Creative Fatigue Read Ad Creative

Creative Fatigue Read

Shared quality bar: ../references/output-standard.md. All numbers cited here live in ../references/thresholds.md.

Fatigue is the most over-used word in creative reporting. Most ads called tired are being served to a smaller audience, or into a more expensive auction, and pausing them fixes nothing.

Use this skill when

One named creative that previously performed is declining across consecutive weeks and the proposed action is to pause or refresh it. Also use it before approving a refresh budget, so the money goes to a creative whose decline was actually caused by the creative.

Required input

• A daily export for that one ad ID: date, impressions, reach, link clicks, spend. Reach is required for both windows and impressions cannot stand in for it. Daily reach summed across each window runs the frequency signals; where the platform reports deduplicated reach for the window, use that and say which of the two you used, since only the deduplicated figure supports the new-reach reading below.
• The reach the creative had already accumulated before the prior window opens, where the export can give it, since new reach cannot be computed without it.
• Placement breakdown for the same dates, or a statement that placements did not change.
• The date of the last creative edit on that ad.
• Which Advantage+ creative enhancements were on during each window and the date any of them changed (test.creative_enhancements_rule). Text variation, image expansion, music and 3D animation change the asset per impression, so an enhancement switched on between the two windows means the second window is a different asset and not a tired one.
• Any budget, bid or audience change inside the comparison period, with dates.

Analysis workflow

1. Confirm the ad has been live longer than fatigue.refresh_floor since its last creative edit. Below that, stop and report that delivery has not settled; anything read now is the learning period.
2. Establish that both windows ran the same asset, using test.creative_enhancements_rule. The rule is about the CHANGE between the windows, not the enabled state: an enhancement on for both is a constant and costs nothing here, while one switched on between them makes the CTR decline a comparison of two different assets. This is the single most likely manufactured cause of a false fatigue verdict on a current account, because these features get enabled account-wide by default and nobody logs the date. Where the answer is unknown, say so, cap the read at low, and do not write the word fatigue.
3. Set the comparison period from fatigue.window and state it as two date ranges, not as a duration. Apply the recalibration rule in that entry only if it fires, and say so if you moved it.
4. Gate on volume before reading any rate. Both windows must clear fatigue.volume_floor. Below it in either window no signal may fire at all, whatever the percentages look like, and the honest output is that this creative has not been served enough to be read.
5. Run ../scripts/fatigue_curve.py over the daily export for the signals it implements, which its own docstring names. This arithmetic is not eyeballed. A rate compared by eye across two uneven windows is the most reliable way to produce a confident wrong answer.
6. Read the CTR change against fatigue.ctr_decline_flag, confirming the figure is link CTR as defined in the denominators table, and confirming the change is in the rate rather than the click count. Without this one there is no fatigue verdict to reach, whatever else moved.
7. Read all three candidate second signals from the same export and the same two windows, since which one discriminates depends on the audience. fatigue.frequency_rise_flag is the relative move and is the one that works on retargeting and any small pool. fatigue.saturation_flag is the new reach decline and is the one that works on broad prospecting; compute new reach as window reach minus the reach the creative already had, and where the export cannot give that figure, say so in the output and fall back to fatigue.frequency_ceiling. Read fatigue.frequency_ceiling windowed, for the same window as the CTR you just used, and reject any lifetime frequency figure outright.
8. Apply fatigue.verdict_rule as revised: the CTR decline plus at least one of those three, inside the same window, both windows above the volume floor. Then apply its void condition. Where the only second signal is fatigue.frequency_ceiling and frequency was already above the ceiling in the prior window, the verdict is void, because a flag that was lit before the decline began did not fire. Say void, not fatigue, and name what the CTR fall points at instead.
9. Read CPM against fatigue.cpm_rise_flag last, and only as context on the auction. It is not one of the second signals in fatigue.verdict_rule and it never carries a verdict. Report it in the same breath as the CTR result or leave it out of the output entirely.
10. Check whether spend, placements or audience moved inside the window. If they did, the finding drops to low confidence under the confidence scale, and the change becomes the leading hypothesis.

Decision rules

• fatigue.ctr_decline_flag crossed and at least one of fatigue.frequency_rise_flag, fatigue.saturation_flag or fatigue.frequency_ceiling crossed in the same window, both windows above fatigue.volume_floor: this is fatigue. cut the creative or test a refresh of the same angle, and say which, on whether the angle has ever produced a winner. Name in the row which second signal fired.
• CTR crossed and the only second signal is fatigue.frequency_ceiling, which the creative was already above in the prior window: the verdict is void under fatigue.verdict_rule, not fatigue. investigate the audience, and say plainly that the ceiling was already lit and therefore did not fire.
• CTR crossed and no second signal moved: investigate the audience and the auction. Label the creative reading hypothesis, not export.
• A second signal moved and CTR did not: investigate audience size. The finding is the audience, and the creative stays live.
• Either window under fatigue.volume_floor: no verdict of any kind. monitor, and state how much longer at current delivery puts both windows over the floor.
• New reach cannot be computed from the export: say so, read fatigue.frequency_ceiling in its place as fatigue.saturation_flag instructs, and mark the row needs_data on the saturation question rather than treating a missing signal as a negative one.
• CPM moved while CTR held: report no CPM row and issue no fatigue verdict, per fatigue.cpm_rise_flag, which is reported alongside fatigue.ctr_decline_flag or not at all. Decision investigate, because the auction is outside the creative.
• An Advantage+ enhancement switched on between the two windows, or the change dates are unknown: no fatigue verdict. investigate under test.creative_enhancements_rule and say that the current window served a different asset from the prior one, so the decline is a comparison of two creatives rather than a decay of one.
• Ad live under fatigue.refresh_floor: monitor, no verdict, state the date it becomes readable.

Output format

Open with one line: is this creative tired, yes, no or void, naming the CTR result and which second signal fired or failed to.

This is the pack's time-series skill, so it labels confidence on the time-series conditions in ../references/output-standard.md rather than on the comparison ones, and it can reach high. Those conditions, stated here so a row can be checked rather than trusted: the signals came from an export, both windows were set by fatigue.window and are stated as dates, both windows clear fatigue.volume_floor, and no creative edit and no Advantage+ creative enhancement change fell inside either window (test.creative_enhancements_rule). All four. Exactly one missing, and not the volume one, is medium with the missing condition named. Either window under fatigue.volume_floor is low on its own, as is any window carrying a creative, budget, placement or enhancement change. test.volume_floor is not read here at all: it counts conversions per variant in a test, and this skill is not reading a test.

Then the decision, then What this could not see, Missing data, Approval gates.

Practical example

Illustrative made-up account data, not a real advertiser.

Input, one ad ID, two consecutive windows set by fatigue.window. Current window: 412,000 impressions, 158,000 reach, 1,030 link clicks, 4,120 spend. Prior window: 380,000 impressions, 152,000 reach, 1,330 link clicks, 3,420 spend. Last creative edit 61 days ago, so fatigue.refresh_floor is cleared. Enhancements: text variation was on across both windows and image expansion was switched on four days into the current one. Under test.creative_enhancements_rule the first is a constant and the second is not, so before any signal is read the output records that part of the current window served an asset the prior window never carried. Both windows sit well above fatigue.volume_floor on impressions, so the signals are allowed to fire at all, which is the first thing checked.

../scripts/fatigue_curve.py output: current link CTR 1,030 divided by 412,000 is 0.250 per cent. Prior link CTR 1,330 divided by 380,000 is 0.350 per cent. The relative change is 0.100 over 0.350, a fall of 28.6 per cent, which crosses fatigue.ctr_decline_flag.

Second signals, all three read on the same two windows. Windowed frequency is 412,000 over 158,000, which is 2.61 impressions per person, and sits below fatigue.frequency_ceiling. Prior windowed frequency is 380,000 over 152,000, which is 2.50, so the relative rise is 2.61 against 2.50, about 4 per cent, far under fatigue.frequency_rise_flag. New reach cannot be computed: the export supplied no cumulative reach before the prior window, so fatigue.saturation_flag is needs_data and, as that entry instructs, the fallback is the frequency ceiling, which is not crossed either. CPM is 10.00 against 9.00, a rise of 11.1 per cent, under fatigue.cpm_rise_flag, and it is reported here only because the CTR result is being reported beside it.

Verdict: no. Two things stop it independently. The CTR decline crossed and not one of the three second signals did, so fatigue.verdict_rule is not satisfied; and the image expansion switched on mid-window means the two windows are not the same asset, so under test.creative_enhancements_rule the decline has a manufactured explanation before fatigue is even reached. This skill cannot conclude fatigue. Falling CTR against flat frequency and unknown new reach points at the audience or the auction, and reach grew alongside impressions, so the leading hypothesis is a placement or audience shift. Decision investigate, not cut. Three things would settle it: the placement breakdown, the creative's cumulative reach, which would let fatigue.saturation_flag be read rather than assumed absent, and a re-read on two windows that both sit entirely after the enhancement change.

Guardrails

• Never declare fatigue on one signal, whatever the client expects.
• Never use lifetime frequency.
• Never compare click counts where the rule is written on rates.
• Never read fatigue on account impressions instead of creative impressions.
• Do not pause anything. Recommend, and let a human approve.