The 20% rule appears in none of the platform’s documentation. Not on the page about significant edits. Not on the learning phase page. Not on limited learning, not on the last significant edit metric, not on the cost reduction guidance.
What makes that absence meaningful rather than merely unproven is that the same documents are full of precise numbers. Fifty results in seven days. Seventeen purchases plus five. The platform is perfectly willing to publish a threshold when it has one. On budget percentages it publishes nothing, and it is not for want of the habit.
This page sets out what is actually documented, what the one budget example says, and how to scale without a rule nobody wrote.
What the platform does publish, precisely
Numbers, in the same pages where the 20% is supposed to live.
The learning phase, verbatim. The period when the delivery system still needs to learn about how an ad set may deliver and perform. Ad sets exit it as soon as they can deliver stably, which usually occurs after about 50 results in the week after the ad set’s last significant edit.
The level it applies to. The ad set, not the campaign.
A second, more specific threshold. For shops advertisements, a minimum of 17 purchases through your website and 5 through the platform is needed for the learning phase to complete.
Limited learning, verbatim. Described as not a penalty, but an indication that your budget is not being spent effectively because the delivery system cannot optimise performance with your current setup. An ad set becomes learning limited when it is unlikely to receive about 50 optimization events in the week after your last significant edit.
What that establishes. The platform publishes exact figures, including an oddly specific 17 plus 5, when it has one to publish.
Which is why the missing percentage matters. An organisation that will tell you it needs 17 website purchases and 5 in-app purchases is not being coy about numbers in general.
The one thing it says about budget changes
There is exactly one worked example, and it is a very wide interval.
The classification. Some edits are always significant: any change to targeting, any change to creative, any change to the optimization event, adding a new ad to the ad set, pausing for seven days or longer, and changing bid strategy.
The conditional ones. Ad set spending limit amount, bid control or cost per result goal or return goal amount, and budget amount. These are significant or not depending on the magnitude of the change.
The example, verbatim. That if you increase your budget from $100 to $101, that is not likely to cause one or more ad sets to reenter the learning phase. However, if you change your budget from $100 to $1,000, one or more ad sets may reenter the learning phase.
What that gives you. A 1% change described as probably safe, and a 900% change described as possibly not. Nothing between.
What it does not give you. 20%. Nor 15%, nor 30%. The interval where the answer changes is undefined, and the platform has chosen to leave it that way.
One exception it does state. Under automatic campaign budget, ad sets within the campaign will not reenter the learning phase as budget is distributed between them.
And the qualitative guidance. Use realistic budgets, set a budget large enough to get enough total results, and avoid frequent budget changes, which can cause an ad set to reenter the learning phase. Frequent is not defined either.
No primary source exists, so the origin has to be inferred. The mechanism is visible if you look at how the claim circulates.
What the platform offers. A qualitative rule, “depending on the magnitude of the change”, illustrated by two extremes.
What an operator needs. A number they can put in a process document and hand to a junior buyer.
What fills that gap. Someone rounds the vague guidance into a memorable figure. It gets repeated, and repetition becomes authority.
How you can watch it happening. Search for the rule and you will find pages that assert it as established practice and, further down the same page, note that it does not appear in the platform’s documentation. Both claims, in one article, without the author noticing the contradiction.
Why 20% specifically. Probably because it is round, feels cautious, and is small enough to be safe under almost any interpretation of “magnitude”. It is a reasonable guess dressed as a policy.
Which is not the same as saying it is bad advice. Increasing budgets gradually is defensible. It is just not a rule, and it should not be presented to a client as one.
Why no fixed percentage could be correct anyway
Even if the platform wanted to publish one, the number could not be a constant, and the reason is in the definition of the learning phase itself.
What the threshold is actually about. Fifty results in a week. Not fifty percent of anything, and not a budget figure. A volume of optimization events.
What follows arithmetically. An ad set producing 200 results a week has enormous headroom. One producing 55 is a single bad week from limited learning. The same percentage budget increase puts those two accounts in completely different positions.
And cost per result moves too. Doubling budget in a thin auction may double results. Doubling it in a saturated one may raise cost per result and deliver fewer additional events than the arithmetic suggests.
So the meaningful question is not the percentage. It is whether the ad set will still clear roughly 50 results a week after the change, at the cost per result the higher budget produces.
Which is computable. Take current weekly results, current cost per result, the new budget, and an allowance for cost per result rising. If the answer stays comfortably above 50, the change is safe in the sense that matters.
Why that is better than any rule. It is specific to your account, it uses numbers you already have, and it explains why the same 30% increase is fine on one ad set and destabilising on another.
The always-significant list is short, checkable, and more useful than the percentage argument.
Always significant. Any change to targeting. Any change to ad creative. Any change to the optimization event. Adding a new ad to your ad set. Pausing your ad set for seven days or longer. Changing bid strategy.
Note what is on that list. Adding a new ad. Which means a creative refresh, the thing most accounts do routinely, is guaranteed to restart learning in a way a budget change may not.
Note the seven-day pause. Turning campaigns off over a holiday period costs you the learning state if the pause runs a week.
Conditionally significant. Spending limit, bid control or cost goal, and budget amount, all depending on magnitude.
What this reorders. If you are worried about learning resets, the budget percentage is the least of it. Your creative rotation and your pause behaviour matter more, and both are documented as always significant.
The practical consequence. Batch your changes. If you are going to refresh creative and raise budget, do both at once rather than a week apart, because you are paying the learning cost either way and you may as well pay it once.
Five practices that survive the absence of a threshold, because none of them depends on one.
Change in increments that matter. A 10% budget increase on a small account is rounding error and teaches you nothing. If you are going to disturb delivery, disturb it enough to observe something.
Leave enough time between changes to reach stability. The documented marker is roughly 50 results in seven days. If your ad set does not produce 50 results in a week at any budget, learning phase discussions are not your problem; volume is.
Judge on cost per meeting held, not on the status label. The learning phase indicator tells you the system is still calibrating. It does not tell you whether the campaign is working.
Batch changes deliberately. One change window per week beats continuous fiddling, because continuous fiddling is exactly the “frequent budget changes” the platform warns against without quantifying.
Write down what you changed and when. Almost nobody does this, and it is the only way to attribute a performance shift to an action rather than to a story invented afterwards.
And if your ad set cannot reach 50 results a week. Consolidate. Fewer ad sets with more volume each is the documented direction, and it is also what the platform’s own limited-learning guidance recommends: combining ad sets, broadening the audience, raising budget or changing the optimization event.
The 20% rule appears in no platform documentation: not on significant edits, the learning phase, limited learning, the last-significant-edit metric or cost reduction guidance.
The same documents publish precise numbers, including 50 results in seven days and, for shops ads, 17 website purchases plus 5 on-platform.
On budget the platform gives one example: $100 to $101 probably will not restart learning, $100 to $1,000 may. Everything between is undefined.
Budget changes are only conditionally significant, “depending on the magnitude of the change”.
Adding a new ad is always significant, so a routine creative refresh restarts learning where a budget change might not.
So is pausing for seven days or longer, which catches anyone switching off over a holiday.
Automatic campaign budget redistribution does not restart learning between ad sets in a campaign.
The guidance on budget is qualitative: realistic budgets, large enough for enough results, and avoid frequent changes. Frequent is not defined.
It appears in none of the platform's documentation. Not on significant edits, not on the learning phase, not on limited learning, not on cost reduction best practices. No primary source for the figure could be found.
What does the platform actually say about budget changes?
That they may or may not be significant depending on the magnitude of the change, illustrated with one example: $100 to $101 is not likely to restart learning, while $100 to $1,000 may.
So what is the real threshold?
None is published. The platform gives a 1% example that is safe and a 900% example that may not be, and nothing in between. Anyone quoting a specific percentage invented it or repeated someone who did.
What is the learning phase exactly?
The period during which the delivery system is still learning how an ad set may deliver and perform. Ad sets exit it once they can deliver stably, which usually happens after about 50 results in the week following the last significant edit.
What counts as a significant edit?
Always: targeting, creative, optimization event, adding a new ad, pausing seven days or longer, and bid strategy. Sometimes, depending on magnitude: spending limit, bid or cost control amount, and budget amount.
Does automatic campaign budget restart learning?
No. The platform states that ad sets within a campaign will not re-enter the learning phase as budget is distributed between them.
What is 'learning limited'?
The platform describes it as not a penalty but an indication that budget is not being spent effectively. An ad set becomes learning limited when it is unlikely to receive about 50 optimization events in the week after the last significant edit.
So how should I actually scale?
In increments large enough to matter and infrequent enough to let the system stabilise, judged on cost per meeting held rather than on whether a status label reappeared. And write down what you changed and when.