Start free
Features › Google Ads Anomaly Detection › Google Ads spend anomaly

Google Ads spend anomaly detection: the night spend does something it never does.

Spend is the metric everyone watches and the one that still surprises them. This page is about the detection side — how AdCortex decides a day's spend is unusual for that entity — and where it hands off to budget pacing and the reviewed budget change. It is not the budget-alerts page, which covers alert delivery, or the budget-pacing report, which covers the plan.

Last verified: September 10, 2026 App version: 50.1 Canonical page for this topic: Google Ads Budget Alerts

Spike, drift, and silence

PatternCaught byCadenceUsual cause
Spike — one day well above normalregime_spike::cost (3× same-regime median); Layer 2 cost modelDailyBudget raised, bid target loosened, a new campaign enabled, a shared budget's other campaigns paused
Drift — steadily higher, no single bad dayWeekly and monthly checks comparing the closed period against prior periodsWeekly (Mon), Monthly (1st)CPC inflation, gradual match-type or PMax expansion, seasonal competition
Silence — always-on spend goes to zeroalways_on_zero::costDailyCampaign paused, ad disapproved, payment failed, budget exhausted early, end date reached

Same-regime comparison

Most accounts have a weekly shape. A B2B account that spends 40% less at weekends would trip an all-days rule every Monday and never at the weekend. The spike rule therefore compares today with the median of days in the same regime — weekday against weekday, weekend against weekend — and fires at 3× that median. On weekly and monthly cadences it records SKIP with a reason, because every weekly observation covers the same mix of days and the comparison would be meaningless.

The cost model

Layer 2 fits a generalised linear model to the entity's spend history and reports the day's expected spend and how improbable the actual figure is. Like the other Layer 2 models it is direction-blind: a day well under expectation is flagged as firmly as one well over, and Layer 3 decides whether under-spend is a problem — usually yes, when it means lost delivery, and sometimes no, when the manager cut the budget on purpose and the change history says so.

Where pacing takes over

Anomaly detection asks whether today is unusual. Budget Pacing asks whether the month is on plan. They are different questions and both are needed: a month can be pacing to 130% of target with every individual day looking normal, and a single spike can land in a month that is otherwise under-delivering. Budget Pacing is a free report: daily spend against the monthly target set for the account, with a projection of where the month lands at the current rate.

When a correction is warranted, the budget optimizer proposes campaign budgets, percentage adjustments shift a set of campaigns together, and the PMax overview lets a campaign's daily budget be edited in place. All of it stages to Checkout. Shared budgets are labelled and queue as one change per budget pool, so two campaigns on one budget cannot be corrected twice.

  • Budget allocation recommendations need at least five conversions of data; below that they report "not checked" rather than guessing.
  • The confirmation before a push names the account and the total daily budget delta in dollars — the number that actually changes tomorrow's spend.

A worked example: the Thursday that cost three Thursdays

A campaign's weekday spend sits between $380 and $460; its weekday median over the regime window is $412. On a Thursday it spends $1,340. The same-regime rule fires — $1,340 is 3.25× the weekday median — at Medium. The Layer 2 cost model, fitted to the same history, reports an expected $418 and a probability score deep in the tail. Layer 3 receives both, the 28-day daily table, and the change history, which shows the campaign's daily budget raised from $450 to $1,500 that morning. The judgement scores it 4 and the short reason says the budget was raised; the alert confirms a deliberate change landed and that clicks rose in proportion, so nothing else moved.

The same numbers with no change-history entry produce a different reason — "spend tripled with no budget change recorded; check for a bid-strategy target change or a shared budget whose other campaigns stopped spending" — and a different first action for the reader.

Shared budgets, PMax budgets, and what a correction looks like

Two campaigns on one shared budget can produce a spend anomaly on one of them when the other stops spending — the pool is unchanged, the split moved. The alert reports the campaign that spiked; the change history and the budget history, which the AI can request, show whether the pool itself changed. When the fix is a budget edit, the PMax overview and Budget Pacing stage it to Checkout, shared budgets are labelled and queue as one change per pool, and the confirmation names the total daily budget delta in dollars before a person pushes. A change queued against a budget that has since been edited in Google Ads is what the reviewer is there to catch.