A newly launched social campaign behaves erratically for a while before its results stabilize. This is a designed stage of delivery rather than a fault in the setup.
The system is estimating, not remembering
Every impression a social platform serves involves a prediction: how likely is this particular person to take the action the advertiser has requested.
For a brand new ad set the platform has no direct evidence about that combination of creative, audience and objective, so its early predictions carry wide uncertainty.
It resolves that uncertainty the only way available, by spending money across varied users and observing which impressions produce the requested outcome.
Exploration costs money before it saves money
During the early period the system deliberately serves people it is unsure about, because a prediction can only improve if the model sees outcomes across a spread of users.
Those exploratory impressions convert at a lower rate than the ones that follow, so cost per result is typically at its worst in the first days.
Advertisers who judge a campaign on those days are reading a measurement of exploration and mistaking it for a measurement of performance.
Edits restart the clock
Meaningful changes to an ad set invalidate what the model has learned, because the thing being predicted is no longer the same thing.
Budget changes of any size, audience edits, new creative and a changed optimization event all commonly reset delivery to an unsettled state.
This produces a familiar failure: an advertiser reacts to unstable early numbers by editing, which resets learning, which produces more unstable numbers.
Conversion volume determines how fast it ends
Stability arrives when the ad set has accumulated enough conversion events for predictions to hold up, which is why platforms publish a minimum weekly event count.
Advertisers with low conversion volume can move up the funnel, optimizing toward an earlier and more frequent action such as an add to cart rather than a completed purchase.
Consolidating budget into fewer ad sets has the same effect, since events accumulate against one prediction problem instead of being split across several.
Structure decisions follow from the constraint
The learning requirement explains why experienced buyers run fewer, larger ad sets rather than many small ones with tightly defined audiences.
It also explains why creative is tested by adding variants inside an existing ad set rather than by launching a new structure for every idea.
The practical rule is that anything which fragments conversion events across more prediction problems slows the whole account down.