I build a Meta creative testing system from one sum: weekly budget divided by cost per purchase. That is what you can buy in a week, and it sets how many ads you can judge. On my accounts an ad needs about 10 purchases to show potential and about 50 for a verdict. At $100 to $500 a day you can judge far fewer ads than a production pipeline makes. So the system is mostly a rule for choosing the few ads you test, and at the low end I would not split testing from scaling. My client accounts are subscription apps and mobile games. I have worked on commerce apps too, but I have not run Meta for a web checkout product, and I say below where that matters.
Making ads got cheap, and budgets did not grow with it. The question I saw came from someone who sells a $25 to $30 digital product and spends about $100 to $500 a day on Meta. Their pipeline can produce many ads. Every new ad costs budget before you can read it. The purchase numbers in this article are my working numbers. They are not benchmarks.
How many ads can your budget judge in a week?
Divide your weekly budget by your cost per purchase. That is what the whole account can buy in a week, for the proven ads and the new ones together.
I use two numbers with different jobs. In my accounts an ad needs about 10 purchases to show potential and about 50 for a verdict. About 10 is a reason to keep an ad running and keep looking. I do not move budget on it. About 50 is where I am willing to give a verdict.
The 50 comes from one account, where I checked how well an ad’s ROAS (return on ad spend) in one week predicted the next. With 50 or more purchases in the first week the correlation was 0.44, which is still only moderate. With five to nine it was 0.05, and with 10 to 19 it was 0.09. So at about 10 purchases an ad has told me very little. Your threshold could be 30 as easily as 100, and that post shows how to find it.
Meta’s numbers point the same way, although its 50 counts something else: results per ad set, where mine counts purchases per ad. An ad set usually leaves the learning phase after about 50 results in the week after its last significant edit. A significant edit is a change that restarts learning. On 21 October 2026 Meta also published a formula. Multiply your cost per result by 50, then divide your total weekly budget by that number. The answer is the most ad sets your budget can support.
The table runs the sum on the budgets from that question. I do not know that seller’s cost per purchase, so I use an illustrative $20. Put in your own number.
| Daily budget | Weekly budget | Purchases a week at an illustrative $20 | Ad sets Meta’s formula supports | Ads the account could bring to about 10 purchases that week | Days of budget for one verdict of about 50 purchases |
|---|---|---|---|---|---|
| $100 | $700 | 35 | 0.7, fewer than 1 | 3 | 10 days |
| $300 | $2,100 | 105 | 2.1, so 2 | 10 | Between 3 and 4 days |
| $500 | $3,500 | 175 | 3.5, so 3 | 17 | 2 days |
Spend never spreads evenly across ads, so the count of ads in the fifth column is a ceiling. The proven ads also take their purchases first, so the number of new ads is smaller still. Look at the last column for $100 a day. One verdict takes 10 days of the whole account’s budget, so a new ad that shares the budget with proven ads is weeks away from about 50 purchases. My post on raising the budget when CPA rises runs the same kind of sum on about $50 a day.
At a budget this size I do not start from a share of budget for testing. I start from the purchases an ad needs and see what the budget can buy.
What changes when you cannot test on installs first?
On apps and games I have a way around the sum: a cheap install test before the purchase test. Many versions of a concept go into an install campaign where every ad gets the same start. The one or two that take clearly more spend go on to a purchase test. The install test only nominates candidates. It does not rank ideas by how well they will sell.
My rule is that the install test earns its cost when there are more versions than purchases to judge them by. For one team, I recommended testing on purchases directly as the better start, because that team made few ads a week. I did not run the two routes side by side on that account, so that was a judgment call. A team with a full pipeline and the budgets in the table has many more versions than purchases, so by my own rule it needs a cheap first test.
A web product has no install to test on first. Learning limited is Meta’s status for an ad set that is unlikely to get about 50 optimization events in a week. For that case Meta’s help page suggests an optimization event that happens more often, such as add to cart in place of purchases. The UK version of the page says add to basket. Meta offers that as a fix for learning limited and does not present it as a way to pick ads. A click test is another way to pick ads. I would bid on the event the platform can learn from and judge the test on a deeper one. On the web the deeper one is the purchase. So I would treat a cheaper event as an experiment, and whatever it picks still has to prove itself on purchases.
The other route is to push many new ads straight into a purchase test. I know the risk from apps. Each ad gets too few events, and an ad that got two purchases by chance can look like a winner.
With one price and no upsell, cost per purchase and ROAS rank ads the same way, and there is no trial or renewal to wait for. Every test ad still costs purchases. For a single order to make money, your maximum cost per purchase has to sit under the price. I call that limit the cost line.
Meta also has its own creative test inside a campaign. Its help page says you can compare up to seven variants, as of this writing, and suggests using no more than 20% of your existing budget. The test does not report a confidence level.
Should new ads get their own ad set or campaign at this budget?
At the low end of these budgets, no. On apps the separate test is the install campaign, and it is cheap. A purchase test has no cheap version. Meta’s guidance of about 50 results a week applies to each ad set, so a second ad set for new ads needs its own 50 results. By Meta’s formula the account in the table supports fewer than one ad set at $100 a day, two at $300 and three at $500, proven ads included.
So I would keep one ad set, put only a few ads in it and add new ones a few at a time. A split halves the events each side gets, and as I understand it, the learning happens inside the ad set. Meta’s 50 a week is about 7 a day. Below that, you could argue an ad set does not learn properly even with one ad. Still, Meta’s own page says learning limited is not a penalty.
This is my reasoning from app accounts on budgets like these. On one account I found the purchase campaign split into an old winner and the new ads. I asked the team to combine them into one ad set and to keep it to a few ads. On that account the new ads still came through a cheap install test before they reached the purchase ad set.
One thing I am still not fully sure about is whether a concept should sit in its own ad set or be mixed with others. As I read Meta’s guidance, it says mix them, and I sometimes see better results when I keep them apart.
Adding ads to the proven ad set has a cost. Meta lists “Adding a new ad to your ad set” among its significant edits, and every significant edit resets the clock on the learning phase. That is a reason to add new ads in small batches, not one at a time. I have also written that the learning phase is overhyped and that 50 is a Meta number, not a law. So I take this as a reason for batches only.
By Meta’s guidance a second ad set fits when it can reach about 50 purchases a week by itself. At the illustrative $20 that is $1,000 a week, about $143 a day. Put the other way round: you can run two ad sets when your cost per purchase is no more than your weekly budget divided by 100. By my arithmetic that is $7 at $100 a day, $21 at $300 and $35 at $500.
Where the budget allows it, I put the best new ads into a second ad set, without the top spenders and under the same campaign budget. I set a minimum and a maximum spend on that ad set, so the campaign budget cannot starve it or overfund it. On apps the best new ads are the ones that earned spend in the install test. Without an install there is no such filter, and the choice is a judgment call. On one account the team and I agreed to try a winners ad set and a test ad set under one campaign budget. I warned that a test ad set with no minimum spend may end up not spending, or may spend more than planned, and I said to expect side effects.
That second ad set sits inside the same campaign. In most large accounts I see, a separate validation campaign sits between the test and the main campaign. New winners prove themselves on purchases there before they join the proven ads. By my rule of thumb I would only add one from roughly $100,000 a month.
Which ads do you test when the pipeline makes more than the budget can judge?
When the budget can judge only a handful of ads, choosing them is most of the work. Most of what the pipeline makes should not launch at this budget. If the ads come from creators, the choice starts earlier: I decide which ad I want made before I look for a creator. My testing has two levels. The first is concepts that differ a lot, and I judge each as a whole idea on purchases. The second is versions of one concept. A version changes one specific thing and runs against the ad it came from.
Concepts come first. On one account the concept alone explained 17.0% of the variation in ROAS and the theme of the opening hook 7.0%. I test ideas that differ a lot before I polish versions. Near identical ads count as one test. I judge a version against the ad it came from, so the original runs next to it. On app accounts I count a version as failed when it spends less than its source. In a purchase ad set with few events a week, the source holds the budget, so I would read that rule with more care.
I say two things about volume that pull against each other, and I hold both today.
The first is that more is better. I have written “Test 100, 500, even 1000 ads if you can produce them”. This year I have told teams that 20 new ads a week is normal to me. In my experience a tiny share of ads brings most of the results, so testing few ads gives few chances to find one.
The second is that cheap production is expensive. I judge a creative program by its cost per winner and by how often it finds a new one. I used to produce a large number of ads with AI tools. My cost per winner did not improve, and in some cases the noise slowed me down.
On apps the install test makes many attempts affordable. Without it, every attempt costs purchases.
How do you give each ad in the test a fair read?
My default is the same start for every ad, and then hands off. On apps that means all ads restarted together, then budgets and ads left untouched. In a purchase campaign my default is a campaign budget with a minimum spend on the ad set I am testing.
Where the money goes is my first read, because ROAS zigzags. Cost comes second. I compare ads inside one group. I do not give each ad a fixed amount and judge it on that, because what the first dollars show can change the next day. When one ad in a group that started together takes nearly all the spend, the test has given its answer.
I do not rank ads on click rate (CTR) or on cost per 1,000 impressions (CPM). The stop rule for apps in the next section still checks the top of the funnel, so on the web I would check that clicks arrive before I blame the checkout.
Reading spend has limits, and I have seen it go both ways. After an equal start in a test group of a few ads, spend is my best early read. Inside a large live campaign a share of spend is not a verdict, because older ads and early luck hold the budget. With one ad set the new ads do not get an equal start either, so a low share of spend for a new ad is not a verdict there.
I have seen Meta get it wrong. I once duplicated a campaign without its five top spenders. None of the five had hit target, and the duplicate found six new creatives above target that never got a chance. That was not a clean split test. I have also written that an ad that got little spend has not lost.
On the other hand, I have told clients that I do not go back to rescue ads Meta barely spent on. I once built a separate campaign from the five lowest spending ads of each ad set, and the result confirmed that Meta had been right to starve them.
An ad’s share of spend can take weeks to settle. In the Horse Racing Solitaire case, Meta walked one ad’s share of the daily budget from a peak of 30% down to 4.5% within 22 days. So I read in stages:
- In the first days I check whether the ad is alive: whether it gets spend and its first purchases. On apps, a strong purchase campaign shows its first trials or purchases within one to three days.
- In the accounts I run, a new ad takes about a week to find its share of spend.
- Budget moves later, on total purchases and a second good week. The next section has the rule.
When do you stop an ad, and when does a winner move?
I write the rule before launch. It has four parts: the metric, the spending limit, the review window and what counts as inconclusive. With one price the metric is cost per purchase against your cost line. The spending limit is cost per purchase times the purchases you want times the ads in the test. At the illustrative $20, three ads that each get about 10 purchases come to $600, most of a week at $100 a day. The review window is the staged read above. When no ad stands out, I record that and keep the ads for a smaller retest later. I do not keep spending to force a winner.
The stop rules on this site are for apps. A purchase test stops after $200 to $300 of spend over about three days when the top of the funnel looks decent and there are no trials or purchases. Set your levels from your own cost per purchase. I have not published a multiple of target cost that I would apply everywhere, and part of the call is judgment.
My rule is that I wait for a second good week before I move budget on the strength of the first. On the budgets in the table that rule protects less. A single ad will often collect fewer than 20 purchases in a week, and below 20 the one account I measured shows little signal from one week to the next. So a second week only protects me from one lucky week. The verdict comes from the purchases an ad has collected in total. The analysis on that account counted purchases inside one week. It did not test adding them up across weeks, so that part is my reasoning. An ad that passes a test has better odds of holding up and has proved nothing yet.
I have also moved winners without waiting for a second good week. On one account my rule for a test of trend ads was that a clear winner moves to the main campaign within three days.
With one ad set there is no move to make, and the decision is which ads leave. The ad set keeps the best handful: by my rule of thumb five to 10, or even three at the low end of these budgets. My answers on that number have varied by account. I take out old winners once the new ads beat them.
On app accounts with room for it, I duplicate the best campaign, raise the duplicate and watch the pair together to scale. On the budgets in the table a duplicate is one more ad set to feed, so the sum at the top applies to it too. A higher budget is also what lets the account judge more ads.
Where has my creative testing gone wrong?
For a fitness app client spending $1,500 a day I tested more than 40 creatives. Meta put 90% of the spend on two of them, and neither hit the target of $15 per trial. At $30 per trial, testing one concept properly cost $300 or more before I got signal, and each test took seven days. The tests I wanted did not fit the budget. Those are one client’s numbers. They are not benchmarks.
On one account my validation campaign let almost nothing through. The team and I ended up with a validation ad set inside the main campaign.
If no ad in a test wins, I change the idea.
I have not run Meta for a web checkout product or used Meta’s creative test. I have also not tried a cheaper web event or a click test as a first round for picking ads, and I have not put new ads straight into one purchase ad set with no install test before it. The purchase sum, the rule written before launch and testing ideas before versions should carry over. The install test does not, so on the web fewer ads get tested and the choice of ideas matters more.
Sources and scope
I opened the five Meta pages below on 14 November 2026 and quote them as they read that day. The $20 cost per purchase is made up for the sums, and I did the arithmetic myself. The purchase numbers are working numbers from the subscription app and game accounts I run. The client patterns are anonymous, and each one is a single account.
- Meta, About the learning phase
- Meta, About learning limited
- Meta, Significant edits and learning phase
- Meta, Set up a creative test in Meta Ads Manager
- Meta for Business, Peak Season 2026: You Don’t Beat The Auction. You Brief It., 21 October 2026
Questions people ask
How many ads should I test at once on Meta?
As many as your budget can bring to about 10 purchases each in a week, counting the ads already running. The 10 is my working number from app accounts. Weekly budget divided by cost per purchase, divided by 10, is the ceiling. The real number is lower, because spend never spreads evenly across ads. I keep an ad set to a handful of ads.
How many purchases does a Meta ad need before I call it a winner?
On my accounts, which are subscription apps and games, an ad needs about 10 purchases to keep running and about 50 for a verdict. On the one account where I measured it, even 50 gave only a moderate read, so I wait for a second good week before I move budget. Your number can differ, so check it in your own account.
Should creative tests run in a separate campaign or ad set?
Only if the test can reach about 50 results a week by itself, which is Meta's guidance per ad set. Below that I would keep one ad set with only a few ads and add new ones a few at a time, because a split halves the events each side gets. With an app, a cheap install campaign can be the separate test.
How much of the budget should go to creative testing?
At $100 to $500 a day I do not start from a share. I start from the purchases an ad needs and see what the budget can buy. Meta suggests no more than 20% of the existing budget for its own creative test inside a campaign.
Does a creative testing method from app accounts work for a web product with one price?
Partly. Three things carry over: sizing the test by weekly budget divided by cost per purchase, writing the stop rule before launch, and testing different ideas before versions of one idea. The cheap install test I run first on apps has no web equivalent that I have tested. I have not run Meta for a web checkout product myself.