---
title: "Should Your Meta Agency Pay to Test New Creative or Keep Funding the Winner?"
description: "When one Meta ad takes most of the budget, is the agency protecting results or starving its challengers? The case for each side—and the testing rule I’d put in writing."
url: "https://groas.com/post/should-your-meta-agency-pay-an-exploration-tax-or-let-the"
image: "https://pub-87da24ecbbfc4c3bad6875f3aa013712.r2.dev/generated-images/59cd30d0-0fb5-4ea3-bea8-2f45891dc097.png"
published: "2026-11-06T05:13:44.616Z"
modified: "2026-11-06T05:13:44.669Z"
---

November 6, 2026 · 9 min read

# Should Your Meta Agency Pay to Test New Creative or Keep Funding the Winner?

[Alexander PerelmanHead Of Product @ groas](https://groas.com/author/alexander-perelman)[LinkedIn](https://www.linkedin.com/in/alexander-433793253/)

![A brightly burning candle, nearly spent, floods its wax over smaller unlit candles, while one tiny candle on the edge has just been lit as its successor.](https://pub-87da24ecbbfc4c3bad6875f3aa013712.r2.dev/generated-images/59cd30d0-0fb5-4ea3-bea8-2f45891dc097.png)

In this article

1. [Position one: let the winner keep spending](#position-one-let-the-winner-keep-spending)
2. [Position two: pay for exploration deliberately](#position-two-pay-for-exploration-deliberately)
3. [The hard disagreement: what is information worth?](#the-hard-disagreement-what-is-information-worth)
4. [My verdict: require an exploration contract, not a creative quota](#my-verdict-require-an-exploration-contract-not-a-creative-quota)

One founder reel took 79% of a supplement brand’s Meta budget in a week. At roughly NZ$75 a day, that left little for the other ads. Was the delivery system protecting the brand’s best performer, or denying its challengers a fair test? Smart advertisers disagree because both costs are real: force spend onto unproven creative and you may lose purchases today; never do it and you may have no replacement when the winner fades. ([the thread](https://www.reddit.com/r/metaads/comments/1xgb1lc/meta_performance_declining_at_similar))

The two camps have receipts. One advertiser put eight creatives in an ad set and watched a single ad take about 90% of spend on day one; top replies treated that as the system working ([that thread](https://www.reddit.com/r/FacebookAds/comments/1wbr3wq/meta_ads_1_ad_set_8_creatives_one_creative_taking)). Another saw every new ad launched beside a four-month winner steal delivery before it had earned any, leaving both ads performing worse: _it literally feels like I kill my own results every time I test_ ([the other thread](https://www.reddit.com/r/FacebookAds/comments/1sqkgbg/every_time_i_test_a_new_creative_my_performance)). Neither complaint is hard to understand.

I would not treat the NZ$75 account as a verdict on Meta. Its numbers were sliding before that reel took over: Meta-attributed ROAS fell from 3.50 in December to 1.05 in the last week of September, cost per purchase climbed from NZ$22.81 to NZ$72.97, and frequency rose from 2.37 to 4.07 ([the same thread](https://www.reddit.com/r/metaads/comments/1xgb1lc/meta_performance_declining_at_similar)). Uneven delivery alone does not explain that decline. But it puts the decision in focus. **Forcing spend into neglected creative can buy information; it can also sacrifice profitable delivery to make a testing report look balanced.** The agency’s job is to defend an exploration budget and a decision rule, not promise equal exposure or obey Meta’s allocation without question.

## Position one: let the winner keep spending

The exploit case starts with a fact that unsettles people new to Meta: **ads in one ad set are not promised even delivery.** An advertiser running eight creatives at ₹1,800 a day watched one take about 90% of spend on day one, and the experienced reply was essentially _hands off_ ([the thread](https://www.reddit.com/r/FacebookAds/comments/1wbr3wq/meta_ads_1_ad_set_8_creatives_one_creative_taking)). Another operator in the same thread reported one ad holding 97% of spend for weeks at strong ROAS. Their verdict: nothing to fix. The delivery system is estimating, not refereeing, and an ad with no purchase to its name may still have assisted a conversion another ad closed. Starved does not automatically mean lost.

Then price the cost of overriding that estimation. The advertiser with the four-month winner watched new ads steal delivery before they had earned any, and both ads came out worse ([the thread](https://www.reddit.com/r/FacebookAds/comments/1sqkgbg/every_time_i_test_a_new_creative_my_performance)). Giving challengers a fair share does not add budget. It takes delivery away from the ad currently paying the bills. Meta’s own creative-test feature recommends spending [no more than 20% of your existing budget](https://www.facebook.com/business/help/1423851372208214) on test ads to reduce the risk of harming overall campaign performance. Even the platform treats new creative as a limited allocation, not an equal-share contestant.

On a small account, the arithmetic sharpens the point. A practitioner guideline is to let a variation spend roughly 1 to 3 times your target CPA before judging it ([Mockingbird House’s math](https://www.mockingbirdhouse.com/guides/facebook-ad-testing-budget)). Multiply that across the four or five variants a testing calendar likes to promise, and an account spending NZ$75 a day is not running a neat experiment. It is running a queue. One analysis estimates [about $1,281 to detect a CTR difference](https://adriselab.com/blog/meta-ad-creative-testing-budget-calculator) at a $15 CPM and concludes that most modest budgets cannot prove conversion-rate differences between individual creatives at all. A fair test of everything at that spend is statistical theater. **Keep funding the ad that converts.**

## Position two: pay for exploration deliberately

Now give the other side its strongest case. An ad splitting the remaining 21% with other challengers has not necessarily lost. It may never have had enough delivery to show what it could do. **If you always let the winner eat, you can accumulate losers without learning why.** One operator describes the endgame of pure exploitation bluntly: 280-plus losing ads and no indication why any of them failed. No diagnosis, no reusable concept, no signal about what the audience responds to ([RevenueCat on the creative volume trap](https://www-docs.revenuecat.com/blog/growth/the-creative-volume-trap-in-meta-ads)). The winner looks efficient in the dashboard. The account gets dumber every month.

The second-order trap is worse. If testing hurts short-term performance, you test less. If you test less, you have no pipeline of challengers. Meanwhile, the NZ$75 account’s frequency climbed from 2.37 to 4.07 as cost per purchase rose sharply ([the thread](https://www.reddit.com/r/metaads/comments/1xgb1lc/meta_performance_declining_at_similar)). Whatever the full cause of that decline, waiting until a winner falters to look for its replacement leaves you buying information when you can least afford the delay. RevenueCat’s proposed measure is _Cost Per Winner_: total testing spend plus production costs divided by the number of creatives that actually won. Its framework pushes most testing budget toward genuinely new concepts, not iterations of the current winner that share its fatigue curve ([their framework](https://www-docs.revenuecat.com/blog/growth/the-creative-volume-trap-in-meta-ads)).

The exploration camp is not asking for equal budgets. It is asking for a deliberate reserve spent against a specific question. Practical frameworks offer two ways to keep that discipline: let each variation spend 1 to 3 times target CPA before judging it ([Mockingbird House](https://www.mockingbirdhouse.com/guides/facebook-ad-testing-budget)), or separate a _learn_ budget for new concepts from an _earn_ budget for proven ads ([Newform’s framework](https://www.newform.com/guides/meta-ad-creative-testing-framework)). Neither makes weak evidence strong. Both make the cost of learning visible. **An account that never spends on discovery is one fading winner away from a blank page.**

## The hard disagreement: what is information worth?

The academic name for this fight is the _exploration-exploitation dilemma_. Early estimates from thin data can push you toward the wrong option, but you cannot spend your whole budget gathering data either ([the framing](https://link.springer.com/article/10.1007/s41060-023-00493-7)). Methods that reserve budget for exploration gather more accurate data and, on tested budgets, convert more on average. That finding does not settle the decision for a NZ$75-a-day supplement brand. It prices the trade. The question is what one more piece of information is worth to this business, weighed against the cost of having no creative pipeline later.

Start by stripping the vocabulary off most agency testing reports. _We tested the new creative_ often means the agency ran both ads for a week and read the dashboard. That is not necessarily an experiment. Meta’s A/B test guidance calls for one variable, a stated hypothesis, non-overlapping audiences, and at least seven days with enough budget and results to determine a winner; a shorter run [may be inconclusive](https://www.facebook.com/business/help/290009911394576). Even a clean 50/50 budget split is not equal evidence, because the auction [may not distribute impressions equally](https://adriselab.com/blog/meta-ad-creative-testing-budget-calculator) across ads in the same ad set. Equal money can buy unequal data. If a report cannot state the hypothesis, spend limit, and rule for making a decision, it is not telling you what was learned. It is telling you what was spent.

Then price the cost of being wrong. Early purchase counts swing wildly at small volume: they move by roughly [their own square root](https://www.prism3.agency/blog/meta-creative-testing-small-shopify-budgets), so a challenger showing two purchases to the winner’s five has not necessarily lost a contest you can call with confidence. A delivery chart also leaves a downstream question untouched: two creatives can produce purchases at similar cost and still bring in different customers. If the winner fades and you have bought no discovery, the account has nowhere obvious to go. **The cost of exploration is visible today; the cost of skipping it arrives later.**

## My verdict: require an exploration contract, not a creative quota

I side with the exploration camp on principle and with the exploit camp on almost every testing plan someone has put in front of me. The deciding consideration is one thing: **information is worth buying only when you can name the question it answers.** A quota says _test four new creatives a month_, which is how accounts end up with a pile of losing ads and no diagnosis. A contract says _we are spending NZ$150 to find out whether a customer-testimonial hook beats our founder monologue on cost per purchase, and we will stop when the evidence is useful or the money runs out_. Only one of those tells you what the spend is for.

### Put four terms in writing before the test starts

If an agency cannot fill in all four, what it calls a test is a donation to the auction.

1. **The question.** What exactly are you trying to learn? _Does this creative work?_ is too vague. _Does a testimonial hook beat a founder monologue at the same target CPA?_ gives you something to judge.
2. **The maximum learning spend.** Set a figure in currency, not just a vague percentage. Meta recommends [no more than 20% of the existing budget](https://www.facebook.com/business/help/1423851372208214) for its creative-test feature. Treat that as a risk-conscious reference point, then name the amount this account can afford.
3. **The primary outcome.** Choose one metric the business trades in: cost per purchase, cost per booked job, or cost per qualified lead. Not click-through rate or thumb-stop rate dressed up as revenue.
4. **The stopping rule.** Plan enough time and budget to get a useful read. Meta’s A/B guidance calls for [at least seven days and enough results to determine a winner](https://www.facebook.com/business/help/290009911394576). If the spend limit arrives before the evidence does, stop and mark the result inconclusive. Do not turn two purchases into a trophy.

![Cut-paper model of a budget splitting into a wide amber river of production spend and a small teal reservoir labelled "learning spend."](https://pub-87da24ecbbfc4c3bad6875f3aa013712.r2.dev/generated-images/74d8de21-c2ee-48e4-86b3-16c0a953453e.png)

### Protect the production budget from the learning budget

Once the contract exists, the account structure has to honor it. Give the test its own ad set or campaign and its own budget, while the proven winner keeps its production line. That is the _learn-or-earn_ separation practitioners describe: one part of the account investigates new concepts, another converts ([Newform’s framing](https://www.newform.com/guides/meta-ad-creative-testing-framework)). If test ads and the winner compete for one limited pool, every learning dollar can displace delivery from a converting ad. That may be a cost worth paying. It should never be a surprise discovered in the monthly report.

The report then owes you three answers:

- **What did we learn?** State the verdict on the hypothesis in the business’s metric: cost per purchase, booked job, or qualified lead.
- **What remains unknown?** At NZ$75 a day, many comparisons buy a directional read, not proof. An agency that cannot say so is selling certainty it does not have.
- **Why does the next dollar go where it does?** Give me a sentence, not a slide. At [groas](https://groas.com/), actions come with logged reasoning for exactly this reason: a report should show what moved and why. Everything else is a receipt.

That supplement founder eventually paused the reel. I understand the impulse. But the account’s delivery chart could not, by itself, say whether another ad deserved more spend, how much that answer was worth, or when to stop buying it. _Pause whatever looks wrong_ is a reaction, not an exploration strategy. An agency collecting a fee to run the account owes its client more than a dashboard screenshot at the end of the month.

So here is the verdict and who it is for. **If you can absorb a week or two of noisy data, pay the exploration tax deliberately: name the question, cap the spend, and protect the winner’s production budget.** If you are spending around NZ$75 a day, do not pretend to run five concurrent tests. Pick one challenger, one hypothesis, and a small reserve you can defend. For either reader, the deciding question is the same: can the person running your account state, in a sentence, what a test dollar is buying and when they will know? If not, you do not have a testing program. You have spend wearing a lab coat.

When you next interview Meta agencies, do not ask _do you test creative?_ Ask them to describe the last test they ran: the question, the budget, the stopping rule, and what they learned. The answer separates operators from vendors faster than any pricing table. The ones worth hiring will reach for the numbers before they reach for the pitch deck.

## Frequently Asked Questions

### Why does one ad take almost all the budget in a Meta ad set with multiple creatives?

Meta's delivery system is estimating, not refereeing: it allocates spend where it expects purchases, so ads in one ad set are not promised even delivery. A starved ad has not necessarily lost, since it may have assisted a conversion another ad closed.

### Does forcing Meta to split spend between a winning ad and new creatives hurt performance?

It can. Giving challengers a fair share takes delivery away from the ad currently paying the bills, and one advertiser saw both the new ads and a four-month winner perform worse when new ads stole delivery before earning it. Meta itself recommends spending no more than 20% of existing budget on test ads.

### Can a small Meta ad budget really test five creative variants at once?

Usually not. A common guideline is to let each variation spend 1 to 3 times your target CPA before judging it, so a NZ$75-a-day account with four or five variants is running a queue, not a neat experiment. One analysis estimates about $1,281 just to detect a CTR difference at a $15 CPM.

### What happens if you never fund new Meta creative and just keep scaling the winner?

You can accumulate losing ads with no diagnosis of why they failed, no reusable concepts, and no signal about what the audience responds to. Meanwhile you build no pipeline of challengers, so when the winner fades you are buying information when you can least afford the delay.

### What separates a real creative test from a week of running two ads and reading the dashboard?

A real test needs one variable, a stated hypothesis, non-overlapping audiences, and at least seven days with enough budget and results to determine a winner, per Meta's A/B guidance. If a report cannot state the hypothesis, spend limit, and decision rule, it is telling you what was spent, not what was learned.

### Can I declare a Meta creative the winner after only a few purchases?

Probably not. Early purchase counts swing wildly at small volume, moving by roughly their own square root, so a challenger showing two purchases to the winner's five has not necessarily lost a contest you can call with confidence.

### What should I require in writing before a Meta agency runs a creative test?

Four terms: the question the test is answering, a maximum learning spend named in currency, one primary outcome such as cost per purchase, and a stopping rule with at least seven days and enough results. If the spend limit arrives before the evidence does, the result should be marked inconclusive.

### How should a test budget be structured so it does not hurt the winning ad?

Give the test its own ad set or campaign with its own budget, while the proven winner keeps its production line. This learn-or-earn separation means a test dollar cannot displace delivery from a converting ad, though any cost to short-term performance should be a known choice, not a surprise.

## Paid Ads & SEO, Run For You

Businesses pay one flat monthly fee per service, agencies resell it under their own brand, and groas does the work either way.

[See Pricing](https://groas.com/pricing)

## Structured data

```json
{"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://graph.groas.com/entity/groas","name":"groas","url":"https://groas.com/","description":"groas is a done-for-you marketing service, not software: elite people working with proprietary technology systems run search ads, Meta ads, SEO and AI search for businesses, and white label for agencies, with a dedicated senior strategist owning every account.","logo":{"@type":"ImageObject","url":"https://groas.com/icon-512.png","width":512,"height":512},"sameAs":["https://www.linkedin.com/company/groas/"]},{"@type":"WebSite","@id":"https://groas.com/#website","name":"groas","url":"https://groas.com/","inLanguage":"en-US","publisher":{"@id":"https://graph.groas.com/entity/groas"}},{"@type":"WebPage","@id":"https://groas.com/post/should-your-meta-agency-pay-an-exploration-tax-or-let-the#webpage","url":"https://groas.com/post/should-your-meta-agency-pay-an-exploration-tax-or-let-the","name":"Should Your Meta Agency Pay to Test New Creative or Keep Funding the Winner?","description":"When one Meta ad takes most of the budget, is the agency protecting results or starving its challengers? The case for each side—and the testing rule I’d put in writing.","inLanguage":"en-US","isPartOf":{"@id":"https://groas.com/#website"},"about":{"@id":"https://graph.groas.com/entity/groas"},"primaryImageOfPage":{"@type":"ImageObject","url":"https://pub-87da24ecbbfc4c3bad6875f3aa013712.r2.dev/generated-images/59cd30d0-0fb5-4ea3-bea8-2f45891dc097.png"},"breadcrumb":{"@id":"https://groas.com/post/should-your-meta-agency-pay-an-exploration-tax-or-let-the#breadcrumb"},"dateModified":"2026-11-06T05:13:44.669Z","mainEntity":{"@id":"https://groas.com/post/should-your-meta-agency-pay-an-exploration-tax-or-let-the#article"}},{"@type":"BreadcrumbList","@id":"https://groas.com/post/should-your-meta-agency-pay-an-exploration-tax-or-let-the#breadcrumb","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://groas.com/"},{"@type":"ListItem","position":2,"name":"Blog","item":"https://groas.com/blog"},{"@type":"ListItem","position":3,"name":"Should Your Meta Agency Pay to Test New Creative or Keep Funding the Winner?","item":"https://groas.com/post/should-your-meta-agency-pay-an-exploration-tax-or-let-the"}]},{"@type":"BlogPosting","@id":"https://groas.com/post/should-your-meta-agency-pay-an-exploration-tax-or-let-the#article","url":"https://groas.com/post/should-your-meta-agency-pay-an-exploration-tax-or-let-the","mainEntityOfPage":{"@id":"https://groas.com/post/should-your-meta-agency-pay-an-exploration-tax-or-let-the#webpage"},"isPartOf":{"@id":"https://groas.com/blog#blog"},"headline":"Should Your Meta Agency Pay to Test New Creative or Keep Funding the Winner?","description":"When one Meta ad takes most of the budget, is the agency protecting results or starving its challengers? The case for each side—and the testing rule I’d put in writing.","datePublished":"2026-11-06T05:13:44.616Z","dateModified":"2026-11-06T05:13:44.669Z","image":{"@type":"ImageObject","url":"https://pub-87da24ecbbfc4c3bad6875f3aa013712.r2.dev/generated-images/59cd30d0-0fb5-4ea3-bea8-2f45891dc097.png"},"author":{"@id":"https://groas.com/author/alexander-perelman#person"},"publisher":{"@id":"https://graph.groas.com/entity/groas"},"keywords":"winner, test, budget, new, cost, creative, meta, spend","wordCount":2011,"inLanguage":"en-US"},{"@type":"Person","@id":"https://groas.com/author/alexander-perelman#person","name":"Alexander Perelman","jobTitle":"Head Of Product @ groas","description":"My background includes Goldman Sachs and Stanford Computer Science.","image":"https://groas.com/media/blog/4f17cac3a81acc224cc1b2cfab8f94bbc295a3089d13e8421125771ef8d4064f.jpg","worksFor":{"@id":"https://graph.groas.com/entity/groas"},"sameAs":["https://www.linkedin.com/in/alexander-433793253/"]},{"@type":"FAQPage","@id":"https://groas.com/post/should-your-meta-agency-pay-an-exploration-tax-or-let-the#faq","mainEntity":[{"@type":"Question","name":"Why does one ad take almost all the budget in a Meta ad set with multiple creatives?","acceptedAnswer":{"@type":"Answer","text":"Meta's delivery system is estimating, not refereeing: it allocates spend where it expects purchases, so ads in one ad set are not promised even delivery. A starved ad has not necessarily lost, since it may have assisted a conversion another ad closed."}},{"@type":"Question","name":"Does forcing Meta to split spend between a winning ad and new creatives hurt performance?","acceptedAnswer":{"@type":"Answer","text":"It can. Giving challengers a fair share takes delivery away from the ad currently paying the bills, and one advertiser saw both the new ads and a four-month winner perform worse when new ads stole delivery before earning it. Meta itself recommends spending no more than 20% of existing budget on test ads."}},{"@type":"Question","name":"Can a small Meta ad budget really test five creative variants at once?","acceptedAnswer":{"@type":"Answer","text":"Usually not. A common guideline is to let each variation spend 1 to 3 times your target CPA before judging it, so a NZ$75-a-day account with four or five variants is running a queue, not a neat experiment. One analysis estimates about $1,281 just to detect a CTR difference at a $15 CPM."}},{"@type":"Question","name":"What happens if you never fund new Meta creative and just keep scaling the winner?","acceptedAnswer":{"@type":"Answer","text":"You can accumulate losing ads with no diagnosis of why they failed, no reusable concepts, and no signal about what the audience responds to. Meanwhile you build no pipeline of challengers, so when the winner fades you are buying information when you can least afford the delay."}},{"@type":"Question","name":"What separates a real creative test from a week of running two ads and reading the dashboard?","acceptedAnswer":{"@type":"Answer","text":"A real test needs one variable, a stated hypothesis, non-overlapping audiences, and at least seven days with enough budget and results to determine a winner, per Meta's A/B guidance. If a report cannot state the hypothesis, spend limit, and decision rule, it is telling you what was spent, not what was learned."}},{"@type":"Question","name":"Can I declare a Meta creative the winner after only a few purchases?","acceptedAnswer":{"@type":"Answer","text":"Probably not. Early purchase counts swing wildly at small volume, moving by roughly their own square root, so a challenger showing two purchases to the winner's five has not necessarily lost a contest you can call with confidence."}},{"@type":"Question","name":"What should I require in writing before a Meta agency runs a creative test?","acceptedAnswer":{"@type":"Answer","text":"Four terms: the question the test is answering, a maximum learning spend named in currency, one primary outcome such as cost per purchase, and a stopping rule with at least seven days and enough results. If the spend limit arrives before the evidence does, the result should be marked inconclusive."}},{"@type":"Question","name":"How should a test budget be structured so it does not hurt the winning ad?","acceptedAnswer":{"@type":"Answer","text":"Give the test its own ad set or campaign with its own budget, while the proven winner keeps its production line. This learn-or-earn separation means a test dollar cannot displace delivery from a converting ad, though any cost to short-term performance should be a known choice, not a surprise."}}]}]}
```