A/B testing splits your traffic between versions of your cart drawer and reports which performs better. Unlike bundle testing, it has no product targeting requirement: any saved drawer can be tested. By the end of this you will have a test running and will know how to read the confidence indicators before deciding anything.
In this article:
Before you start
The drawer must be saved. If it is not, you will see A/B testing is disabled, save the cart upsell to create A/B testing variants.
You also need traffic. The results view is explicit about how much: at least 1,000 views for reliable results.
Step 1: Enable A/B testing
In the cart drawer editor, select
A/B testingin the left navigation, badgedNew.Select
Enable A/B testing.
Atlas confirms with A/B testing enabled and active in the store, proceed to edit or create variants.
Step 2: Create a variant
Select Add variant. The variant is a copy of your drawer that you then modify. They appear under A/B test variants, alongside the Original cart upsell.
Change one thing. A variant that differs in its threshold, its upsell products and its colours all at once will tell you that something worked without telling you what.
Step 3: Let it run
Atlas splits traffic automatically. There is nothing to configure.
A test can be paused, in which case you will see A/B testing is paused. All variants are in draft mode. with Resume A/B test to restart it.
⚠️ Do not edit a drawer while it is in a test. Changing the thing being measured makes the results before and after incomparable, and there is no way to separate them afterwards.
Run it across at least one full week. Weekend and weekday shoppers behave differently, and a test that ran Tuesday to Thursday can point the wrong way.
Step 4: Read the results
Open A/B test analytics to compare variant performance and identify the winning configuration for your cart upsell.
The headline. When a variant is ahead: {Variant} is your winner! with Expect uplift of ... over control, quantified in conversions and revenue.
The confidence indicators, which are the most valuable part:
Indicator | What it means |
| Results are statistically significant with 95%+ confidence. You can trust this winner. |
| Results show a trend but haven't reached statistical significance yet. Keep the test running. |
| You have enough data to make a reliable decision. |
| At least 1,000 views are needed for reliable results. Continue running the test. |
The comparison table shows per variant: Impressions, Checkouts completed, Conversion rate, Revenue per visitor, and Uplift against the Baseline. Other panels cover Revenue per type, Upsell products by variant, Traffic split, Sample size and how many Days the test has run.
💡 Wait for Significant before acting. A variant that is ahead while still Trending is as likely to be noise as a real difference, and acting early is how you convince yourself of things that are not true.
Step 5: End the test
Select
End A/B test.In the modal, review each variant's
Impressions,CheckoutsandConversion rate.Select
Chooseon the winner.Select
End test.
What ending a test destroys
⚠️ Once you select your winning variant, all other variants will be deleted and all traffic will be directed to the selected variant. The losing variants are gone permanently.
Write down every variant's numbers before you end the test, including the losers. Knowing that a $75 threshold earned less per visitor than $50 is worth as much next quarter as the winner is today, and Atlas will not keep that comparison for you.
What to test
Highest-value tests first, based on what changes behaviour most.
Test | Why it matters |
Reward bar threshold, for example $50 against $75 | Directly moves average order value in both directions |
Reward bar on against off | Tells you whether the module earns its place at all |
Upsell products, manually chosen against AI recommended | Answers a question you cannot reason your way to |
Upsell position, above items against below | Small change, sometimes surprisingly large effect |
Shipping protection price | Take-up against revenue per order |
Checkout button text with the total against without | Cheap to test, sometimes meaningful |
ℹ️ Judge on revenue per visitor rather than conversion rate. A variant with a lower threshold will convert better and can still earn you less per visitor, because you gave away shipping on smaller orders.