We could not work out why campaigns started well and declined within weeks, despite no change to targeting or budget. We assumed it was auction dynamics.
It was creative fatigue, it was faster than we had assumed, and once we measured it properly the fix was straightforward.
What the curve looked like
We tracked performance by days since launch across a large number of campaigns.
The pattern was consistent. Strong initial performance, a plateau, then a steady decline that began earlier than anybody expected — in many cases within two weeks, and in high-frequency campaigns within one.
By four weeks, most executions were performing at a fraction of their initial rate.
The decline tracked frequency far more closely than it tracked elapsed time, which was the key finding. Campaigns with a small audience and high budget decayed fast; the same creative against a large audience lasted much longer.
Why it happens
Two mechanisms, and they compound.
The audience one. People who have seen an advertisement several times and not acted are unlikely to act on the fifth exposure, and are increasingly likely to be irritated. Response rate falls as the responsive portion of the audience is exhausted.
And the platform one. Delivery systems optimise toward predicted response. As response declines, the system either delivers less or reaches further into less relevant audiences to maintain volume, both of which look like performance decay.
The second mechanism means the decline can accelerate, because the system's response to falling performance makes performance fall faster.
The production mismatch
The insight that changed how we work.
Our production cycle was roughly six weeks from brief to live. Our creative was exhausted in three.
Which meant we were permanently running tired creative while waiting for the next batch, and treating the resulting poor performance as a media problem.
No amount of bidding or targeting adjustment fixes creative the audience is bored of.
What we changed
Volume of executions, principally. We moved from three or four per campaign to fifteen or twenty, produced in a single session from one setup.
That is the change generative tooling has genuinely enabled, and it is the most valuable thing it has done for us.
Rotation rather than replacement. Rather than running one execution until it dies, we run several concurrently and swap continuously, which extends the life of all of them because frequency per execution stays low.
Frequency caps set deliberately, which the platforms allow and which most people leave at default.
And a production calendar sized to the observed decay rate rather than to whatever the agency cycle happened to be.
What counts as a new execution
A question we had to resolve, because the answer determines how much production is required.
Our testing suggested that meaningful refresh requires a visibly different execution — different imagery, different opening, different structure.
Changing a headline while keeping the same image produced very little recovery. The visual is what the audience recognises and tires of.
Different crops of the same image performed as the same execution.
Which means the production requirement is real and cannot be met by variation of copy alone, which was our first and cheapest attempt.
The audience size lever
The other half of the equation and the one people forget.
Frequency is spend divided by audience size. Doubling the audience halves the frequency at the same spend, and extends creative life proportionally.
Which is an argument for broader targeting than the instinct suggests, and it aligns with what the platforms have been recommending for other reasons.
Very narrow targeting produces high frequency, fast fatigue and a constant production requirement, and the precision gained is frequently smaller than assumed.
What we monitor now
Frequency, daily, as a leading indicator rather than performance as a lagging one.
Performance by days since launch for each execution, which makes the decay visible before it becomes a problem.
And a rule that any execution past a frequency threshold is paused regardless of how it is performing, because by the time performance drops the audience irritation has already happened.
The creative testing consequence
One further implication that changed our process.
If creative decays within weeks, a testing programme that takes six weeks to declare a winner is declaring a winner that is already tired.
We shortened tests substantially, accepting less statistical confidence, on the basis that a fast approximate answer is more useful than a precise answer about something no longer running.
That is a genuine trade and it was the right one for this channel specifically. In channels where creative lasts longer, the opposite applies.
The new-audience exception
One case where the decay curve does not apply in the same way.
Campaigns targeting audiences that continuously refresh — new market entrants, people newly matching a life-stage signal, expanding geographies — see much slower creative fatigue, because the audience being exposed is not the same audience.
Which means the production requirement varies enormously by campaign type, and applying a single refresh cadence across everything wastes effort on some campaigns and starves others.
Measuring frequency per campaign rather than assuming a schedule is the practical answer.