How long should a test run?

Long enough to clear the learning phase and to cover a full weekly cycle, because weekday and weekend behaviour differ. Two weeks is a common floor. The conversion count decides it, not the calendar.

Should I test at my target CPA or above it?

Fund the test at the level that produces a readable result, then judge the result against your target. Underfunding a test to protect the target average is how you never find out whether the idea worked.

How many creatives should I test at once?

As many as can each get meaningful delivery inside the budget you actually have. That number is usually smaller than the number of ideas you walked in with.

Can I test on $500 total?

You can test something. At a $40 target CPA, $500 buys roughly a dozen conversions, which is enough to spot a disaster and not enough to pick between two reasonable options. Set expectations accordingly and say out loud which question you are answering.

Does pausing and restarting reset the learning phase?

Significant edits and restarts generally send an ad set back into learning. Plan the test so that you do not need to touch it once it is running.

How much should I spend testing a new campaign?

Short answer

Enough to reach a decision, which is a conversion count question rather than a dollar one. Meta's own documentation puts the exit from the learning phase at roughly 50 optimization events per ad set per week, so a test funded below that measures the learning phase rather than your idea. Work backwards from there: target cost per acquisition multiplied by the events you need, spread across the window you are willing to wait.

The arithmetic, worked

Take your target cost per acquisition. Multiply it by the number of conversions you would need before you would actually change your mind. Divide by the number of days you are willing to give it. That is your daily budget, and if the answer is uncomfortable, it is still the answer.

At a $40 target CPA, 50 conversions costs about $2,000. Over two weeks that is roughly $140 a day for one ad set. Somebody planning the same test at $20 a day is planning to spend $280, collect seven conversions, and reach a conclusion a coin could have reached faster and cheaper.

The threshold matters because it is not arbitrary. Below it the platform itself says delivery has not stabilized, so what you are measuring is the learning phase and not the creative.

Cheap tests are the expensive ones

A test that never exits learning measures the learning phase, not the idea.
Small samples swing wildly. Six conversions against four is not a fifty percent difference. It is two conversions.
You will kill a winner. This is the real cost and it is invisible, because nobody audits the ads they cancelled.
You will keep a loser for the same reason, and pay for it every week afterwards.
Slow tests get contaminated. Two weeks in, someone changes the landing page, a sale starts, the season turns, and the comparison is gone.

What to do when you cannot afford the honest number

Sometimes the budget is what it is. There are legitimate ways to make a smaller test readable, and every one of them works by reducing what you are asking the test to prove.

Test fewer things. One variable, two variants. Four creatives against three audiences is twelve cells and you can fund none of them.
Optimize for an earlier event. Add to cart happens more often than purchase, so volume arrives sooner. You are trading precision for readability, on purpose, and you should say so.
Consolidate. One ad set with real budget beats four ad sets splitting it and all sitting in learning together.
Test creative rather than audiences. Creative differences are usually larger and show up sooner.
Extend the window rather than lowering the threshold. Slow is worse than fast. Both are much better than wrong.

Write the decision rule before you start

Say in advance what number, over what window, makes you scale, iterate or kill. If you cannot write that sentence before the test runs, you are not testing. You are spending money and then telling a story about it afterwards.

Then honour it. The most common way a test fails is not statistical. It produces an inconvenient answer and gets extended until it produces a convenient one.

Guardrails while it runs

Do not touch the budget mid-test. Large edits reset delivery learning and restart the clock you are paying for.
Set the kill line in advance, in writing, including the spend ceiling at which you stop regardless of what the chart is doing that morning.
Do not read the result on day two. Conversions report late, and day-two data is a mood rather than a measurement.
Change one thing between rounds, so the next test inherits a clean comparison instead of a mess.
If an assistant is drafting or launching the test, state the ceiling explicitly in the brief. Models do not infer the constraint you forgot to mention. Every write in Muze previews before it runs and new campaigns are created paused, which gives you a second look at the budget field before anything spends.

Keep reading

Creative testing, definedStructuring a test you can actually read.What is a good ROAS?The target the test result gets judged against.How to scale a winning campaignWhat to do once a test wins.Should I let AI change my ad budget?Why the ceiling belongs in the brief.Campaign launchBuilding the test, created paused.Break-even ROAS calculatorThe floor a test result has to clear.

Ask your own account instead

Muze connects Meta Ads, Google Ads, Amazon Ads and Shopify to ChatGPT or Claude, so the answer comes from your numbers rather than a general one. Free to start, and every write previews before it runs.

See how Muze works