Most creative fatigue advice has no source behind it

A client asks why cost per purchase drifted up last month. Someone on the call says the creative is fatigued. Frequency crossed 2.5, so it is time for a refresh. Everyone nods. The slide behind the claim has a curve on it. The curve has no axis labels. It has no source line either. The meeting moves on, and a month of production gets rebooked on that curve.

I have sat in that meeting for years. At some point I started asking where the number came from. The answer is always another deck. Follow the citation back and you land on an agency blog post. Follow that back and you land on a webinar. Follow that back and you land on nothing. No account level test. No spend behind it.

The published record is thin

So I went looking for the study that should exist. A scan of scholarly work from 2025 onward on ad fatigue and creative wearout surfaces very little. What is there is mostly non US. None of it is joined to real spend data from a live account. The Google Scholar listings from 2025 onward are easy to check yourself. It takes about ten minutes.

Two examples stand out. Beede, in the Journal of Marketing Communications in 2026, works with Super Bowl Ad Meter likeability scores from 2013 to 2024. Takaoki and Yamaguchi presented work on ad fatigue in RTB auctions at an IIAI conference in 2026. RTB means the programmatic market, where each impression clears in an auction milliseconds before it loads.

Both are honest pieces of work. Neither one tells you when to swap out a Meta ad. Ad Meter records how much a panel liked a Super Bowl spot. It never touches a cost per purchase inside a buying account. The RTB paper lives in a different auction with different pacing and different inventory. Nothing in either maps onto your refresh cadence.

The numbers in your deck came from somewhere softer

Take the line that top DTC brands ship 50 to 70 new Meta ads a week. That comes from Motion's 2025 Ad Creative and Creative Strategy Trends report. Motion surveyed more than 500 advertisers who chose to respond. DTC means direct to consumer, brands that sell on their own site. A survey of volunteers is a fine thing. It is not a census of spend.

Self selected matters more than it sounds. The brands who answer a creative volume survey are the brands who already care about creative volume. So the number reports what an engaged group says about itself. It does not report what spend actually looks like across the market. It certainly does not tell you the right weekly count for your account.

Or take the claim that Meta's Andromeda retrieval system raised model capacity 10,000 times. That figure moves through 2026 decks with no paper, no link and no document title behind it. Common Thread Collective ran it in 2026. Meta's own post on the ranking stack is dated 27 March 2025. It reports things like an 8 percent gain in ad quality. Retrieval is the step where the system narrows millions of ads down to a shortlist to score.

Follow the wearout number back far enough and the trail ends at another deck.

The measurement layer under all this is shakier than people assume. In 2025 the Ehrenberg-Bass Institute published an open access report by Hartnett, Bellman, Beal, Kennedy, Charron and Varan. They tested eye tracking, heart rate and skin conductance against EEG, to judge quality of attention to video ads. Heart rate discriminated best. Eye tracking did not. That lands on the thumbstop family of metrics. Thumbstop rate is the share of viewers who stop scrolling in the first seconds. It is eye tracking adjacent by construction. A fatigue rule built on it inherits the weaker signal.

The best case against me

Here is the strongest objection, stated properly. Fatigue is real, and every buyer watching an account sees it. An agency touching hundreds of accounts sees patterns that one account cannot. A rule of thumb drawn from 300 accounts beats a noisy read from yours. Most accounts do not have the conversion volume to see a clean curve anyway. So use the shared number and move on.

I agree that fatigue is real. I see it every week. My claim is narrower than it sounds. The constants do not travel. A benchmark pools accounts with different offers, different prices, different audience sizes and different campaign structures. The average of those curves belongs to no account in the set. On the volume point, you are not chasing a published result. You need a decision rule you apply the same way each week. A rough rule applied consistently beats a borrowed constant applied with confidence.

What to do on Monday

• Pull the last 90 days at the ad level. Take spend, impressions, weekly frequency and weekly cost per purchase.
• Plot cost per purchase against cumulative spend for each ad, not against days live. Spend is the thing that wears an ad out.
• Mark where the curve bends up and stays up for two straight weeks.
• Write that spend figure down for your five biggest ads. That is your account's wearout point, in dollars.
• Re-run it next quarter. The number moves when the offer or the audience moves.

You will end up with a number. It probably will not match the one in the deck. It will be built on your own spend, in your own account. That is the only version worth defending on a call.