We ran a purchase-optimized campaign on a brand-new ad account and a brand-new dataset. Structure was correct on every axis: the campaign carried the sales objective, the ad set optimized for purchases against the right dataset, targeting was a single large market, both ads passed review and sat ACTIVE. It spent nothing. Zero impressions, from launch to the next morning.
Nothing was broken. That outcome is what purchase optimization on an empty dataset is supposed to look like, and it is worth understanding before you spend a week debugging a campaign that has no defect.
A brand-new dataset has no purchase history, and purchase optimization needs that history to find anyone worth bidding on. The campaign can be correct on every axis and still not spend, because the request itself is unanswerable. Start the ad set on an event the account can actually produce, then move up to purchase once the volume exists.
What purchase optimization actually asks for
When you optimize for purchases, you are not asking Meta to show your ad to people who might buy. You are asking it to find people who resemble the people who already bought — resemble them in the specific, high-dimensional way Meta's models define resemblance. That instruction is only executable if the dataset holds purchase events to learn from.
On a dataset with zero purchases, the instruction has no referent. Delivery does not fail loudly; it simply never finds a bid it is confident in, and the campaign sits there looking perfectly healthy in the UI.
Why the campaign still looks fine
- Status reads ACTIVE at every level — nothing is paused, nothing is rejected.
- There is no error, because no rule was violated.
- Spend stays at zero, which reads as "hasn't started yet" rather than "can't start".
- Learning phase never begins, because learning needs events and no events are arriving.
The tell is spend, not status. An ACTIVE ad set that has not spent anything after a full day in a large market is not warming up. It is being asked for something it cannot deliver.
The ladder that works
Optimize for an event your dataset can actually produce today, and move up as signal accumulates. Each rung gives Meta more frequent events to learn from, at the cost of being further from the money:
- Landing page views — abundant from day one, weakest intent. Useful only to prove the tracking chain works end to end.
- Add to cart — mid-funnel, usually 5–10× more frequent than purchases, and a real signal of intent.
- Initiate checkout — closer to the money, still meaningfully denser than purchases.
- Purchase — the one you want, once the dataset has enough history for Meta to model.
The rough gate people use is roughly 50 optimization events per ad set per week. Below that the model has too little to work with, and delivery either starves or stays erratic. If your budget and your average order value cannot plausibly produce 50 purchases a week, purchase optimization is the wrong rung — not because your product is wrong, but because the arithmetic is.
A quick sanity check before you launch
- Does the dataset already hold events of the type you are optimizing for? If it holds zero, pick a lower rung.
- Daily budget ÷ expected cost per event × 7 — does that clear ~50 a week? If not, pick a lower rung.
- Is the event firing at all? Check it in Events Manager before spending, not after.
Cold start is not a bug you can configure around. It is the price of a new dataset, and the only way through it is to spend on an event that exists while the one you actually care about accumulates.