Don't guess,test it.
Run two ad variants side by side on Meta. The scorecard calls a winner only once confidence passes 95% and each variant reaches its sample threshold.
Rather hand the job over? In our Google Ads Management service we run this tool for you; scope and monthly price are published.
Google Ads agencyHow an A/B test runs in the panel
You don't start from a blank form. Pick one of six ready templates; each brings a hypothesis and the metric to watch. You can rewrite both to fit your own test.
- 1Pick a template
Hook, creative format, call to action, offer, thumbnail or audience.
- 2Set the metric
Choose the success metric and, if you like, a guardrail metric with a threshold.
- 3Write the variants
Image, ad copy, headline and description for A and B.
- 4Go live
The two variants open on Meta as two separate campaigns.
- 5Measure
“Measure now” on the panel pulls Meta results for both campaigns; a test can be measured once every five minutes.
- 6Read the scorecard
Confidence, sample plan and the leading variant on one panel. You mark the winner.
Which statistical test each metric uses
Every metric has its own test. Rate metrics use a two-proportion test; cost and return metrics use the event rate per unit of spend. Because you can't know in advance which variant will lead, the test is two-sided.
| Metric | What it measures | Statistical test | Better direction |
|---|---|---|---|
| Click-through rate (CTR) | Clicks / impressions | Two-proportion z-test | Higher |
| Conversion rate | Conversions / clicks | Two-proportion z-test | Higher |
| Cost per click (CPC) | Clicks per unit of spend | Poisson rate test | Lower cost |
| Return on ad spend (ROAS) | Purchases per unit of spend | Poisson rate test | Higher |
For ROAS, significance is measured from the purchase rate; the spread of revenue itself can't be known from aggregated ad data, and the screen says so.
Meta explains how its own tool works on its about A/B testing page. Our scorecard answers the same question from your own data: is the gap real, or is it chance?
Four thresholds before a winner is declared
- 1Confidence must reach at least 95%. No test gives 100%; the screen caps the figure at 99.9%.
- 2Each variant needs at least 100 clicks on its own. A combined threshold would accept 199 clicks on one side and 1 on the other.
- 3Each variant must have spent at least $2.50.
- 4The test must stay live for the minimum duration you set. The default is 3 days.
The sample plan is sized to detect a 20% relative difference with 80% power. The scorecard shows how much sample you need, how much you have and how many days are left at the current pace.
Guardrail metric and decision log
A guardrail metric is the number that must not break even when the primary metric wins. If the winning variant crosses the guardrail threshold, the scorecard warns you before you scale it.
- The log records starts, pauses, winner picks, manual stops and scaling.
- Each entry keeps the confidence and sample size at that moment, so later data can't make a decision look right in hindsight.
- A decision made while the system said "not enough data" is flagged separately.
- Ending a test by hand asks for a reason; the last 40 decisions per test are kept.
Moving budget after a winner
The "Scale winner" button records the winner and takes you to the ad automation screen. You move budget there with a rule that has clear limits; budget is not shifted on its own.
New campaigns swing in their first days, so reading results then misleads (why you shouldn't touch a campaign in learning). If you plan to test on purchases, set up server-side conversion tracking first.
- If the brand has a test type it has never tried, that type is suggested next.
- Once every type has been tried, the one with the highest measured lift is suggested again.
- With no test history, you start with a hook test, the quickest to read.
Who it suits, and where it stops
| Situation | Is this screen a fit | Why |
|---|---|---|
| You're torn between two Meta ad texts or images | Yes | Variants differ by image, copy, headline and description |
| You want the same ad tested on two audiences | Yes | In an audience test, A runs on a broad audience and B on the interests you enter; the ad copy is shared |
| You want to test bids or landing pages in Google Ads | No | Tests open on Meta only; Google has its own experiments tool |
| Your budget can't spend $2.50 per variant | No | No winner is declared below the threshold |
A/B testing needs the Pro, E-commerce Standard, Plus Starter or Creator Pro plan (or a plan above one of them in the same family). Compare plans on the pricing page. To follow every campaign in one place, use the ad management screen.
Frequently asked questions
Which platform does A/B testing run on?
Meta, which means Facebook and Instagram ads. The two variants open as two separate traffic campaigns optimised for link clicks.
How do the variants differ?
By image, ad copy, headline and description. In an audience test the copy stays the same and targeting differs: A is a broad audience, B uses the interests you enter. If those interests can't be found on Meta, the test doesn't open. In other tests targeting is the same for both; if the brand has no target countries set, Turkey and ages 18-65 are used.
How do results come in?
While the test is live you press “Measure now” on the panel; Meta results for both campaigns are pulled and the scorecard is recalculated. A test can be measured once every five minutes. No winner is picked on its own and the campaigns aren't touched; you end the test.
How is the winner chosen?
By the success metric you pick. Confidence must pass 95%, each variant needs at least 100 clicks and $2.50 in spend, and the test must have run its minimum duration.
Why is the threshold per variant instead of combined?
A combined threshold would call 199 clicks on one side and 1 on the other "enough data". A per-variant threshold makes sure both sides are measured.
Can I end a test early?
Yes. You pick the winner by hand and write a reason. The decision is logged with the confidence and sample size at that moment, and flagged if the system said the data was not enough.
What does the guardrail metric do?
It protects a number that must not break even when the primary metric wins. If CTR rises while CPC crosses the threshold you set, the scorecard warns you.
Does budget move automatically once there's a winner?
No. "Scale winner" records the winner and sends you to the ad automation screen, where you move budget with a rule that has clear limits.
Which plans include A/B testing?
A/B testing needs the Pro, E-commerce Standard, Plus Starter or Creator Pro plan (or a plan above one of them in the same family). It is off on the free account; the check runs on the server.
What if my Meta account isn't connected?
The test stays in draft mode and no campaign opens in your ad account. Connect the ad account and Page, then restart the test to open both campaigns.
Sources
The official documentation behind the technical points on this page:
- [1]Meta Business Help Center: About A/B testing
How Meta's own A/B tool splits the audience and how results are read.
- [2]Google Ads Help: Set up a custom experiment
How campaign experiments are set up in Google Ads with Google's own tool.
- [3]Meta Business Help Center: About the learning phase
Why a newly launched campaign gives unstable results at first.
Related pages
- Ad automationMoves budget to the winning variant with rules that have clear limits.
- Conversion trackingSends purchase and lead events from the server to ad platforms.
- Ad managementOne screen to follow and manage every campaign.
- Google Ads management serviceCampaign setup, experiments and daily decisions handled by our team.
- Scene MultiplierScene multiplier turns your product and scene description into 5, 10, 15 images that show the same subject with different camera, light, setting, time of day and mood.
Stop guessing, measure the winner.
A/B testing needs the Pro, E-commerce Standard, Plus Starter or Creator Pro plan (or a plan above one of them in the same family). Compare plans on the pricing page.