Google Ads Anomaly Detection — A Practical Guide

Why anomaly detection matters

The single largest source of PPC budget waste is delay. A negative-keyword conflict, a runaway PMax asset group, a campaign budget cap that triggered after-hours, or a Google policy disapproval can all double your daily spend or zero out your conversions overnight — and most teams discover it 5-10 days later in the weekly review.

Anomaly detection compresses that detection window from days to hours. Done right, it pays for itself within the first month it catches anything. Done poorly, it floods you with false alerts and you turn it off.

The 4 metrics worth alerting on

1. Spend (cost_micros) — abrupt up moves usually mean budget cap reset, new keyword unlock, or competitive bid surge. Down moves usually mean disapprovals.
2. Conversions — most direct revenue signal. Sudden drops mean tracking broke, landing page broke, or a key campaign paused.
3. Conversion value — when conversions look fine but value is off, attribution is shifting (low-value lead form replacing high-value purchase).
4. CTR — high-volume metric, useful early-warning for ad disapproval (CTR drops fast when one variant in a campaign goes dark).

You can add more (impressions, average CPC, conversion rate) but the four above catch ~90% of accidents.

The statistical method that actually works

The naive approach — flag when today is N standard deviations above the rolling 7-day mean — fails in two ways: 1) seasonality (Mondays differ from Saturdays), 2) outlier sensitivity (a single bad day inflates SD).

The better method, what we use inside MarqOps, is median + Median Absolute Deviation (MAD) over a 14-day window:

1. Pull the last 14 days of daily values for the metric.
2. Compute median(values).
3. Compute MAD = median(|value - median|).
4. Compute the robust z-score for today: z = (today - median) / (1.4826 × MAD).
5. Flag if |z| ≥ 2.5 (medium severity) or |z| ≥ 4 (high).

MAD is robust to outliers — one bad day doesn’t inflate the baseline. The 1.4826 multiplier scales MAD to be comparable to standard deviation under normality.

Thresholds that don’t cry wolf

Pure z-score thresholds aren’t enough — small absolute changes can have huge z-scores when MAD is tiny. Combine with a minimum-percent-change floor:

• Spend: z ≥ 2.5 AND change ≥ 25%
• Conversions: z ≥ 2.5 AND change ≥ 30%
• CTR: z ≥ 3 AND change ≥ 20%
• Conversion value: z ≥ 2.5 AND change ≥ 30%

Add a baseline-floor: ignore campaigns with median spend <$5/day or median conversions <1.5/day — anomalies on tiny baselines aren’t actionable.

Handling false positives

You will get false positives. The way to keep alerting useful:

• Tier the severity. Daily digest for medium; same-day push for high.
• Decay good-direction alerts. If CTR went up dramatically, that’s a "good anomaly" — log it, don’t page anyone.
• Annotate known events. Use timeline annotations ("launched promo on the 12th") so future regression analyses can pattern-match. MarqOps’s annotation feature handles this.
• Acknowledge / dismiss / resolve. Don’t just leave a noisy backlog of "open" anomalies; build status workflow.

Tooling: build vs buy

Build if: you have a data engineer, you want to integrate with internal alerting (PagerDuty), or your account is >$500K/mo and you want full control.

Buy if: you want results in <1 day, you don’t want to maintain stats code, you want the alerts wired into Slack/email/in-app notifications without doing the plumbing.

The best-in-category tools all ship anomaly detection now: Optmyzr, Opteo, Adalysis, MarqOps. The differences are in severity tiering, false-positive handling, and integration depth (Slack, webhooks, email digest customization).

For a no-cost start: the free MarqOps tier includes daily anomaly detection on connected Google Ads accounts. Or run the free audit to see your current account snapshot.