Meta ad creative testing

What to test

One variable at a time

If two ads differ in image, headline and offer, a difference in results cannot be traced to any one of them. Change one thing per test. Start with bigger questions, such as which concept or offer works, before testing smaller details like button wording. Big differences are easier to detect and usually matter more.

Setting up a fair test

Meta's split testing divides an audience into groups that do not overlap, so each version is seen by different people under similar conditions. Simply putting several ads in one ad set lets Meta's delivery favour one early, which is useful for performance but less reliable as a test. Decide which approach you need: a clean comparison to learn from, or letting delivery choose to get results quickly.

A testing cycle

1. 01

   Question

   One thing to learn.

   Output: Test hypothesis

2. 02

   Measure

   The result that decides it.

   Output: Success measure

3. 03

   Build

   Versions differing in one way.

   Output: Test ads

4. 04

   Run

   Until enough results arrive.

   Output: Test data

5. 05

   Record

   What was learned and next test.

   Output: Test log

Choosing the measure

Decide before launch what counts as winning: cost per purchase, cost per qualified lead or another outcome that matters to the business. Click-through rate is easy to move but may not lead to sales. Meta's guidance is to set the measures and confidence level before the test starts, so the result is not chosen after the fact.

Example: testing two concepts

Testing checklist

How long to run

Run a test until each version has enough results to compare with some confidence, not for a fixed number of days. With few conversions a day, that may take weeks. Stopping early because one version leads can mean picking a winner by chance.

Creative fatigue

Even winning ads tire as the same people see them repeatedly. Rising cost per result on an unchanged ad is a sign. Testing is therefore continuous: each round finds the next version before the current one wears out.

Keeping a test log

Record each test's question, versions, measure, result and what you will do differently. Over time the log becomes the most valuable thing your testing produces, because it tells you what your customers respond to.

Audience size and tests

Small audiences produce results slowly, so tests take longer. If your audience is small, test bigger differences, accept longer test periods or use a measure that arrives sooner, such as add-to-basket, while checking it relates to sales. Running many tests at once on a small audience splits the data further and slows every one of them.

Claims in test versions

Every version must be accurate. Test different angles, not claims you cannot support.