Quick definition: A funnel is a defined sequence of customer stages used to measure progress, conversion, and loss from an initial opportunity to a final outcome.
What is a funnel?
A funnel models how users, sessions, or accounts move through an ordered journey. A commerce funnel might be product view, cart, checkout, and paid order. A B2B onboarding funnel might be account creation, integration, first successful task, and subscription. It is a measurement model, not proof that every customer follows one linear path. Valid alternate routes must be explicitly represented or they will look like loss.
Every stage needs an event, unit, eligibility rule, order rule, and time window. “Unique users who successfully start checkout after adding a cart item” is a defined stage. “Interested customers” is not. Decide whether people may enter midway, repeat a step, complete on another device, or skip a conventional route. These choices determine the population and the interpretation.
A funnel differs from a single conversion rate, which compares a final outcome to a broad population. It differs from a customer journey map, which can describe motivations and touchpoints without requiring instrumented stages. Funnel analysis provides counts and conditional rates; it should be supplemented by qualitative research and operational evidence when explaining why customers do not progress.
Funnel formulas and denominators
For sequential stages A and B:
stage conversion A→B = unique eligible units reaching B after A / unique units reaching A × 100
Overall conversion is unique units completing the final outcome / unique units entering the defined funnel × 100. Use one consistent unit. Events, sessions, users, and accounts cannot be casually mixed. A step rate can be diagnostically useful, but calculate the final rate directly rather than multiplying rounded percentages.
| Decision question | Useful denominator | Key risk |
|---|---|---|
| Where does checkout fail? | Unique checkout starters | It omits earlier losses. |
| Does a landing experience create orders? | Eligible assigned visitors | Needs a mature attribution window. |
| Does onboarding create value? | New eligible accounts | Requires account identity rules. |
Funnels in A/B testing
Funnels explain mechanisms in experiments: a variant may increase checkout starts, reduce payment failure, or change the mix of people reaching a late stage. Select one primary decision outcome, usually final conversion among all assigned eligible units, and use stages as secondary diagnostics. If a product-page change alters who begins checkout, comparing only checkout starters is post-treatment selection and does not estimate the variant’s full effect.
Predefine stage definitions, order, window, primary outcome, and guardrails. A flow can raise paid orders but create refunds, fraud, latency, complaints, or low-quality leads. Primary and guardrail metrics provides a useful framework. Preplanned sample size and duration should include time for the final outcome to mature; see A/B test duration.
Worked funnel example
A retailer assigns 20,000 eligible sessions to each checkout design. Control records 8,000 cart additions, 4,000 checkout starts, and 2,400 paid orders. Treatment records 8,200 cart additions, 4,510 starts, and 2,620 orders.
control overall conversion = 2,400 / 20,000 = 12.0%treatment overall conversion = 2,620 / 20,000 = 13.1%overall difference = +1.1 percentage pointscontrol checkout completion = 2,400 / 4,000 = 60.0%
The stage pattern suggests treatment helped customers enter checkout, but it is not evidence by itself that the payment step changed. The team estimates uncertainty for the preselected paid-order outcome and checks order value, cancellations, stock, payment errors, and shipping costs before making a decision.
Data-quality caveats
Funnels commonly break at client-server handoffs and identity changes. A browser can log checkout start while a billing system confirms payment later; a customer may finish on another device. Use stable identifiers where possible, reconcile final outcomes to a source of truth, deduplicate retries, and document matching logic. A payment webhook and confirmation page must not double-count the same order.
Event semantics must survive redesigns. If a variant no longer emits the old “checkout started” event, a stage change can be instrumentation rather than behavior. Audit stage counts, missing values, ordering, and timestamp delays by arm. Bots, internal traffic, outages, campaigns, and device mix can affect the journey. Randomization does not repair variant-specific tracking; inspect unexpected allocation with SRM diagnostics.
Practical interpretation
Use the funnel to locate high-volume, meaningful constraints, then investigate causes with error logs, usability work, path analysis, and support feedback. Prioritize a stage only when fixing it can improve the final outcome. A high conditional loss may be harmless if it represents an alternate path; a modest early loss can dominate business impact because many more people encounter it.
Keep separate views for overall and conditional performance. Overall conversion answers whether the journey creates the intended result from the available opportunity. Conditional stage rates explain where a change may have operated, but their denominators are influenced by earlier behavior. This distinction prevents a team from celebrating a late-stage improvement that is offset by loss before the stage.
Funnel trends should use comparable cohorts. A newly launched campaign, stock outage, product-price change, device release, or season can change the mix of people entering the journey. Annotate these events and compare like with like. Where a customer can take several valid routes, create a path-aware measurement design rather than forcing all activity through one canonical sequence.
Use an experiment funnel as a quality-control tool as well as a diagnostic tool. Abrupt variant-specific breaks, impossible stage totals, or changes concentrated immediately after a tracking release should trigger instrumentation investigation before behavioral interpretation. Preserve event definitions and source timestamps so a later reader can reproduce the analysis.
This preserves trustworthy decisions over time.
Review both successful and unsuccessful paths before changing a stage definition.
Common funnel mistakes
- Assuming a linear journey: valid paths become artificial leakage.
- Mixing units: stage rates cannot be compared.
- Using later-stage subsets as causal evidence: treatment can change entry.
- Ignoring cohort maturity: late outcomes disappear.
- Changing event definitions: tracking creates fake movement.
Frequently asked questions
How many stages should a funnel have?
Use the fewest stages that represent meaningful decisions or operational handoffs.
Can funnel rates be multiplied?
Only with compatible units and rules; calculate the final rate directly whenever possible.
Is the narrowest drop-off always the priority?
No. Consider traffic volume, final impact, alternate paths, and fixability.
Why do funnel totals not match?
Identity fragmentation, repeat events, time windows, ordering rules, and tracking loss are common causes.
Summary
A funnel measures defined progress through an ordered journey. Trust it only when stage events, units, identity rules, and time windows are explicit. In experiments, use final conversion among assigned eligible units for the decision and stages for diagnosis.
Sources
- Google Analytics Help: Funnel exploration
- Nielsen Norman Group: Journey mapping
- UK Government Digital Service: Performance data