95% of Meta creatives fail. Guess who pays to find out.

Day one. You ship two ads you believe in.

Day four. One is quietly winning. The other isn’t dead, isn’t obviously wrong. It’s just being fed a little less.

Day nine. Meta throws a Learning Limited banner on the loser, and by then you have already spent whatever you were going to spend on it.

That is the tuition. Not the CPM. Not the CPA. The money Meta charges you to find out which creative wasn’t going to win.

Nineteen out of twenty ads don’t win. That is the number I keep coming back to. Motion’s 2026 study of 578,750 Meta creatives puts the winner rate at roughly 5%.^[1] The other 95% are what everyone means when they say creative testing is expensive. Below is the actual math on how expensive, sourced from Meta’s own earnings, WordStream’s 2025 benchmarks, and Motion’s dataset. It is not a comforting read.

How the money leaks

The mechanism is not exotic. When you launch a creative, Meta’s optimizer wants roughly 50 conversion events in the first week to exit what it calls the Learning Phase.^[11] Below that, delivery is volatile and the algorithm is hedging. You are paying for its uncertainty.

Here is what that costs, on WordStream’s 2025 numbers. All-industries US cost per lead on Meta is $27.66, up nearly 21% year over year.^[3] Multiply by fifty. About $1,383 to run one creative in one ad set until Meta will render a verdict. That is the price of finding out. If the creative is one of the 95% that don’t clear Motion’s winner threshold, the $1,383 is gone and Meta has already booked the revenue.

Outside the US the same math produces a very different bill. Superads publishes CPM medians across roughly $3B in aggregated Meta spend for July 2025 through July 2026,^[10] which lets us scale the US floor by geo. Local CTR and CVR move the true number up or down, so treat the non-US rows as first-order estimates.

And that is the floor, not the average. Most teams buffer above 50 events for confidence, run several ad sets in parallel per creative, and rebuild the Learning Phase every time they materially edit anything. Doubling the floor for a real kill decision is common. Some agencies plan on three to five times.

The 5% number, and why it ruins every plan

Motion’s Creative Benchmarks 2026 is the largest published dataset on Meta creative performance. 578,750 creatives, 6,015 accounts, $1.29 billion in Meta ad spend, September 2025 through January 2026.^[1] It is the closest thing to an independent audit of what happens after you click Publish.

Their definition of a winner is blunt. Ten times the account’s median ad spend, plus at least $500 in total delivery. Not “top-performing.” Not “shows promise.” The algorithm has decided to keep feeding this one.

About 5% of creatives clear that bar. Nineteen out of twenty do not. The bimodality is not incidental. Meta’s optimizer exists to identify a small handful of winners and starve everything else, and it does. Across the dataset, 55% of total ad spend concentrates on the winning 5%. At enterprise accounts, 64%.^[1]

“~5% of creatives are winners (≥10× account median spend, ≥$500).”

Motion Creative Benchmarks 2026 · 578,750 creatives, $1.29B in Meta spend

Losing creatives don’t merely underperform. They are what the account pays so the algorithm can find the winners. Which is fine, until you count what the finding actually costs.

Three costs your ad manager won’t show you

The media spend is only the visible layer. Three more compound on top of it, and none of them appear anywhere in the ad manager column headers.

Learning Phase resets are new tuition bills

Meta’s Learning Phase restarts every time an ad set undergoes what Meta calls a “significant edit,” including creative swaps, targeting changes, and material budget shifts.^[11] So: you spend the $1,383 to reach 50 events, decide the creative is a loser, and swap in a fresh one. The ad set goes back to zero. The next 50 events cost the same amount again. Fast iteration sounds cheap on a whiteboard. In practice, every kill is a fresh tuition bill.

The auction is getting more expensive

Meta’s Q4 2025 earnings release documents an average price-per-ad increase of 9% year over year across all of 2025, and 6% in Q4 alone. Ad impressions delivered grew 12% for the year.^[2] Meta is selling more ads and more expensive ads at the same time. The $1,383 US test does not buy in 2026 what it did in 2024. Compounded at 9% a year, the same floor costs about 19% more two years out.

The winner you didn’t fund

Because winners capture 55% of Meta spend across the Motion dataset, and 64% at enterprise scale,^[1] every ad set running a loser is displacing budget that would otherwise have been feeding a winner. The true cost of a bad creative is not just the money it burned. It is the return of the winner your budget didn’t reach in time. That number never shows up anywhere Meta will let you see.

Why the biggest spenders just run more creatives

Look at what the largest Meta advertisers actually do, and the answer isn’t “make better ads.” It is “make more of them.”

Motion’s data shows enterprise-tier advertisers launching 18.8 creatives per week and reaching an 8.2% winner rate, versus 3.8% at micro accounts.^[1][13] Enterprise creative isn’t inherently smarter than what a small brand ships. The advantage is arithmetic. At five creatives per week and a 5% hit rate, an account produces roughly one winner every four weeks. At twenty per week, one every week. Test more, surface more, feed more.

This is why creative volume has quietly become a structural input to media performance rather than a nice-to-have. It isn’t a philosophy. It’s a response to the winner rate.

What actually changes the math

If 5% of creatives are going to win, and each losing test costs somewhere between about $80 (India) and $1,400+ (US) just to reach the algorithm’s decision point, the operational question stops being can we make better ads. It becomes can we kill losers before we pay for the tuition.

Three responses show up in real practice.

• Test more, on purpose. The 18.8-creatives-per-week rhythm in the Motion data isn’t wasteful. It’s the volume game working as intended. At a 5% hit rate, more shots at goal is the only reliable path to more winners.
• Test off the auction first. Static concept tests, small-panel surveys, moderated focus groups, and synthetic panels (Splitroom is one of these; there are others) try to filter obvious losers before they touch Meta’s Learning Phase. None of them replace real delivery. Every one of them can replace a $1,383 tuition bill for a loser you would have caught anyway.
• Treat creative as a portfolio. Know your baseline hit rate. Cap the tuition per test. Kill fast. Most performance-marketing agencies have already converged on this. Motion’s benchmarks make the math visible, and Meta’s earnings make it clear the pressure isn’t relenting.

The 5% hit rate isn’t going down. The CPM isn’t going down. Meta’s price per ad is up 9% a year on Meta’s own reporting. Whatever an advertiser does about creative testing, the cost of ignoring it is now sourced and specific.