Creative testing on Meta that stops the losers early and scales the winners.

Every new winner costs a run of test ads that never pay back. Nupact reads each Meta test ad every day, at every step of the funnel, and gives it one of three verdicts: stop, keep testing or move to scale.

By Federico Baravalle, co-founder of Nupact · Published

How does Nupact decide when to stop a test ad?

It compares the ad with your account’s recent winners, matched on landing page and format where possible, step by step from the first seconds of a video to the purchase. A model trained on your past tests reads the early signals together. Only clear evidence stops an ad or moves it to scale, so an ad with too little data stays in test.

Get your backtestSee the result
In the product

One test ad, read step by step.

This video beat its benchmark on 15- and 30-second views and on clicks, then fell behind at landing page views and purchase. Its benchmark ads send traffic to the same page, so the page itself is ruled out.

Nupact · experimentsMeta
Test ad · video, 30–60 s · launched 9 Oct · €765 spent
Verdict on 18 Oct: stop. Behind the benchmark at landing page views (44% of clicks against 57%) and at purchase.
3 s viewIgnoredA deeper step passed
15 s viewPassed · +54%13 Oct
30 s viewPassed · +36%13 Oct
ClickPassed · +44%14 Oct
Landing pageFailed · −22%18 Oct
PurchaseFailed · −71%18 Oct
Proposed: pause this ad. The rest of its ad set keeps running.ApproveKeep testing
How it works

How a test ad gets its verdict.

Each ad is judged against your own winners.

A kill rule waits until an ad has spent enough without a purchase. Nupact reads every step of the funnel against the ads that already made money for you, matched on landing page and format where possible, so the early signs count from day one.

01

Learn the account

Benchmarks come from your recent winning ads, matched on landing page and format where possible. A model learns from your past tests which early signals point to a winner or a loser, and how soon they show.

02

Read every test ad

Each live test ad is checked at every step of the funnel, from video views through the click, the landing page and add to cart to the purchase, using only what the account knows that day: views the same day, 1-day click purchases the next day, 7-day click value after a week.

03

Give a verdict

The model weighs each step against the benchmark and reads the early signals together. When one step is weak but the ad could still reach your target, it stays in test. Only clear evidence gives a stop or a move to scale.

04

Act on the day

The saving comes from the days after a verdict. Most teams start with verdicts as alerts or proposals to approve, then let Nupact act inside the guardrails they set.

The backtest

411 test ads, replayed day by day.

A German DTC brand launched 411 test ads on Meta from January to October 2026 and spent €239k on them. We replayed every day of every test with only the data known on that day.

Test spend saved, net
12%€28k of €239k in ten months
Clear losers stopped
66 of 67none sent to scale
Clear winners scaled
13 of 19of the clearly profitable ads went to scale
Winners stopped by mistake
1 of 19the cost of deciding early
VerdictAdsClearly losing (67)Clearly profitable (19)Unclear (325)What happened next
Stop248661181Median stop on the day after launch
Move to scale3301320€1.23 of purchase value per euro after the move
Keep testing13015124Too little evidence for a verdict, so a person decides

Return is purchase value divided by spend on Meta’s 7-day click, and break-even is a return of 1. An ad counts as clearly losing or clearly profitable when the 95% range of its whole-run return lies below or above 1, which is known only once the run ends and was never shown to the verdicts.

How much there is to save depends on how much your tests lose. In the first quarter this brand’s test ads returned €0.45 per euro and the verdicts saved 44% of the spend. From the third quarter they ran near break-even and the share fell to 4–6%, with the fourth quarter counted to 20 October.

Get the full case study

Four pages on how each verdict is reached and what the saving was in each quarter. We email you the PDF.

How sure we are

Tested on ads the model never saw.

New ads only
The model learned from ads launched up to the end of 2025, so every 2026 test ad in the backtest was new to it. The rules around the model were set on the 2026 ads.
No hindsight
Each day’s verdict used only what the account knew that day. It never looked at how a run ended.
10,000 redraws
We reshuffled the 411 ads 10,000 times, as if the brand had launched a different mix. Every run saved money: between 8% and 16% of test spend in 95% of them.

What else do people ask about creative testing?

Which ad platforms does it work on?

Meta, for now. Verdicts need a steady flow of test ads and Meta's funnel data, from video views to purchases. Fatigue on ads that have already scaled is watched on Meta, TikTok and app networks.

What do you need for the backtest?

Read-only access to your Meta ad account, with about two years of history. Nothing in the account changes. We agree with you how your test ads are marked, for example a testing campaign or a naming rule, and send you the result within 72 hours of your data arriving.

Does the backtest cost anything?

No. The backtest is free, and the access can be removed whenever you like.

What break-even does the backtest use?

Yours. The case study used break-even, a return of 1 on Meta's 7-day click. On your account we draw the line at your own target; set it higher and more ads get stopped.

How is the saving calculated?

It is what the stopped ads spent after their stop day, minus the sales those days brought. It counts the spend avoided, not what that money would have earned in other ads.

Will it stop a winner by mistake?

Sometimes. In the backtest, 1 of the 19 clearly profitable ads was stopped and 5 stayed in test. Most teams start with verdicts as proposals to approve, so a person sees every stop until they trust the calls.

Does Nupact stop ads on its own?

Only if you let it. A verdict can arrive as an alert, as a proposal you approve, or as an action Nupact takes inside guardrails you set. Every change is logged with a revert control.

How is this different from Meta's A/B tests?

Meta's A/B tests split an audience between a few versions for a set period and name a winner at the end. Most creative testing is a steady stream of new ads in a testing campaign, with no end date. Nupact reads each of those ads every day and calls it once the evidence is clear.

What happens to our ad data?

Client data is stored in the European Union and segregated per client. Our AI model providers do not use client data to train their models.

Your own backtest

Your test ads, replayed in 72 hours. Free.

Give us read-only access to your Meta ad account with about two years of history, and tell us how your test ads are marked. We replay them the same way, against your own target, and send you the net amount you could have saved and the ads it would have moved to scale, with the day of each call.