Hypothesis
If a low-volume campaign optimises for a more frequent upstream event instead of purchase, it will leave the learning phase and deliver purchases at a lower cost, because delivery gets enough signal; if the upstream event predicts purchases poorly, purchases will cost more. The decision behind it is concrete: optimise a low-volume campaign for purchases or for a more frequent upstream event such as add to cart.
Both halves of the hypothesis are live. Meta describes the Learning PhaseThe period after launch or a significant edit when an ad set's delivery is still being optimised. as the period in which the delivery system is still exploring how best to deliver an ad set, and the learning phase guide explains why a campaign with few purchases may never leave it and lists a more frequent event as one remedy. The remedy has a cost: delivery then finds people who add to cart, and in some shops those people buy at a steady rate while in others many are browsing, comparing or saving for later. Which is true for a given shop is an empirical question, and this test answers it.
Metrics and guardrails
The primary metric is cost per purchase in both arms: spend divided by purchases, with purchases counted the same way for both, ideally reconciled with the shop's own orders. CPACost of each conversion. on purchases is the only fair basis, because the upstream arm will always win on its own event. The CPA calculator puts both arms on the same footing. Where the winning event will also be used on other platforms, the guide to conversion events and values across ad platforms records it in a signal register against the result it stands for.
Three guardrails explain the result. Learning phase status shows whether the upstream arm achieved what it was meant to: if it is still learning at the end, the mechanism the hypothesis relies on did not happen. Purchases per optimised event, the share of add-to-carts that became purchases in the upstream arm, shows how well the upstream event predicts buying; a falling ratio means delivery is finding people who add to cart and stop, and the shop's normal Add-to-Cart RateShare of sessions in which a product was added to the cart. funnel is a useful reference point. Average order value is the third: an arm that finds cheaper purchases of smaller baskets may not earn more.
Design
Use the platform's A/B test tool with the optimisation event as the single variable. Build two ad sets that are identical in audience, creative, placements, budget, schedule, bid strategy and attribution setting; one optimises for purchase, the other for the chosen upstream event. Both arms track purchases with the same pixel and server events.
| Design element | Choice | Reason |
|---|---|---|
| Unit of randomisation | User, assigned by the platform's A/B test tool | Keeps each person in one arm |
| Arms | Optimise for purchase versus optimise for an upstream event | One change: the event delivery is told to find |
| Everything else | Identical audience, creative, budget and bid strategy | Isolates the event from other levers |
| Primary metric | Cost per purchase in both arms | The outcome the business pays for |
| Event tracking | Same purchase and upstream events, checked before launch | Miscounted events decide the test |
| Read window | After both arms leave or settle in learning | Learning-phase cost is not representative |
Choose the upstream event by how closely it tracks purchase in the shop's own data. Add to cart is common; initiate checkout sits closer to purchase and fires less often. The campaign objectives guide explains which events each objective allows, and the TikTok campaign structure guide shows where the optimisation goal sits in a TikTok ad group if the test runs there instead. Before launch, confirm that the event used is the same Key EventA GA4 event marked as important to the business, the property's measure of success. the business reports in analytics; the GA4 key events and conversions guide covers how those are defined, so the platform and analytics count the same thing.
Sample size reasoning
The test is read on purchases even though one arm optimises for something else, so the scarce event sets the sample. That is the uncomfortable centre of this design: the campaign has too few purchases to optimise for them, and the test needs enough purchases per arm to tell the arms apart. The sample size and significance guide shows how a low event rate stretches the run, and its section on tests that cannot answer the question applies directly.
If purchases are too few to separate the arms within a tolerable period, the honest answer is a longer test or a decision on judgement. A short test that reads noise as a result is worse than either, because it gives a false reason to change the account.
Duration
Run both arms through their learning phase, then at least two full weekly cycles. Extend rather than read early when purchases are sparse. Do not edit either ad set during the test; a significant edit returns that arm to learning and puts the two in different states. If the shop has a long consideration period, allow for purchases that arrive after the read window closes by fixing an attribution window before launch.
Reading the results and the decision rule
Read the test once, at the planned end. Compare cost per purchase between arms with an interval, then check learning status, purchases per optimised event and order value. Write the decision rule into the plan before launch:
| Result | Decision |
|---|---|
| Upstream arm has lower cost per purchase and order value holds | Switch to the upstream event; recheck when purchase volume grows |
| Upstream arm leaves learning but purchases cost more | Stay on purchase; the upstream event predicts buying poorly here |
| Upstream arm cheaper but order value falls | Stay on purchase, or weigh the margin before switching |
| No reliable difference | Stay on purchase and solve volume by consolidating ad sets |
| Neither arm leaves learning | Consolidate first; the event alone cannot fix the structure |
Consolidation is often the better remedy whatever the result, because one ad set with the whole budget gathers purchases faster than several with a share each. The academy capstone on running AI in production for a year at a children's shoe retailer practises applying this to a wide-fit range whose ad sets are merged before any event is tested, within a plan that keeps every platform test and its reading clear of the back-to-school season.
A ceramics shop tests add-to-cart optimisation
Wickerby Ceramics, an invented online pottery shop, records about 40 purchases a month from Meta and its ad set stays in learning limited. It runs an A/B test at £60 a day per arm for six weeks: one arm optimises for purchase, the other for add to cart. The purchase arm records 34 purchases at £74 each and never leaves learning; the add-to-cart arm leaves learning in its second week and records 41 purchases at £61, with order values within a few pounds. The interval on the difference is wide and crosses zero. Wickerby consolidates its three prospecting ad sets into one, keeps purchase optimisation, and plans to retest at the higher volume.
Pitfalls
Judging each arm on its own optimised event is the pitfall that the design exists to prevent, and platform reports invite it, because each ad set's headline result is the event it was given. Set the report's columns to show purchases and cost per purchase for both arms before launch. Too few purchases are the second: a test that starts with a volume problem cannot escape it, so plan for a longer run or a judgement call.
Pixel or server events miscounted are the third. If the add-to-cart event fires twice on some pages, or the purchase event is missing for some payment methods, the arms are judged on bad data. Test every event on the live site before launch and compare platform purchases with the shop's orders weekly during the test.
Common mistakes
Event tests fail on their reading more often than on their setup. The audience targeting breadth test uses the same A/B test structure to compare targeting, and the experiment library lists every design.
- Reading the upstream arm on add-to-carts. It wins by construction; read purchases.
- Changing audience or creative with the event. Two changes answer neither question.
- Reading inside learning. Calibration cost distorts early figures.
- Ignoring order value. Cheaper purchases of smaller baskets may not pay.
- Expecting the event to fix a fragmented structure. Consolidate ad sets first.
- Trusting events without testing them. A double-firing tag decides the result.
