# What is anomaly detection in ads?

> Anomaly detection flags ad numbers that break their own normal pattern, instead of a fixed line you set. How it differs from rule-based alerts, with an example.
- **Author**: Marxx
- **Published**: 2026-11-07
- **Category**: Alerts
- **URL**: https://marxx.ai/library-glossary/anomaly-detection

---

Anomaly detection in ads means a tool learns what normal looks like for a number, such as daily spend, and flags it when today breaks that pattern. A rule-based alert does something simpler: it fires when a number crosses a line you chose.

## Rules vs anomaly detection

| | Rule-based alert | Anomaly detection |
|---|---|---|
| You set | A fixed threshold, such as spend under INR 15,000 | Little or nothing, the tool learns the range |
| Knows about weekends and seasons | No, unless you write separate rules | Yes, if it has enough history |
| Easy to explain | Yes: "it crossed INR 900" | Harder: "unusual for a Tuesday" |
| Needs history | No | Weeks of steady data |
| Good for | Costs with a clear target | Spend, clicks and other numbers with a pattern |

## The formula

Anomaly score = (today's value − usual value) / usual day-to-day spread

The usual value is the average for similar days. The spread is how far a normal day strays from that average. A score above about 3 means "far outside normal".

## A worked example

Say an account usually spends about INR 20,000 on a weekday, and normal weekdays stray by about INR 1,500 either way.

1. On Tuesday it spends INR 26,000.
2. Score = (INR 26,000 − INR 20,000) / INR 1,500 = 4.
3. That is far outside normal, so anomaly detection flags it.

Now Sunday. The account usually spends INR 9,000 on Sundays. A rule that says "alert if spend falls under INR 15,000" fires every Sunday for nothing. Anomaly detection compares Sunday with past Sundays and stays quiet.

## Which one to use

Start with rules. They are clear, and you know exactly why each one fired. Add anomaly detection for numbers with strong weekly or seasonal patterns, such as spend or [CTR](/glossary/ctr), where a fixed line causes alert fatigue. For costs such as [CPA](/glossary/cpa), a rule tied to your target is usually enough.

Tools label these differently. Opteo, for example, lists "Unexpected Changes" among its Slack alerts. We make Marxx, whose alerts are rules you set on metrics, with a severity; see [Meta ads alerts in Slack](/guides/how-to-set-up-meta-ads-alerts-in-slack).

## FAQ

### Is anomaly detection the same as AI?

Not always. Many anomaly checks are plain statistics on past numbers. Some tools use machine learning, but the idea is the same.

### Can anomaly detection miss real problems?

Yes. If a problem grows slowly, the tool may learn it as the new normal. Keep a rule on your most important cost as a backstop.

---
- [More Alerts articles](https://marxx.ai/library-glossary/category/alerts)
- [All articles](https://marxx.ai/library-glossary)