blamed on the market, sometimes weeks later, after budget has been spent buying traffic that could not convert.
The same alerting catches the opposite problem. Runaway spend, a duplicated tag inflating conversions, or a competitor entering the auction all move a number sharply, and the sooner that is seen the cheaper it is to fix.
Where anomaly detection goes wrong
Too many alerts is the usual failure. Set the sensitivity high and every ordinary quiet Saturday triggers a message, until people mute the channel and miss the one alert that mattered. Small accounts suffer most, because low daily volumes swing naturally and look dramatic beside a typical day.
The second failure is alerting without a next step. A flag saying conversions are low is not useful unless someone knows who checks it, what they check first, and what counts as resolved. The third is forgetting seasonality, so a genuinely predictable holiday drop is investigated every year.
How to act on it
Start narrow. Pick the handful of measures whose failure would actually cost money, alert on those only, and set the sensitivity so that a normal week is quiet. Route the alerts to a person, not a shared inbox nobody owns.
Pair every alert with a short checklist of what to inspect: recent website releases, tag firing, consent changes, campaign status. Annotate known events so predictable swings stop being flagged. And treat detection as a complement to data validation rules, which stop bad values entering in the first place. Getting the underlying measurement dependable enough to alert on is the point of a solid analytics and tracking setup.