Reviews
Part of Reading ecommerce conversion optimization with a skeptical eye
Common ecommerce conversion optimization questions: denominators
Common ecommerce conversion optimization questions: which denominator to use for session, user, cart, and checkout rates, and how to record the answer.
What to take away
- A conversion rate is meaningless until numerator and denominator are explicit.
- Research and experiments answer different questions.
- Do not test whether the store should keep a known harmful defect.
Common ecommerce conversion optimization questions often use the same word for different events. A click, selected variant, payment, accepted order, retained order, and repeat purchase are not interchangeable. Define the state and eligible opportunity before comparing rates.
Which denominator should be used?
Google Analytics' dimensions and metrics reference distinguishes user key event rate from session key event rate and defines each within its product. These are analytics calculations, so the team should preserve event, identity, consent, filter, and session settings and connect them to merchant outcomes.
For an ecommerce purchase-conversion rate, use eligible shopping sessions as the practical default when the question is whether a shopping visit produced an accepted order. Use eligible users when the question is whether a person purchased across visits; use carts for cart-to-order conversion, and checkout starts for checkout completion. An eligible opportunity is a defined session, user, cart, or checkout that could reach the stated outcome under the report's inclusion rules; set those rules before comparison and apply them consistently. The denominator should match the outcome and unit you intend to measure, not whichever option produces a more favorable rate.
Consider a shopper who views a product, leaves, then returns before ordering. A session rate counts eligible visits as the unit, while a user rate counts the person across visits; a checkout-completion rate starts with checkout starts, not all storefront visits.
Formulas for the common rates
Name the acceptance rule first, such as "paid and not cancelled for fraud", and count each order once under it. Then use these denominators:
- Purchase conversion rate, session:accepted orders divided by eligible sessions.
- Purchase conversion rate, user:users with an accepted order divided by eligible users, in the same window.
- Add-to-cart rate:sessions with at least one add-to-cart divided by product-view sessions. Do not use all sessions, because many visits never view a product.
- Cart-to-order rate:accepted orders divided by carts created.
- Cart abandonment rate:carts not converted divided by carts created. It equals one minus the cart-to-order rate.
- Checkout completion rate:accepted orders divided by checkout starts.
- Checkout abandonment rate:checkout starts without an accepted order divided by checkout starts. It equals one minus the checkout completion rate.
- Email signup rate:signups divided by sessions that reached the signup form.
- Revenue per session:accepted revenue divided by eligible sessions.
- Average order value:accepted revenue divided by accepted orders. This is an average, not a rate, so its unit is the order.
Use the same window and the same acceptance rule for numerator and denominator. Mixing a session numerator with a user denominator is an easy error to make, and it produces a rate that looks plausible but measures nothing you intended.
User vs Session Key Event Rate
User key event rate
- Denominator
- Users
- Event counted
- Key event
- Scope
- Per user
- Settings
- Identity, consent, filters
- Use
- Merchant outcomes
Session key event rate
- Denominator
- Sessions
- Event counted
- Key event
- Scope
- Per session
- Settings
- Session settings
- Use
- Merchant outcomes
| Question | Working answer |
|---|---|
| Should every visitor convert? | No. Only eligible people with the relevant task belong in the opportunity. |
| Is more checkout starts better? | Only if payment, acceptance, returns, and customer outcomes hold. |
| Does significance prove value? | No. Effect size, uncertainty, costs, risks, and durability matter. |
| Can research replace testing? | Research explains behavior; a controlled test can estimate a specific change. |
| Who owns CRO? | Product, design, research, engineering, analytics, commerce, service, and business owners share evidence. |
What belongs in an experiment plan?
NIST's statistical handbook sets out a planning workflow for experiments, from objectives and variables through design, runs, assumption checks and analysis, and advises keeping all the raw data. It is general statistical guidance, so teams should choose a design with appropriate expertise for the actual decision.
Use a decision card for each test. It holds the customer task, current evidence, the change, the mechanism, the unit, eligibility, the primary outcome, guardrails, the sample plan, the duration, the stopping rule, implementation checks, the analysis, the owner, and action thresholds. Record why an A/B test is or is not suitable.
Known legal, accessibility, security, price, and customer-control defects should be corrected promptly. Testing can compare safe remedies, but a harmful version should not be preserved solely to create a control.
When evidence is weak, choose a reversible pilot and state uncertainty. A declared unknown is safer than a confident universal benchmark.
Revisit these answers when the store changes its catalog, checkout, markets, consent settings, payment methods, or definition of an accepted order.
Resolve the question with a decision record
Consent settings change the counts that every rate depends on. Google's Consent Mode changes how analytics and ad tags behave when a visitor declines cookies, so the session and user totals can shift without any change in shopping behavior. Record the consent setting with each rate.
A useful answer identifies the decision owner, the eligible population, the evidence period, the method, and the action. It also states what the result cannot establish. Save the definition and date with the answer. If the same question returns, compare the records before changing the metric. That step exposes a changed population or source instead of treating every difference as new behavior.
Store the evidence with the decision it produced so a reviewer can reproduce the reasoning without memory. Date each answer's next review and name the change that would justify an earlier check.
Keep rejected options on record with the chosen path; the constraint that ruled them out may change. One set of definitions serves the interface, export, meeting note, and correction log; retire unused reports and permissions when the decision cycle ends.
Common questions
What is a good conversion rate?
There is no universal rate. Task, product, audience, device, market, denominator, and order maturity differ, so the useful comparison is the store's own rate for the same task and measurement window. If an external benchmark range is used, record its source, definition, sample, and date before treating it as a target.
How long should a test run?
Run until the sample set in the plan is reached and at least one full weekly cycle has passed, so weekday and weekend shoppers both count. The sample size depends on the baseline rate and the smallest change worth acting on. Your testing tool or a standard sample-size calculator can compute it from those two inputs.
Allow downstream outcomes such as returns time to mature before you judge the test. Set the stopping rule before launch, and do not act on a favorable peek at the data.
Can CRO reduce sales?
A clearer experience can reduce unsuitable orders while improving correct selection, retained value, service load, and trust.
