Funnel Analysis
Date: 2026-08-16
Where people leave an ordered sequence. The mechanics are trivial; the choices — window, ordering, re-entry, unit — move the answer more than anything users do, and they’re usually made by a tool’s defaults.
What it is
Funnel analysis measures how many units reach each step of an ordered sequence and where they drop out.
step users step CR cumulative
product_viewed 120,000 — 100%
add_to_cart 24,000 20.0% 20%
checkout_started 14,400 60.0% 12%
add_payment_info 10,080 70.0% 8.4%
purchase 7,560 75.0% 6.3%
Step conversion is the diagnostic; cumulative is the headline. The 60% add-to-cart to checkout step is where this funnel loses most, and it’s invisible if you only read the 6.3%.
The four choices that move the answer
1. The window. How long does someone have to complete the sequence?
- Same session — tight, and undercounts considered purchases where people return
- Fixed period (7, 30 days) — needs identity to survive that long, which for anonymous traffic it usually doesn’t — Browser Privacy Restrictions
- Unbounded — overcounts, and never settles because yesterday’s cohort can still convert
2. Strict or loose ordering. Must the steps occur in sequence, or just all occur? A user who views a product, adds to cart, browses more, then checks out counts under loose ordering and may not under strict.
3. Re-entry. Someone who abandons and returns twice — one journey or three? Most tools count first entry per user per window, which is usually right and is rarely stated.
4. The unit. Users, sessions, or the entity itself. Session-scoped funnels are the trap: the denominator moves with the timeout and with event volume, so a more interactive variant can improve the funnel while selling nothing — Sessionisation, Randomisation Unit.
In plain terms: two analysts with the same data and different defaults will produce different drop-off numbers, and both will be right. Which is why the choices belong in the metric definition rather than in the tool.
Reading one properly
- Step conversion, not cumulative. The biggest absolute loss is usually the first step, because it has the most people. The biggest problem is the step with the worst rate relative to what’s reasonable
- Segment before concluding. Mobile and desktop funnels differ enough that a blended view describes neither, and a shift in traffic mix looks like a performance change — Simpson’s Paradox
- Compare to itself over time, not to a benchmark. Funnel definitions differ so much between businesses that industry comparisons are close to meaningless — Benchmarking
- Check the step exists before believing a 95% drop. A step that fires unreliably produces a cliff that’s a tracking bug — Guide - Auditing a Tracking Plan
What it can’t tell you
- Why. A funnel locates the loss and says nothing about the cause. That’s Session Replay, Form Analytics and Usability Testing — quant to find, qual to explain
- Whether the loss is bad. A product page to add-to-cart rate of 20% isn’t a problem if 80% of viewers were browsing. Not everyone entering a funnel intended to complete it
- What would happen if you fixed it. The drop-off is not the size of the opportunity. People who left may have had no intent
That last point is the most common misreading. A 40% drop at delivery options does not mean fixing delivery options recovers 40%.
Where it’s strongest
Comparing the same funnel across segments or over time. The definition choices stay constant, so differences are real:
- Mobile against desktop, same funnel
- New against returning
- Before and after a change, with the date annotated — Annotation and Change Logs
- Between test arms, on the unit you randomised
Funnels are weak as an absolute measure and strong as a comparative one. Read them that way and the definitional problems mostly cancel.