Practical guide
What Is an A/B Test Used for on a Website?
An A/B test compares two versions of the same page or element across different visitor groups.
It is used to verify a specific hypothesis: does the change being tested actually alter the behaviour we measure?
Illustration
Same page, same measurement
Split traffic, one hypothesis, one observed event.
Variant A
ReferenceVariant B
1 changeMetric
Form submission
Split
Split traffic · 50 / 50
Status
Illustration
The principle
Two variants, one hypothesis, one shared measurement
An A/B test does not pit two randomly chosen pages against each other. It compares a reference version to a variant built to test a specific hypothesis.
The difference must remain sufficiently identifiable so the result can be interpreted.
Hypothesis
A specific change being evaluated.
Variant A / B
A reference and a version built to test.
Same audience
Traffic distributed without selection bias.
Same measurement
An identical event or indicator observed.
Comparison
Careful reading of the observed difference.
Before the test
Testing without a hypothesis mainly produces results that are hard to interpret
Observation
Step 1Few visitors reach the form.
Hypothesis
Step 2Presenting the contact point earlier could increase the number of started forms.
Test
Step 3Compare the current page to a variant where the contact point appears earlier.
An observation can first come from a heatmap that shows where behaviours concentrate, or from a broader analysis.
To understand how a user journey actually unfolds over time, a session replay can help identify a recurring blocker or hesitationbefore building a hypothesis.
Illustration · Example
Same page, one clearly identified difference
The reference page and its variant differ at only one main point. All other structure, copy and proof remain identical so the measurement can be interpreted.
Status
Test in progress
Result
Result not established
Type
Illustration
Good context
A test becomes useful when the question is specific enough
- Volume
A page receives enough traffic to be compared. Without sufficient volume, observed variations can be too unstable to draw conclusions.
- Measurement
A measurable action exists. Click, form started, request sent or other clearly defined event.
- Hypothesis
A hypothesis already exists. We know what is being changed and why we think it could modify behaviour.
- Stakes
The change is genuinely worth testing. A low-stakes micro-detail does not always need formal experimentation.
Poor context
Sometimes, fixing directly is more rational than testing
- Bug
Obvious bug. A broken button or a form that returns an error does not need an A/B test.
- Low traffic
Too little traffic. A rarely visited page may require another observation method.
- Too many variables
Too many changes at once. If the variant modifies the title, design, form and offer, the result becomes difficult to attribute.
- Vague question
Poorly framed question. "Is this page better?" is not a sufficiently precise hypothesis.
Interpretation
An observed difference must be interpreted with caution
Result ≠ definite cause
A variant can perform better without us knowing exactly which psychological mechanism explains it.
One period ≠ all periods
Campaign, season, traffic source or audience composition can change.
One metric ≠ the whole journey
Increasing clicks on a CTA is not useful if request quality or the next step degrades.
Four levels of reading
Analytics, heatmap, session replay and A/B test do not answer the same question. Each tool has its place in a progressive reading.
Analytics
Level 1How many?
Measures volumes, sources and overall trends in a user journey.
Heatmap
Level 2Where?
Shows where clicks, scroll and attention concentrate on a page.
Session replay
Level 3How?
Replays an individual user journey to understand blockers and hesitations.
A/B test
Level 4Does the change produce a better result?
Compares two variants built to evaluate a specific hypothesis.
Scope
A test can cover far more than a button colour
Information order
Title or value proposition
Proof placement
Landing page structure
Form length
Field order
CTA
Intermediate journey step
Mobile display
Functional component variant
Execution
Experimentation quickly becomes a technical problem
Creating a variant properly does not always mean changing a few lines of text. Sometimes conditional rendering, events, variant persistence, tracking or a deeper interface difference must be managed.
A marketing team can define the hypothesis and the metric. Atomic Design can handle the technical part needed to build the variant and correctly measure the experience.
More advanced case
Some experiments require more than a page variant
When an experiment touches a feature or multiple interface states, a feature flag approach can be more suitable than a simple content change.
This type of implementation makes it possible to control who sees a variant, disable it quickly, and more cleanly separate technical deployment from marketing evaluation.
Related pages
Related pages
Some supplementary reading to extend the thinking without replacing a discussion about your situation.
What Is a Heatmap Used for on a Website?
Where do behaviours concentrate to build a solid hypothesis?
Open page →What Is Session Replay Used for on a Website?
How does the journey actually unfold to understand a result?
Open page →How to Track Conversions Correctly with GA4 and GTM
Correct conversion measurement is required for an A/B test to be interpretable.
Open page →Practical intervention
Have a hypothesis to test but not the implementation?
We can build the variant, connect the necessary measurement and verify the test rests on a technically comparable journey.