AI Anomaly Detection for Google Ads —
Know Before It Costs You

What is anomaly detection — and why does it matter?

Anomaly detection is the automated process of finding data points that don't fit an account's normal pattern. Every Google Ads account has a rhythm: weekday traffic differs from weekends, CPCs drift with auction pressure, conversions ebb and flow with seasonality. An anomaly is a day that breaks that rhythm — spend that doubles with nothing to show for it, a conversion tag that silently stops firing, an impression share that falls off a cliff because a competitor raised bids overnight.

The reason it matters is simple: the cost of a problem in paid media compounds by the hour. A broken conversion action doesn't just lose tracking — it starves Smart Bidding of signals, and the algorithm starts optimizing toward the wrong outcome. A disapproved ad in your top campaign doesn't just pause one asset — it reroutes budget into weaker ads. The longer an issue runs undetected, the more expensive it becomes and the longer the recovery takes.

Most advertisers find out about these problems one of three ways:

• Manually — someone happens to look at the right report on the right day. This doesn't scale past a handful of accounts.
• Retroactively — the month-end report shows the damage after the budget is already spent.
• From the client — the worst way. "Why did our leads stop last Tuesday?" is a conversation no agency wants to have.

For exactly what runs each night, layer by layer, see how Google Ads anomaly detection works.

Statistical outliers, filtered by AI, grounded in your change log

Pure statistics catch too much. Pure human review catches too little, too late. AdCortex layers the two: deterministic rules and per-metric statistical models do the tireless nightly scanning that no human can, and an AI judgment layer applies the context a statistician can't — so what reaches your inbox is only the identifiers that actually deserve attention. Each layer only sees what the one before it produced.

Why human judgment still matters — and how we encode it

A model can tell you Tuesday's spend was unusual. It cannot tell you that Tuesday was Black Friday, that you launched a new campaign Monday night, or that the "conversion collapse" is actually a known tag migration. That contextual reasoning has always been the human's job — which is exactly why pure statistical alerting tools drown users in false alarms until they stop reading them.

AdCortex encodes that judgment into the pipeline. The AI layer sees the statistical evidence and the account's change log side by side, the same way you would, and asks: is this surprising given what we know changed? Anomalies explained by deliberate changes are named as such; anomalies with no explanation get escalated with the context attached. You stay in the loop for the decisions that matter — nothing an alert says is ever applied to the account without a person — and the system makes sure the only things reaching you are worth your attention. Every mechanism that keeps alerts quiet →

Account averages hide problems. Segments expose them.

The most dangerous anomalies are the ones invisible at the account level. If one campaign's conversion rate collapses while another surges, the account-wide average barely moves — the dashboard looks green while a real problem burns underneath. This is Simpson's paradox in action, and it's why detection that only watches account totals misses the failures that matter most.

AdCortex runs detection at three grains — the account, every campaign, and every ad group — with the rules tuned to each: ad groups get the median-based rules because their history is short, while account and campaign grain get the statistical models too. A campaign under 4% of the account's 28-day spend is folded into the account's judgement rather than alerted on its own, so a two-conversion campaign cannot page you. When an alert fires, it names the exact identifier that triggered it, so you're never left diffing reports to figure out where the problem lives — and the reports below carry the diagnosis into the segment that moved. For a whole MCC, the portfolio view shows whether it is one client or everyone: monitoring multiple accounts →

Every segment draws from the same nightly data pipeline that powers AdCortex reporting, so an alert is never based on stale or sampled data — it's the same finalized numbers you'd see in the report the next morning.

Quantifying outliers: a model per metric, robust baselines, and significance

"That looks high" is not a detection strategy. Every anomaly AdCortex surfaces is backed by explicit statistical evidence — here's what's actually running under the hood, and why each model earns its place. A z-score on a rolling mean is the obvious tool and the wrong one for ads data: counts are over-dispersed, rates depend on their denominators, costs are skewed, and a rolling mean is dragged by exactly the outlier you are trying to catch.

Rates: beta-binomial, because the denominator sets the certainty

CTR and conversion rate are successes out of trials. One click from ten impressions and a hundred from a thousand are the same rate with wildly different certainty, and a rate model has to know that. The beta-binomial asks how surprising today's rate is given today's denominator:

Conversions-out-of-clicks is the same model as clicks-out-of-impressions with different columns, which is why one block covers both rates.

Counts: negative-binomial with an exposure offset

Clicks, impressions, and conversions are counts whose variance grows faster than their mean — Poisson noise plus the day-to-day variation of an auction. The negative-binomial captures that over-dispersion, and an exposure offset means expected conversions depend on the clicks actually bought that day:

Four conversions from 400 clicks on a campaign that usually gets four from 100 is a drop. A flat average calls it normal; the count model does not. How a conversion drop is caught the same night →

Costs and values: lognormal deviation, centred on the median

CPC and conversion value are positive, right-skewed, and move multiplicatively — a 2× day is as likely as a 0.5× day, which no symmetric test respects. The lognormal model works in log space and centres on the median, scaled by median absolute deviation:

Median and MAD are what make the baseline robust: with a mean and standard deviation, one day of runaway spend widens the band enough that the next runaway day looks normal — the detector is blinded by the incident it should have caught. Spend anomalies, spikes and silence →

Ratios: decomposition, so the alert names which half moved

Cost per conversion and ROAS are ratios of two noisy quantities, undefined when the denominator hits zero — exactly when you most need the alert. Modelling them directly produces a detector that fires on every low-conversion day and explains nothing. AdCortex models the numerator and denominator separately and reports the ratio with the side that drove it:

That is conversions falling two-thirds while spend held — a tag, not a bid problem — and it is the sentence the email prints. Why did my CPA spike? →

Statistical significance ≠ practical significance

Here's the part most tools get wrong: a statistically significant change is not automatically an important one. A campaign spending $3/day can triple its CPC with total damage of a few dollars — statistically extreme, practically irrelevant. The statistical layer is deliberately high-recall and direction-blind; the decision about materiality is made downstream, with evidence:

• Two gates before the AI. An entity is judged only if a rule or model failed, and a campaign under 4% of the account's 28-day spend folds into the account rather than being judged alone.
• Direction is handled explicitly. A conversion rate six times better than expected is detected and then held — the judgement is told that improvements are not incidents. On the live model it scored 1 of 5.
• A check that cannot run says SKIP, never PASS. Volume floors and thin history are reported as coverage gaps, not counted as clean.
• The AI applies the practical test. Even a statistically airtight anomaly is weighed against the 28-day history and the change log before it earns a place in your inbox — and the send threshold of 4 is where its decision is written down, not a dial.