How long should a Facebook ad run before you judge it?
Judge it by spend, counted in target CPAs, since days say little on their own. Below about 5 times your target, only an ad with no purchase or one purchase can be called a loser with any confidence.
From about 20 times target the line tightens to a CPA around 1.4 times target, and from 100 times to about 1.15 times. The table below gives every step, for Meta ads on Facebook and Instagram and for any other platform that counts purchases.
Why should the kill line move with spend?
At $250 on a $50 target, an ad that costs exactly $50 a purchase expects 5 purchases, and chance alone leaves it with 2 or fewer about one time in eight. A $125 CPA at that spend says little about the ad.
At $5,000 the same ad expects 100 purchases, and the count swings far less in proportion. Here 87 or fewer happens about one time in ten, so a $57 CPA carries about the weight of evidence that $125 did at $250.
A fixed CPA threshold has to pick which end it serves. If it suits a $250 read, it lets an ad running at $70 keep spending well past $5,000.
What are the cut and scale lines at each spend?
Read your ad's spend as a multiple of your target CPA, then compare its purchases with that row. Because only spend divided by target enters, the table works in any currency. Count the purchases your target is set on, for example Meta's 7-day click purchases, and judge prospecting ads apart from retargeting.
| Spend, in target CPAs | At $50 | Cut at or below (90%) | Daily checks: cut (99%) | Scale at or above (90%) |
|---|---|---|---|---|
| 5× | $250 | 2CPA 2.50× or more | 0no purchase | 9CPA 0.56× or less |
| 10× | $500 | 6CPA 1.67× or more | 3CPA 3.33× or more | 15CPA 0.67× or less |
| 20× | $1,000 | 14CPA 1.43× or more | 10CPA 2.00× or more | 26CPA 0.77× or less |
| 50× | $2,500 | 41CPA 1.22× or more | 34CPA 1.47× or more | 60CPA 0.83× or less |
| 100× | $5,000 | 87CPA 1.15× or more | 77CPA 1.30× or more | 113CPA 0.88× or less |
| 200× | $10,000 | 181CPA 1.10× or more | 167CPA 1.20× or more | 219CPA 0.91× or less |
Our calculation. Each figure is a count of purchases, with the CPA it implies as a multiple of your target. Cut means a one-sided mid-p Poisson test rejects "the ad costs exactly target" at 10%, or at 1% in the daily column; scale is the same test on the cheap side. Between the columns, keep spending and watch where the ad lands. With no purchase at all, an ad crosses the 90% cut line from 2.3× target, $115 on a $50 target.
Curtis Howland publishes cut points for a $50 target: cut at $250 of spend if CPA is $121 or more, at $1,000 if it is $72 or more, and at $5,000 if it is $58 or more, each at 90% confidence on a Poisson model (LinkedIn, 10 March 2026, repeated 13 July 2026; both checked October 2026). In purchases those are at most 2, 13 and 86. All three hold at 90% or more on this test, at 91.7%, 94.7% and 92.2%, with his $1,000 and $5,000 points one purchase on the cautious side. His scale points, $250 at a $32 CPA, $500 at $36 and $1,000 at $39, come out at 89 to 90.5%.
How does this compare with a flat 2 to 3 times target rule?
A common rule of thumb is a kill line at 2 to 3 times target CPA once an ad has some spend behind it. At small spends it sits close to the statistical line: at $250 on a $50 target, the 90% line cuts at $125, which is 2.5 times target.
At larger spends the two part ways. At $5,000 the 90% line cuts at a CPA of about $57, while a 2 to 3 times line waits for $100 to $150, 1.7 to 2.6 times higher. An ad settling at $75, 50% over target, stays under a 3 times line about 70% of the time; in the other 30% an unlucky start pushes it over.
Buyers keep the buffer for reasons a Poisson count leaves out:
- Conversion lag. Purchases for the newest spend are still arriving, so the last few days make every ad look dearer than it is.
- Attribution noise. Platform-reported purchases include modeled and view-through conversions, and daily counts swing more than a Poisson count allows.
- Order value. A CPA test treats every purchase alike, so an ad that brings larger baskets can earn a higher CPA.
- Learning phase. Meta says ad sets in learning "are less stable and usually have a higher CPA" (Meta Business Help Center, checked October 2026).
All four weigh most at small spend and in the newest days, which is where the line in the table is already lenient.
Does checking every day change the line?
Yes. The 10% is the chance of a wrong cut at one look, and every extra look gives an ad that is on target another chance to dip below the line.
We simulated ads checked after every $50 of spend up to $5,000, about daily for an ad spending $50 a day. An ad costing exactly $50 crossed the 90% line in 44% of runs. Checked weekly, it was 32%. The 99% line brings daily checks down to about 8%, level with a flat 3 times line, and still catches far more of the ads that cost too much.
| Ad's true CPA, $50 target | Flat 2× from $250 | Flat 3× from $250 | 90% line | 99% line |
|---|---|---|---|---|
| $50, on target | 21% | 8% | 44% | 8% |
| $60 | 37% | 15% | 87% | 46% |
| $75 | 63% | 30% | over 99% | 95% |
| $100 | 93% | 57% | 100% | 100% |
Share of ads cut by $5,000 of spend, checked after every $50, in a Nupact simulation of 40,000 runs per cell with purchases arriving as a Poisson process. The flat lines cut an ad with no purchase, or a CPA at or above the line, from $250 of spend.
The price is slower cuts: on daily checks, the 99% line cut a $75 ad at a median of about $1,450 of spend, against about $400 on the 90% line. Pick the column by how often you look. The scale side drifts the same way, so a daily scale call wants the stricter line too.
Should one ad past the line flag the whole account?
Expect a few ads past the line by chance. With 20 ads that all cost exactly the target, each read once at $5,000, at least one sits past the 90% line about 89% of the time.
Flag the account when more ads are past the line than chance would put there, or when one ad is so far past that it stays past after you allow for how many ads you checked. The simplest allowance divides the 10% by the number of ads, which statisticians call a Bonferroni correction.
What does the line assume?
- Poisson purchases. Purchases arrive independently at a steady rate, so the count at a given spend is Poisson with a mean of spend divided by target. Daily counts in an account swing more than that, which makes the line somewhat too quick to cut.
- No conversion lag. Leave out the last few days' spend, or wait until their purchases have had time to arrive.
- One-sided at 90%. The cut line asks only whether the ad costs more than target, and the scale line only whether it costs less.
- Mid-p test. It counts half the probability of the observed count. It cuts at 2, 14 and 87 purchases where the exact test cuts at 1, 13 and 86, which would put Curtis Howland's $250 point at 87.5%. For an ad exactly on target its false-alarm rate runs a little above 10%: 12.5% at 5× target, 10.5% at 20× and 10.4% at 100×.
- Floors on small budgets. Mid-p's false alarms climb when few purchases are expected, so an ad with no purchase is cut only by the exact test, from 2.3× target, and an ad counts as clearly cheap only with two purchases or more.
- One ad, one look. Daily checks and many ads need the adjustments above.
Each creative watched, and the recut or replace call drafted for approval.
Nupact reads fatigue from frequency and CTR drift on Meta and TikTok and from IPM decay on gaming and app networks, comparing each creative with its own history. A flag comes with the recut-or-replace call drafted, and the rotation waits for your approval in Slack unless you have let that action type run on its own inside your guardrails.
Every rotation is logged with its trigger, reasoning, approver and outcome, with a revert control on each row. If a conversion feed goes stale, automation pauses and the team is paged. See creative fatigue and anomaly detection.