Detecting Creative Fatigue Before It Shows Up in CPI

By the time cost per install rises on a fatiguing creative, the fatigue has usually been visible in the data for a week or two already. Creative fatigue detection built around CPI as the trigger is, structurally, a lagging system: CPI moves after the auction has already repriced the impression, which happens after engagement on the creative has already fallen, which happens after the audience has already seen it enough times to tune it out. A UA budget planning template can set the guardrails, as we set out in "A UA Budget Planning Template Built for a Flat Market," but the team still needs to see fatigue coming before it hits the number the budget is protecting.

Why CPI is the wrong first signal

CPI is a downstream metric. It reflects an auction's collective judgement about an impression's value, and that judgement changes only after enough advertisers have seen enough evidence that a given creative or audience segment is worth less than it was. By the time that repricing shows up in a dashboard, the behavioural signals that caused it have usually been declining for one to two weeks. You end up reacting to something that started a fortnight ago, and paying for the lag in between.

The signals worth watching sit upstream. Hook rate, meaning the share of viewers who watch past the first two to three seconds. Completion rate on video. Click-through rate segmented by frequency band rather than blended across the whole audience. Each of them moves before CPI does, because each measures audience response directly rather than the auction's downstream reaction to it.

Building a frequency-banded view

The single most useful change most teams can make is splitting performance by frequency rather than looking at a blended average. A creative shown to a user for the first time and the same creative shown to that user for the eighth time are not the same event, and blending them hides fatigue inside a healthy-looking top-line number for as long as new audience keeps entering the pool.

Segment reporting into frequency bands: first exposure; two to four exposures; five to eight; nine or more. Then track hook rate and click-through rate within each band over time. A creative holding steady on impressions one through four while collapsing from exposure five onward is fatiguing among the audience it has already reached, even if blended CPI still looks acceptable because fresh impressions are propping up the average.

A three-signal early warning system

Rather than building one composite fatigue score, which tends to obscure which input actually moved, track three signals separately and treat any two moving together as the trigger for review. The first: hook rate in the two-to-four exposure band dropping more than 15% from its seven-day average. The second: completion rate declining for three consecutive days rather than one, since single-day dips are usually noise. The third: the frequency at which cost per click starts climbing within a segment, tracked separately from blended CPI.

Any one of them alone can be a false alarm. A hook rate blip from an audience mix shift, a completion dip from a platform reporting glitch, a cost climb from a competitor's temporary bid push. Two moving the same way inside the same window points much harder at the creative itself than at the environment around it.

Frequency capping touches all three signals, so set it deliberately instead of leaving a platform default in place. A cap set too loose lets a creative accumulate exposures nine through eleven against a shrinking slice of the audience well before the team's dashboard would otherwise flag fatigue in that band, quietly dragging down blended engagement while the frequency-banded view for lower exposure counts still looks healthy. Too tight, and the cap starves a creative of the repeat exposure some formats need to convert, playables with a slower-to-land mechanic above all. The right cap depends on genre and format. Testing two or three cap levels against the three-signal framework beats adopting whatever a network happens to default to.

What a worked week looks like

Take a hypothetical hybrid-casual creative that has been live for three weeks. In week one, hook rate across all bands sits close to a strong 45%, completion rate at 30%, and blended CPI is stable. By the start of week three, hook rate in the five-to-eight exposure band has fallen to 36%, a drop of roughly 20% against its own seven-day average, and completion rate in that same band has declined for three straight days. Cost per click within that specific segment has started climbing too. Blended CPI across the whole campaign has moved only marginally, since fresh first-exposure impressions still hold the average up. Two of the three triggers have now fired together. That is the moment to rotate a fresh variant in against that segment specifically, days before blended CPI would have moved enough to prompt a review under a CPI-only approach.

What to do when the signal fires

Match the response to how much creative inventory sits behind the fatiguing concept. If three to five variants of the same hook have not run yet, rotate a fresh one in before pausing the fatiguing creative, so a competitor doesn't simply take the audience segment while a new concept goes into production. If the fatiguing creative carries a disproportionate share of spend and no variant sits ready, that is a portfolio risk worth flagging on its own: a single concept should rarely be load-bearing for more than 30 to 40% of a channel's spend, because its fatigue curve then dictates the whole channel's efficiency.

Set refresh cadence by category norms, not by a fixed calendar. Hyper-casual and hybrid-casual creative fatigues faster, often within two to three weeks, because audiences see more competing ad units in the same session context. Mid-core and 4X creative can hold for four to six weeks. Playables and longer-form video especially, where the format itself takes longer to feel repetitive. Build a refresh calendar around one fixed interval for every game type and you waste production capacity on formats nowhere near fatiguing, while leaving the faster-fatiguing formats exposed for far too long.

Making creative fatigue detection the default view

The practical change for most teams isn't more data. Hook rate, completion rate, frequency: most platforms report all of it already. What changes is which view opens first, so frequency-banded engagement becomes the dashboard someone sees rather than a report they have to remember to pull. A team that reviews blended CPI daily and frequency-banded hook rate weekly has the ordering backward. The frequency-banded view should catch the problem while there is still time to rotate a variant in, not after the auction has already told everyone else the impression is worth less.

Related archive reading

These articles provide related context and remain subject to their stated review status.

• Building a Creative Fatigue Detection Cadence That Works
• Creative Fatigue Leading Indicators to Watch Before CPI