Find the widget that converts best.

1. Pick 2 to 4 widgets. They must be active and belong to the same campaign. Their targeting (countries, languages, devices, browsers, schedule) has to match so every variant reaches the same visitors. Page rules and triggers can differ, with a warning.
2. Choose what counts as a win. A widget action or a conversion URL, explained below.
3. Choose whether to measure incrementality. Optional, off by default.
4. Set the length and the ending. Pick 7, 14 or 30 days, or an end date, and what happens when the test ends.

Before you start, a summary restates everything in one paragraph. While the test runs, each tested widget carries an A/B Test badge in your widget list, linking to the results.

A widget can be in one running test at a time. Changing a tested widget mixes old and new results, so the editor warns you. Pausing or deleting a tested widget ends the test.

Widget action

A click on the widget's button, an email or a form submitted, or feedback given. Closing the widget never counts as a win. We record closes separately and show the share of visitors who closed each widget.

Conversion URL

A visit to a page such as your thank-you or order confirmation page. Enter one or more addresses, one per line. A page counts when its address contains your text, capitals ignored, so merci/ matches /fr/merci/. The GrowthTools pixel must be installed on that page. Add an average value per conversion to see results in money.

A test tells you that B beats A. It cannot tell you whether either one brings anything at all. Maybe those visitors would have bought anyway.

Turn on the control group and 10% of visitors see none of the tested widgets. Say 5 in 100 visitors who see a widget buy, against 3 in 100 who see nothing. Your widgets add 2 sales per 100 visitors. Those are your incremental conversions.

With an average value per conversion, the top of your dashboard shows one line: This month: +$7,965 incremental revenue (estimated). It is an estimate from your own numbers, and says so. When the data is too thin, it says too early to tell instead of showing a figure.

Each visitor gets a random identifier and is assigned a variant once. They see that same widget on every page and every visit, so results are not mixed.

The chance to be best comes from Bayesian statistics. "96%" means that, given your data, B is the best widget in 96 out of 100 plausible scenarios. It is easier to read than a p-value.

Early numbers move a lot, so we hold back. A winner is named only when all three conditions are met:

Until then the page reads Not enough data yet, or No clear winner yet when two widgets are close. When you create the test, we estimate how long you need from your recent traffic and warn you if your chosen duration looks too short.

You choose the ending when you create the test, and can change it while the test runs.

Pause the losing widgets

The winner keeps running alone and the test closes. Without a clear winner, nothing is paused, the test keeps running and the email tells you so.

Keep the test running

Same split, same control group, and the results keep updating. You can stop it any time.

Pause all tested widgets

Always applied at the end date, winner or not.

Then we email you: the verdict in one sentence, the results table, the incremental figure if you used a control group, and the action we applied (for example "As requested, we paused Variant A. Variant B keeps running."). If visitors without a widget converted more than those with one, the email says so plainly.

On the results page, Keep this one picks a winner by hand and pauses the others. Stop test ends the test and pauses all tested widgets.

The test stores a random identifier in the visitor's browser, on your own site. It holds no personal data, it is never shared between sites, it expires after 13 months, and it is used for statistics only. You stay responsible for your cookie banner.

If you prefer, tick Wait for consent on a test. Your banner then calls growthtools('consent', true) when the visitor agrees. Until then, visitors see Variant A and are not counted.

Finished tests do not count toward the limit.