Quick definition: A conversion funnel is an ordered model of the eligible steps people or accounts take from an initial opportunity to a defined outcome, with counts and rates at each step.
What is a conversion funnel?
A conversion funnel maps a customer journey into measurable stages. An ecommerce example may be eligible product view, add to cart, checkout start, payment attempt, and paid order. A B2B product may use signup, workspace creation, data connection, first successful task, and paid subscription. The funnel helps locate where people leave, but it is a model—not proof that every person follows one linear path.
Every stage needs an event, unit, timestamp, and eligibility rule. A funnel is only interpretable when step order and time window are explicit. “Checkout-to-purchase conversion within 48 hours for unique sessions that started checkout” is a measurable stage transition. A list of unrelated page views is not. Define whether a person may enter in the middle, repeat steps, use different devices, or complete the outcome without the conventional route.
Funnel conversion differs from overall conversion rate. Overall conversion compares a final outcome with a broad eligible population. Step conversion compares one stage with a prior stage. Click-through rate measures response to an impression, and completion rate measures finishing a specific process. A high late-stage rate cannot compensate for a large loss earlier in the funnel; the overall outcome depends on both volumes and conditional rates.
Funnel formulas and denominators
For consecutive stages A and B, calculate:
step conversion A→B = unique eligible units reaching B after A / unique eligible units reaching A × 100
For the overall funnel, use final conversion = unique eligible units completing final outcome / unique units entering the defined funnel × 100. Use the same unit consistently. Sessions may be appropriate for a short web purchase flow; accounts may be appropriate for a multi-user setup journey. Mixing users in one stage with events in another creates rates that cannot be multiplied or compared.
| Funnel question | Useful denominator | Risk |
|---|---|---|
| Where is checkout lost? | Unique checkout starters | Does not explain earlier traffic loss. |
| Does landing experience create customers? | Eligible assigned visitors | Needs full attribution window. |
| Does onboarding lead to value? | New eligible accounts | May need account-level identity resolution. |
| Does an email drive purchase? | Delivered eligible recipients | Delivery and scanner traffic need controls. |
Sequential funnels impose order. If customers can buy before creating an account, a strict signup-before-purchase funnel will falsely classify valid paths as drop-off. Consider a flexible path analysis or a separately defined outcome when the journey has legitimate branches.
Conversion funnels in A/B testing
Funnels are valuable experiment diagnostics. A new checkout field may lower checkout completion, while a new product page may affect add-to-cart rate. Select one decision-relevant primary outcome, such as purchase per assigned eligible user, then use stages to explain its mechanism. Do not choose a favorable intermediate stage after seeing the data and call the test successful if the final outcome did not improve.
Analyze the final rate among everyone randomized and eligible, including people who do not enter later steps. Restricting analysis to checkout starters after testing a product-page change is post-treatment selection: treatment might alter who starts checkout. Conditional stage rates are useful descriptive diagnostics, but they do not generally estimate the causal effect of the variant on the original population.
Predefine the funnel window, stage order, and important guardrails. A redesign could improve purchase but increase payment errors, refunds, page latency, or support contacts. It can also move traffic between channels rather than create incremental outcomes. Our primary and guardrail metrics guide describes how to retain a decision focus. For credible results, check allocation and exposure integrity as described in the sample-ratio mismatch guide.
Worked funnel scenario
A retailer tests a guest-checkout design among 20,000 assigned eligible sessions per arm. Control has 8,000 add-to-cart sessions, 4,000 checkout starts, and 2,400 paid orders. Treatment has 8,200 add-to-cart sessions, 4,510 checkout starts, and 2,620 paid orders.
control cart rate = 8,000 / 20,000 = 40.0%control checkout completion = 2,400 / 4,000 = 60.0%control overall purchase = 2,400 / 20,000 = 12.0%treatment overall purchase = 2,620 / 20,000 = 13.1%
Treatment’s overall lift is 1.1 percentage points. The stage view suggests it improved checkout starts and retained a similar purchase-through-checkout rate. The team nevertheless checks average order value, payment declines, cancellations, fraud, shipping cost, and whether the event that records checkout start changed with the design. It estimates uncertainty around the primary purchase rate rather than multiplying rounded conditional rates to claim precision.
Funnel data-quality caveats
Funnel data frequently breaks at handoffs: client events may precede server confirmation, identity can change after login, and a user can return on another device. Use stable IDs where possible, reconcile final transactions to a system of record, and document the identity stitching rule. Deduplicate retries and retries caused by errors. A final order should not be counted twice because a payment webhook and confirmation page both emit events.
Event semantics matter. “Checkout started” might mean clicking a button, loading a page, or successfully creating an order draft. If one variant changes the UI, the old client event can become incomparable. Instrument a semantic backend event or version the event carefully. Validate each stage with counts, null rates, timestamps, and plausible ordering before reading experiment results.
Traffic mix and time affect funnels. Mobile and desktop paths, countries, stock availability, campaign intent, bots, and page performance can alter stage rates. Randomization reduces many differences but does not repair missing or variant-specific tracking. Segment responsibly and avoid declaring a winner from an unplanned stage among many examined cuts; multiple comparisons explains the risk.
Common conversion-funnel mistakes
- Assuming every journey is linear: valid alternate paths can look like abandonment.
- Mixing units across stages: events, sessions, users, and accounts are not interchangeable.
- Using conditional stages as the primary causal result: later-stage entry can be treatment affected.
- Changing event semantics with the UI: apparent lift may be instrumentation.
- Ignoring time windows: delayed conversions disappear from recent cohorts.
- Optimizing one stage alone: faster entry can create lower-quality outcomes later.
Frequently asked questions
How many stages should a funnel have?
Use the fewest stages that represent meaningful decisions or handoffs. Too many noisy micro-events obscure the decision; too few hide the constraint.
Can funnel rates be multiplied?
Only when stages use compatible units, a common population, and sequential rules. Prefer calculating the final rate directly from eligible entrants.
Should an A/B test optimize the narrowest drop-off?
Not automatically. Improve the step only when it is causally connected to the final customer and business outcome and guardrails remain healthy.
What is funnel leakage?
It is the share that does not progress between defined stages. It can reflect choice, friction, ineligibility, alternate paths, or measurement loss.
Why do funnel totals not match?
Identity fragmentation, different time windows, repeated events, stage-order rules, and tracking gaps can all cause differences; investigate before interpreting them.
Summary
A conversion funnel models ordered progress from a defined opportunity to an outcome. Trust it only when stages have consistent units, event meaning, identity rules, and time windows. In experiments, use the overall outcome among assigned eligible units for the decision and use funnel stages to diagnose mechanisms and data problems.
Sources
- Nielsen Norman Group: Journey mapping
- Google Analytics Help: Funnel exploration
- UK Government Digital Service: Performance data