The short answer
Decide ad by ad against a target CPA, and ask two questions: how likely is it that this ad's true CPA is worse than target, and does it have enough data to be sure? Scale ads that are confidently under target, keep testing ads with too little data, and cut ads that are confidently over. Scale in modest steps and set stopping rules before you look.
Key takeaways
- Judge every ad against one explicit target CPA, not against the account average.
- Treat CPA as an estimate with uncertainty; low-volume ads need more spend before a verdict.
- Scale in modest steps and change one thing at a time; the common 15-20% step is a rule of thumb, not a platform rule.
- Decide stopping rules before the test starts: checking early and stopping on a good-looking result inflates false positives, including with Bayesian methods.
- Look at reach and fatigue signals, not only CPA; CPA is a lagging indicator.
Most paid-media waste comes from two mistakes made in opposite directions: scaling an ad that only looked good for a few days, and cutting an ad that simply had not spent enough to prove itself. A consistent decision rule removes both. This guide sets out one, and is honest about how strong the evidence behind each part is.
Start with a target, not an average
Every ad needs to be judged against one explicit number: the cost per acquisition (CPA) you can afford. Comparing ads with the account average hides the point, because an entire account can be above or below what the business can sustain. Work the target backward from margin, lifetime value and payback period, then hold every ad to it.
Treat CPA as an estimate
An ad with five conversions and a CPA of $30 is not the same as an ad with 200 conversions and a CPA of $30. The first number is mostly noise. The useful question is not "what is the CPA?" but "how likely is it that the true CPA is worse than my target?"
That framing gives three outcomes:
- Confidently under target: candidate to scale.
- Confidently over target: candidate to cut.
- Uncertain: keep spending, within limits, until it resolves.
Platforms offer a rough anchor for "enough data". Meta's help documentation says an ad set typically needs about 50 optimization events in the week following its last significant edit, and describes ad sets unlikely to reach that as "Learning limited". Meta has also said, in a post reported by PPC Land, that ad sets below that threshold ran about 28% more expensive per purchase. Treat that figure with care: the report notes Meta published no method, sample or controls, and smaller budgets, which also struggle to reach 50 conversions, are an obvious confounder. The threshold is a sensible floor; the 28% is a claim, not a finding.
Scale in steps, and change one thing at a time
Once an ad is confidently under target, scale it gradually. Meta lists changes to an ad set's optimization event, audience or creative, and in some cases its bid strategy or budget, as "significant edits" that may restart learning, with larger budget changes more likely to count.
Practitioner guidance converges on increases of roughly 15 to 20 percent at a time, spaced a few days apart. Sources disagree on the spacing, and the 20% figure is a heuristic rather than a number published by either Meta or Google. A sound approach:
- Raise budget by 15 to 20 percent.
- Wait several days and change nothing else.
- Check the learning status and CPA against target.
- Repeat only if the ad stays under target.
Note that scaling is not free. As spend rises, the cost of the next conversion (marginal CPA) usually climbs faster than the average. An ad can look healthy on average while the added spend is unprofitable, so judge increases by what the extra dollars bought, not by the blended figure.
Do not peek and then decide
The most common testing error is checking results repeatedly and stopping as soon as an ad looks like a winner or a loser. That inflates false positives. A common defence is that Bayesian methods avoid the problem. The evidence says otherwise:
- A 2018 paper by Markus Loecher, "Are Multi-Armed Bandits Susceptible to Peeking?", found that two commonly used Bayesian measures, posterior probability and value remaining, are severely affected by repeated testing, and depend strongly on the prior.
- A 2025 simulation by Alex Molas tested Bayesian A/B tests with no true difference between variants and a 95 percent posterior threshold. Peeking after every 100 observations produced a false-positive rate of about 80 percent. This is a single practitioner's simulation, not a peer-reviewed study, but it illustrates the mechanism clearly.
The practical rule: decide your sample size or stopping rule before launch. If you must look early, use a method built for it, such as a sequential test, rather than assuming a probability threshold protects you.
Watch the signals CPA is too slow to show
CPA is a lagging indicator. By the time it has visibly worsened, an ad may have been tiring for a while. Vendor sources commonly point to rising frequency combined with falling click-through rate as the earlier warning, and they warn that scaling speeds fatigue because it pushes frequency up faster. The specific frequency thresholds quoted vary from about 2.5 to 4 and come from vendor rules of thumb rather than controlled studies, so use them as prompts to look, not triggers to cut.
Also weigh reach. An ad slightly over target that is reaching many new people can be worth keeping for a short, defined period. A grace rule makes this explicit: for example, allow an ad to run up to 15 percent over target while it is reaching lots of new customers, and cut it if it stays there for 14 days. That is a judgment call, not a research result; the value is in writing the rule down so it is applied consistently.
A decision table
| Situation | Action |
|---|---|
| Confidently under target, enough data | Scale in 15-20% steps, watch CPA and frequency |
| Near target, strong reach | Keep spending under a time-limited grace rule |
| Uncertain, low data | Keep testing; do not scale or cut yet |
| Confidently over target | Cut, or raise the target only if the economics justify it |
| Under target but frequency up, CTR down | Refresh creative before scaling further |
How strong is the evidence?
| Claim | Source type | Confidence |
|---|---|---|
| About 50 events per week stabilizes delivery | Platform documentation | Moderate to high, as a floor |
| Under-50 ad sets cost about 28% more | Platform claim, no method published | Low |
| Budget changes can restart learning | Platform documentation | High; size thresholds not published |
| 15-20% budget steps | Practitioner consensus | Low to moderate; a heuristic |
| Bayesian tests are vulnerable to peeking | Academic paper plus simulation | Moderate to high |
| Frequency thresholds for fatigue | Vendor blogs | Low |
Putting it into practice
You can run this by hand in a spreadsheet. If you would rather not, XDO Media Optimization applies the same logic: upload ad-level results, set a target CPA, and it shows the probability each ad is over target and recommends whether to scale, keep testing or cut.
Whichever way you run it, the discipline is the same: one target, honest uncertainty, small steps and rules set before you look.
Sources
- Meta Business Help Center: About learning limited and Last significant edit
- PPC Land: Meta says ad sets under 50 weekly conversions cost 28% more per purchase
- Loecher, M. (2018). Are Multi-Armed Bandits Susceptible to Peeking? Zagreb International Review of Economics and Business, 21(1), 95-104
- Molas, A. (2025). Bayesian A/B testing is not immune to peeking