How long to run Facebook ads before changes

An ad set on €80 a day. Same offer, same audience, same three creatives all month. Monday it brings four leads at €20 each. Tuesday it brings one, and the dashboard says €80 a lead. Nobody touched anything.

Tuesday evening someone raises the budget by ten percent, because sitting still feels like negligence.

I have made that edit. It is the most expensive habit in a small account: it does nothing for the ad, and it wipes out the only stretch of comparable data you had.

Short answer: Wait until the ad set has produced enough conversions for its numbers to mean something, which is a different number of days for every account. An ad set doing four conversions a day needs about two weeks to reach fifty events. One doing forty gets there before lunch on day two.

The takeaways

• The waiting period is an event count, not a number of days. Four conversions a day and forty conversions a day both hit day seven on Sunday, with 28 events in one account and 280 in the other.
• A small account swings harder, and that is arithmetic. The wobble on a daily conversion count runs at roughly one over its square root: about 50% at four a day, about 16% at forty.
• An early edit costs more than the wait. Meta's "About learning limited" documentation puts a stable read at roughly 50 optimization events per week, and a budget or targeting change sends the ad set back to collecting them.

Why does the same ad cost €13 one day and €40 the next?

Because a daily conversion count is a small number, and small numbers wobble. That €80 ad set at a €20 cost per lead produces four leads on an average day. Two on a quiet day and six on a good one are both ordinary. That alone prints €40 a lead on Tuesday and €13 on Thursday, with the creative, the audience and the auction all sitting perfectly still.

The size of the wobble is knowable in advance. Counting noise on a count of n runs about the square root of n, so the swing as a percentage is one over that square root. Four conversions a day gives you 50%. Forty gives you 16%. Four hundred gives you 5%.

Which explains why the accounts posting the wildest daily screenshots are the small ones, and why "just be patient" reads as such a weak answer. Patience is not a virtue here. It is the only mechanism you have for turning a jumpy count into an average worth reading.

How many days does your account need?

Divide the events you want by the events the ad set produces in a day. That is the whole calculation, and it lands in a different place for every account.

Take Meta's roughly 50 optimization events per week as your floor. At four conversions a day, that is thirteen days. At forty a day, it arrives on day two.

So the seven-day rule every guide repeats hands one account 28 events and the other 280. Same calendar, two completely different levels of confidence, and only one of them buys you the right to act. The rule is not wrong so much as unfinished: it answers a question about sample size with a unit that has nothing to do with sample size.

Do the division before you launch and write the date on the brief. A number you agreed to in advance is much harder to argue yourself out of at 11pm.

Why does checking every morning make it worse?

Because every look is another chance to mistake noise for a signal. Seven independent looks at a 95% threshold leave you roughly a 30% chance of at least one false alarm (1 minus 0.95 to the seventh). Those looks are not independent, since each day shares most of its data with the day before, so the real figure sits lower than 30%. The direction holds anyway.

The practical version is simpler. If you open Ads Manager every morning with the power to intervene, you will intervene, and it will usually be on the worst day of an ad that was doing fine.

Read a rolling window instead of yesterday. Seven days trailing, updated daily, moves slowly enough to be readable and fast enough to catch a real collapse. If a number has to live on a dashboard, put the trailing window there and hide the daily column.

What does an early edit cost you?

The comparison you were building. A budget change, a targeting change or a swap of the optimization event restarts learning, so the days before and the days after stop being the same experiment. You have spent a week of budget to produce two half-samples, neither of them readable. I wrote up what the learning phase is doing to your data separately.

That is the trap under the 2% budget nudge. It feels harmless because the number is small. The reset it triggers is the same size as the reset from doubling the budget.

Our own automation refuses to touch anything inside a learning-phase lockout: under five days or under €200 spent, no rule fires, whatever the metrics say. After that, kills stay reversible for 24 hours. Both rules exist for the reason a human should wait, which is that the numbers underneath are not ready to carry a decision yet.

When is waiting the wrong answer?

When the ad is bad and you are paying to confirm it. Waiting buys you a readable number; it does not improve the ad. An ad set that has spent three times your target cost per result with zero conversions has already told you something, and no further calendar days will rescue it. Set that floor before launch as well.

The other exception is a real decline. One bad day is noise. A click-through rate sliding for five days straight is a slope, and a slope carries information, provided the rest of your account and the wider market did not slide with it. Adscalr reads fatigue off a multi-day slope with a market-wide check, so a bad day for everyone does not get pinned on your creative, and it pulls a new ad's early score toward what its format normally does using format-specific priors, so a lucky opening day cannot win a ranking by itself. The ad-intelligence page covers how that scoring behaves on thin samples.

Which of two ads won is a related question with its own trap, and I worked through it in winner or lucky streak. This post is about the one before it: how many days of data buy you the right to touch anything at all.