In an A/B test, the control is the existing version of a page, email or app screen, and the variant is the changed version you are testing against it.
Visitors are randomly split between the two, and the control's result is the baseline that tells you whether the variant helped, hurt or made no difference.
What "variant" means in an A/B test
A variant is any version of the experience that a visitor can be assigned to.
In everyday CRO talk, "variant" usually means the new version (version B), while "control" means the current one (version A).
Ron Kohavi's Practical Guide to Controlled Experiments on the Web uses stricter terms.
It calls the existing version the Control and the new versions Treatments, and treats "variant" as the word for either one.
You will see "treatment," "challenger" and "variation" used for the same idea in testing tools and papers.
The guide also explains why the split matters: if the experiment is set up properly, the only consistent difference between the groups is the change itself.
So a statistically significant gap in results can be pinned on the change rather than on timing or traffic source.
How control and variant work in practice
- One change per variant, ideally. If the variant changes the headline and the image, a win tells you the package worked but not which part did.
- The control stays untouched. Editing the control mid-test breaks the comparison.
- You can run several variants at once. An A/B/n test pits two or more variants against one control. Each extra variant splits traffic further, so each needs more time to reach a reliable answer.
- The control is your fallback. If a variant has a bug, everyone goes back to the control.
- Control vs control is a check, not a test. An A/A test shows the same version to both groups to confirm your setup splits traffic and counts conversions correctly.
Results are reported as the variant's lift over the control.
For example, a variant at 3.3% conversion against a control at 3.0% is a 10% relative lift.
The control's rate is the baseline conversion rate for that test.
Real examples from the database
Kasper Dolk tested a benefits-first hero against the control hero.
Revenue per visitor rose from €1.20 on the control to €1.47 on the variant, a 23.05% increase.
Dave Diederen ran a three-way test on the first product image, with a blonde model and a brunette model as two variants against the control image.
The blonde variant converted at 9.55% against the control's 9.32%, while the brunette variant came in at 9.01%, below the control.
One test, one winning variant and one losing variant.
The control does not always lose. Blend Commerce replaced a subscription frequency dropdown with pill buttons on a coffee brand's product page, and the variant lost overall with a 7% drop in conversion rate.
It did win on desktop, so the brand shipped it on desktop only.
That is why the control exists: without it, the drop would have looked like a normal slow week.
@DaveDiederen
@kasper_dolk
Blend CommerceControl vs variant at a glance
| Control | Variant | |
|---|---|---|
| What it is | The current version | A changed version |
| Other names | Version A, baseline, champion | Version B, treatment, challenger, variation |
| How many per test | One | One or more |
| Role in results | Sets the baseline | Measured as lift against the control |
Related terms
See lift for how the difference is calculated, A/B/n testing for tests with several variants, and how to run an A/B test for the full process.
Browse every term in the glossary, or filter all test results to see more variants that beat, matched or lost to their control.