Conversion · Measurement

Can add-to-cart rate pick a landing-page test winner?

By Mirow · November 6, 2026 · 6 min read

A higher add-to-cart rate is useful evidence about a page's buying path. It is not enough, by itself, to name the winner of a test intended to increase sales. Keep the purchase outcome as the decision metric and use cart activity to find out what changed along the way.

A more visible button can make adding an item easier without resolving doubts about fit, delivery or total cost. When purchases are sparse, the tempting move is to promote the metric that changes fastest. That can turn a useful diagnostic into an unsupported sales claim.

Name the decision before the metric

For a sales-focused page test, write the decision as: “Should we replace the current page because the new version produces a worthwhile purchase improvement?” Then define a purchase rate using the people assigned to each version, including those who leave without adding anything.

Keep two jobs separate in the test brief:

• Choose the page: the preselected purchase outcome, its uncertainty and the business limits that would make rollout acceptable.
• Explain the buying path: cart additions, checkout starts, errors and other steps that help locate friction.

This distinction follows the metric problem described in Microsoft's experimentation discussion: improving an early step need not improve the outcome farther along the process. The ecommerce application here is Mirow's recommendation, not a finding about your store.

If your immediate task is simply to repair a broken cart button, verify that repair directly. You do not need to pretend an obvious functional fix is a demonstrated conversion lift. If you are still choosing what kind of page to build, start with the product-page versus landing-page decision.

Count people consistently

Google Analytics defines separate events for adding an item, beginning checkout and making a purchase. Its ecommerce implementation guide describes when those events should be sent. An event count is not automatically a count of distinct buyers or visitors.

For the example below, the unit is one identifiable experiment visitor, assigned once and kept in the same version. The two rates are:

• Cart-add rate: assigned visitors who add an item at least once ÷ all assigned visitors.
• Purchase rate: assigned visitors who complete at least one purchase ÷ all assigned visitors.

Repeated additions do not create extra cart-adders. Repeat orders do not create extra purchasers in this binary purchase-rate measure. Revenue and order counts can be recorded separately. Choose an identity and assignment method the test system can actually support; browser identities may split one person across devices.

Before launch, confirm that both versions record the same actions, keep the assigned version attached to the outcome, and measure the same eligible population. Do not compare one version's sessions with the other's users, or divide raw purchase events by raw add-to-cart events. Check purchase duplication too: Google documents transaction-ID deduplication, with a unique ID per order. That does not replace visitor-level counting for this test.

Read the whole example

Consider a hypothetical test of a sticky add-to-cart button, unrelated to any client. Eligible visitors are randomly assigned to the current page or the challenger. Each gets a seven-day observation window from assignment; every visitor included below has completed that window. Seven days is an example definition, not a universal recommendation. Pick the real window before launch from the product's buying cycle and measurement constraints.

In this simplified example, every purchaser previously added an item, all events are observed, and each visitor is counted at most once in each row. Price, offer and checkout stay the same. The numbers illustrate how to read the metrics; they are not a forecast or a sufficient sample-size rule.

Scroll horizontally to see all columns.

The cart-add rate is six percentage points higher, or 60% higher relative to the current page: (16% − 10%) ÷ 10%. The purchase-rate difference is 0.2 percentage points, based on two additional purchasers. “Cart additions rose 60%” cannot be rewritten as “the page sells 60% more.”

The last row also needs care. The button can change who enters the cart, so the cart-adders are no longer the original randomized groups. Comparing only those subsets cannot establish that checkout became worse. The observed ratio falls, but several explanations remain possible: more undecided visitors enter the cart, an implementation issue changes the measurement, or random variation affects the counts. Inspect those possibilities before assigning a cause. See the changing-denominator examples in Microsoft’s metric-interpretation paper.

Keep all assigned visitors in the primary purchase comparison. Use the analysis method and stopping rule chosen before the test to assess the 22-versus-20 result. The table alone establishes neither a purchase winner nor that the versions are equivalent.

Turn the signal into a decision

The following rules are a practical way to separate repair work, continued observation and rollout. Set the acceptable downside and the smallest worthwhile purchase improvement before looking at results.

Scroll horizontally to see all columns.

A purchase-rate win can still be commercially weak if orders become smaller, more heavily discounted or more costly to serve. Review net sales after discounts and refunds, variable order costs and contribution per assigned visitor using consistent definitions. Record unavailable costs as unavailable. A page decision should not depend on an assumed profit improvement that the test never measured.

When purchases will not answer in time

Plan sample size for the purchase effect that matters, rather than the easier-to-move cart metric. If the available traffic cannot support that decision within an affordable period, narrow the comparison, choose a more substantial change, or use usability work to identify a specific repair. None of those choices creates evidence of sales lift automatically.

A cart metric can help prioritize what to investigate or which idea deserves a later purchase test. Using it as a substitute rollout criterion would require evidence that changes in that metric reliably predict the purchase outcome for comparable changes. A correlation between cart activity and sales on ordinary days does not supply that experiment-level validation.

Before the next page test, add four lines to the brief: the purchase decision, the visitor denominator and outcome window, the cart signal that will help diagnose the result, and what you will do if purchases remain inconclusive. For the measurement checks behind that brief, use the ad-to-page diagnostic. The useful result may be a better question for the next test, even when there is no winner to roll out.