Your best ad died two weeks before you noticed

We watched a creative die once, in the data, in slow motion. It was the account's best performer, a video ad that had carried a DTC brand's prospecting for most of a quarter, and on an engagement to clean up their marketing measurement we reconstructed its whole life from the daily logs. Day 1 to day 38: cost per acquisition drifting gently around target. Day 39: a bend. Not a crash, a bend, the CPA starting to climb a few percent a day, compounding quietly under the weekly averages. The team caught it on day 54, in a Monday report, by which point the last two weeks of that ad's spend had bought conversions at nearly double the price of its prime.

The two weeks cost about 40,000 dollars over what the same budget would have earned before the bend. Nobody had done anything wrong by the book. The book just checks weekly, and averages are excellent painkillers.

Creative fatigue is not a cliff you fall off. It is a slope that starts on a specific day, and a weekly dashboard meets it three reports late.

Fatigue hides inside noise. Daily performance jumps around for a dozen honest reasons, day of week, auction pressure, a competitor's launch, so a human watching the chart waits, sensibly, for confirmation. Confirmation takes days. Then the discussion takes days. The platforms' own frequency metrics do not save you either: fatigue is not purely about how often one person saw the ad, it is the audience's aggregate boredom, and it arrives on a different schedule for every creative, every audience, every placement.

The result, across every account we have audited, is the same asymmetry: winners get killed too late. The spend between "started decaying" and "got paused" is a tax paid on every single creative that ever worked, which means the better your creative team, the more this tax collects.

(Losers, by contrast, get killed with impressive speed. Nothing motivates a Monday meeting like an ad that was bad from birth.)

The statistical tool for "when did this series change behaviour" is changepoint detection, and it is a different question from "is today bad." A changepoint model watches each creative's daily performance against its own established behaviour, accounting for day-of-week rhythm and normal noise, and estimates the probability that the underlying process has shifted, that the bend has happened, even while individual days still look defensible. It is the difference between noticing a fever and noticing the moment the temperature started rising.

Wired to the ad platforms, it becomes an early-warning channel: the day the model's belief in "this creative changed" crosses a threshold, the alert lands with the evidence, here is the curve, here is the bend, here is the daily cost of leaving it. Paired with a simple survival view of how long creatives in this account historically last by format and audience, it also answers the planning question nobody can currently answer: how many new creatives do we need in the pipeline per month just to replace natural death?

The alert on day 40 instead of day 54 is worth two weeks of your best ad's budget, on every winner you will ever run.

We built exactly this monitor after that reconstruction, and the interesting cultural effect was not the paused ads. It was that creative refresh stopped being an argument. The debate about whether the hero ad was "still fine" ended, because the bend had a date and a cost attached, and the brief for its replacement started that day.

It needs your daily per-creative history, a year if you have it, and nothing else exotic. It does not need a new attribution model, a data warehouse migration, or anyone to change how they buy media. It is a listener. At TensorLabs we fit it to the account's own noise first, because an alert system that cries wolf gets muted by week three, and a muted alert is worse than none.

Reply with roughly how many active creatives you run and how you currently decide when one is done, and we will tell you honestly whether a changepoint monitor would pay for itself on your account size, with the arithmetic shown. For some accounts the answer is no, and we will say so.