An A/B test changes the page a visitor is already on and tells you which version did better. You make the change by pointing and clicking. No code.
Every test is built in six steps. You can close a test halfway through and come back to it: what you have filled in is saved as a draft.
Before you start, the tag has to be on your site and at least one metric has to exist. See Install the tag and Metrics.
The six steps
| Step | Name | What you do there |
|---|---|---|
| 1 | Setup & hypothesis | Name it, say what you expect, pick the page it runs on. |
| 2 | Traffic & variants | Make your variations and split the traffic between them. |
| 3 | Goals & metrics | Choose the number that decides it. |
| 4 | Audience & targeting | Choose who sees it. |
| 5 | Configuration | Schedule, traffic sharing, connected tools. |
| 6 | Preview & confirm | Look at it on the real site, then start it. |
How to set one up
- Setup & hypothesis. Type a name and the page the test is built on. That page is what the editor opens, so use a real address from your own site.
- Traffic & variants. Save the step first, then click Visual editor on a variation. It opens your own page with editing panels around it. Make your change, and it saves against that variation.
- Goals & metrics. Pick the one metric that decides the test, and up to five more to watch. Add guardrails if you want to be told when something else gets worse.
- Audience & targeting. Leave it empty to run on everybody, or pick an audience.
- Configuration. Optional, all of it. No schedule means you start and stop it by hand.
- Preview & confirm. Open the preview link and look at each variation on the real site. Then press Start A/B test.
The Visual editor button is disabled until step two is saved, because there is no variation to write against yet.
What each label means
| Step | Label | What it means |
|---|---|---|
| 1 | Experiment name | What you will recognise it by. Up to 120 characters. |
| 1 | Hypothesis statement | What you expect to happen and why. Up to 1000 characters. Write it before you look at a result. |
| 1 | Tags | For finding it later. Up to ten, each up to 40 characters. |
| 1 | The page this test is built on | The address the editor opens. It has to be on a domain you have added. |
| 1 | Include pages, Exclude pages | Where the test runs, beyond that one page. Leave both empty to run on that page alone. Up to twenty rules. |
| 1 | A URL to check against these rules | Paste an address and it tells you whether the test would run there. Nothing is saved. |
| 2 | Share of matching traffic, percent | How many of the visitors who match get into the test at all. The rest see your page untouched. |
| 2 | Variation name | Shown on every report. Up to 80 characters. |
| 2 | Split | What share of the test each variation takes. |
| 2 | Baseline | The variation the others are read against. The original, unless you move it. |
| 3 | Goal | The one metric the test is judged on. |
| 3 | Guardrails | Metrics watched for harm. They never decide the winner. |
| 5 | Start it myself, Start at a time | Whether it waits for a time you set. |
| 5 | Run until I stop it, Stop at a time | Whether it stops itself. A stop time cannot come before a start time. |
| 5 | Mutual exclusion | A group of tests that share traffic, so a visitor sees only one of them. |
| 5 | Integrations | Which of your connected tools this test sends to. |
A test needs at least two variations and takes at most ten. The share of matching traffic goes down to 0.1%.
Before it can start
Step six lists what has to be true. Some lines stop you, others only warn you.
| Check | Stops you starting |
|---|---|
| Snippet found on your site | Yes |
| Every change resolves on the page | Yes |
| One goal decides the test | Yes |
| A person has looked at the variation | Yes |
| The group has enough free traffic | Yes, if the test is in one |
| Text changes are scoped to the right pages | No, a warning |
| No other running test touches the same elements | No, a warning |
| Goals are receiving data | No, a warning |
| Consent mode | No, for information |
"A person has looked at the variation" means you. Open the preview link and see it on the real site. A test nobody looked at is how a broken layout reaches half your visitors.
Take the warnings seriously anyway. "Goals are receiving data" failing means the event under your goal has not arrived in a day. The test will run and answer nothing. Check it on the Events screen first.
Good to know
- Each visitor is placed in one variation and stays in it. Somebody who comes back on the same browser sees the same version every time.
- A change that resolves nowhere is a variation that changes nothing while still taking its share of the traffic. That is what the "Every change resolves" check is for.
- If the test runs on more than one page, the scope warning matters. A text change written against one page can match something you did not mean on another.
- Once it starts you can pause it and resume it. A paused test keeps the visitors it already placed, so resuming does not reshuffle anybody.
- Write the hypothesis at step one and leave it alone. It is the record of what you expected, and it is worth much less written after you have seen the numbers.