◀ Course contents Part 4 · Module 4-02

Funnel Analysis

Find the step where users quietly leave

A funnel is an ordered set of steps toward one outcome, measured by how many people survive each step. Built carefully it points straight at your biggest leak. Built carelessly it turns a tracking bug into a redesign. This guide covers both halves.

Ready?

1

What a Funnel Is, and When It Fits

A funnel is a sequence of steps people pass through in order, where some drop out at each one. Visit, sign up, activate, pay. Four steps, four numbers, and the gaps between them show where you are losing people.

It is the most useful analytical tool in product work and the most frequently misapplied, for one reason: it only works when the behaviour genuinely is a sequence.

An airport is a funnel. A supermarket is not. People wander, double back, pick things up and put them down, and leave empty-handed for reasons that attach to no particular step.

Force supermarket behaviour into five ordered stages and you will still get numbers. The numbers will describe a journey nobody is taking, and the drop you find at step three will often be people doing step four first.

So before drawing one, check the shape of the behaviour. Do people really do these things in this order, and does each step require the one before it? If yes, a funnel will show you exactly where to look. If no, you want a different tool — and noticing early is the difference between a useful quarter and a quarter spent optimising a step nobody is stuck on.

Good fit

Bounded, ordered path to one outcome

Checkout, signup, onboarding, KYC, upgrade. Steps are discrete and instrumented.

Poor fit

Open-ended or compounding journeys

Browsing behavior, long multi-channel journeys, or growth that feeds itself, use a journey map or a growth loop.

It's worth knowing what a funnel is not. A journey map covers emotion and context across touchpoints; a funnel only counts. A growth loop lets output feed back into input; a funnel is strictly one-directional, so it can never model compounding. And a funnel answers where users leave, never why.

A funnel is a sequence with drop-off Four stages narrowing downward with counts: Visited, ten thousand; Signed up, two thousand two hundred; Activated, nine hundred; Paid, one hundred and thirty. Each step shows the conversion rate from the step above. Visited 10,000 Signed up 2,200 · 22% Activated 900 · 41% Paid 130 · 14%
Each row exists only because the row above it happened first, and that is the entire requirement for drawing one. Force supermarket behaviour into this shape and you still get four numbers — describing a journey nobody took, with a drop at step three that is really people doing step four first.

Everyday example

An airport is a funnel: kerb, check-in, security, gate, plane. Everyone passes through in the same order and each stage loses a few people. A supermarket is not a funnel — people wander, double back and leave with nothing, and forcing that into five ordered stages will invent a story that is not there.

The funnel that was not one

A team drew a five-step funnel for a tool people used in no fixed order, then spent a quarter "fixing" step three. The drop at step three was simply people doing step four first. The funnel had created the problem it was measuring.

Quick check

A PM builds a funnel with the steps "visited site", "was interested", "signed up". What's the defect?

2

Defining a Funnel You Can Trust

Most funnel disputes are definition disputes wearing a disguise, and four decisions cause nearly all of them.

What counts as entering a step. Page loaded, or element seen, or action started? What the denominator is. Everyone who entered the funnel, or everyone who reached the previous step? These give very different numbers and both get called “conversion”. The time window. Does someone who signs up today and pays in three weeks count as converted — and if so, when? Who is excluded. Internal staff, bots, returning users, people on unsupported browsers.

None of these has a universally right answer. What matters is that the answer is written down somewhere findable, and that everyone quoting the number is quoting the same one.

A practical habit that saves a great deal of time: put the definition next to the number, every time you report it. “Checkout conversion: 23% (of sessions reaching the cart page, within 7 days, excluding internal traffic).” It looks pedantic on a slide and it prevents the three-week argument.

1

One outcome

Name the single event that counts as success. Everything upstream is a step toward it.

2

Observable steps

Each step is an event the product actually fires. No inferred mental states.

3

Conversion window

How long a user gets to finish. Same session, 24 hours, 30 days, this alone can double the reported rate.

4

Unique users, not events

Count each person once per step, or retries on a failing step will make it look healthy.

5

Strict order or any order

Match the rule to real behavior. Strict ordering silently drops users who took a valid but different path.

Four decisions, and the habit that settles them A checklist of the four definition choices that cause most funnel disputes — what counts as entering a step, what the denominator is, the time window, and who is excluded — and beneath them the habit that settles all four: publishing the definition next to the number. Defined Undefined Define what counts as entering a step State the denominator explicitly Fix the time window List who is excluded Publish the definition with the number "Conversion" with no definition Two teams, two denominators Open-ended attribution window Bots and staff left in silently A number quoted with no footnote
Two teams once reported the same funnel fourteen points apart, and both numbers were correct — one counted signed-in sessions and one did not. Nothing in the right-hand column is a measurement error; every line of it is a number computed properly from an unstated choice, which is why the fifth row settles the other four.

The window decides the answer

A considered purchase is often abandoned, then completed two days later from a reminder email. Under a same-session window those users are counted as losses, and the reminder email looks worthless. Under a 7-day window they're wins and the email is one of the highest-ROI things the team owns. Same users, same behavior, opposite conclusions.

Two definitions, two answers

Two teams reported the same funnel with a fourteen-point difference in conversion. One counted anyone who loaded the page; the other counted anyone who loaded it and was signed in. Neither was wrong, and the argument lasted three weeks before anyone compared the definitions rather than the numbers.

Quick check

A funnel counts every event fired rather than each unique user reaching a step, and many users retry a failing payment three or four times. What happens to the reported numbers?

3

Reading Drop-off, and Ranking the Opportunity

Your funnel has four steps. One loses 30% of the people who reach it; another loses 8%. Which do you fix?

The instinctive answer is the 30%, and it is often wrong, because a percentage on its own hides how many people it applies to. The 30% step might sit at the bottom of the funnel where only a few hundred people arrive. The 8% step might sit near the top with forty thousand.

So rank by absolute loss — the number of people, not the rate. That is the size of the prize. But then apply a second filter that is just as important: how movable is it? A step losing many people may be losing them for a good reason. A pricing page that loses people who were never going to pay is doing its job, and “fixing” it moves unqualified traffic further down the funnel where it costs more to handle.

The two questions together — how many people, and can we plausibly keep them? — are what turn a funnel from a description into a decision. The biggest percentage is a headline; the biggest recoverable number is the work.

Rank by people lost, not by percentage Two funnel steps compared. A step losing thirty per cent of a few hundred users loses fewer people than one losing eight per cent of forty thousand. Step A · 8% of 40,000 3200 lost the smaller percentage Step B · 30% of 900 270 lost the alarming one
The bar lengths are the argument: the alarming step loses 270 people, the boring one loses 3,200. Twelve times the prize, at a quarter of the drop rate — and a percentage on a slide gives you no way at all to see that, because it has thrown away the only number that sizes the work.

Everyday example

A queue losing 50 people out of 100 is a worse queue than one losing 500 out of 5,000, even though the second loses ten times as many people. Absolute loss tells you where the volume is; rate tells you where the problem is. Ranking by the wrong one sends the whole quarter to the top of the funnel.

The biggest percentage isn't the biggest prize

A step where 30% of users drop looks worse than one where 5% drop. But if only 200 users reach the first and 50,000 reach the second, the "small" leak costs 2,500 users a month and the "big" one costs 60. Rank opportunities by drop rate × users reaching the step, then temper by how fixable the cause looks.

Percentage tells you how bad a step is. Volume tells you how much it's worth fixing.

Add a second dimension while you're there: time to complete each step. A step people survive but spend four minutes on is friction that the conversion rate can't see, and it often predicts drop-off further downstream.

The 30% that was worth less than the 8%

One step lost 30% of a few hundred users; another lost 8% of forty thousand. The second was worth roughly ten times more in absolute terms. The team had spent a quarter on the first because the percentage was more alarming.

Quick check

A funnel shows 10,000 users at step 1, 4,000 at step 2, and 3,600 at step 3. Where should the team look first?

4

Traps, and What to Do With a Drop-off

Three traps catch people reading funnels, and all three produce confident conclusions in the wrong direction.

The first is survivorship. The people in your funnel are the ones it has not yet removed. If a step at the very front excludes a whole group — a browser that fails, a country you do not support — that group is invisible in every number after it, and you will conclude they were not interested.

The third is averaging across segments. A flat overall conversion rate can easily be one channel doubling while another halves. The blended number is real and it is describing nobody.

The habit that defends against all three is the same: before acting on any drop-off, segment it and ask what a legitimate reason for this drop would look like. If you can imagine one, go and check whether that is what is happening — because the alternative is spending a quarter making the number look better while the business gets worse.

Instrumentation gaps

A near-total drop appearing overnight with no support tickets is a broken event until proven otherwise.

Blended segments

62% on desktop and 24% on mobile average to a flat 41% that hides both the problem and any progress.

Relocated drop-off

Pushing more users past step 2 means nothing if they simply leave at step 4. Judge on end-to-end.

Borrowed benchmarks

An external "3% is normal" assumes someone else's funnel definition, traffic mix and price point.

Once you've located a real leak, narrow it with a micro-funnel: break the offending step into its sub-steps so "users drop at checkout" becomes "users drop at card validation". That's a specific enough where to hand to qualitative work, session replays, a survey on exit, five user interviews, which is the only thing that will give you the why.

Before acting on a drop-off A chooser for interpreting a drop: check whether a legitimate reason could explain it, whether a segment is hiding inside the average, and whether the funnel itself excluded people upstream. A step just lost a third of its traffic. What now? could it be correct? A price breakdown that filters early is a better business is it one segment? A flat average can be one channel up and one down who never arrived? A barrier upstream makes a group invisible entirely
One checkout step lost a third of its traffic after a release and total completed orders went up — the release had added a complete price breakdown. A worse-looking funnel and a better business.

Everyday example

A shop with a step at the door loses wheelchair users at the entrance, and its floor data will show almost none inside. The drop-off is not evidence that they did not want to shop there.

The drop-off that was the fix working

A checkout step lost a third of its traffic after a release, and the team prepared to roll back. The release had added a complete price breakdown, and the people leaving were the ones who would previously have got to the final screen and abandoned there. Total completed orders had gone up. A worse-looking funnel and a better business.

Quick check

A team improves a landing page, 40% more users reach step 2, and final purchases stay exactly flat. What most likely happened?

Drill what you learned
Scenario 1 easy

A PM builds a funnel with the steps "visited site", "was interested", "signed up".

What's wrong with this funnel?

Scenario 2 easy

A signup funnel shows 10,000 users at step 1, 4,000 at step 2, and 3,600 at step 3.

Where is the largest drop-off?

Scenario 3 easy

A team reports "our funnel converts at 4%" without saying over what period a user has to finish.

What's the missing definition?

Scenario 4 easy

A funnel counts every event fired rather than each unique user reaching a step, so users who retry a failed payment are counted multiple times.

What happens to the reported conversion rate?

Scenario 5 easy

A PM finds a step where 30% of users drop, but only 200 users reach that step each month.

How should this compare against a 5% drop at a step 50,000 users reach?

Scenario 6 medium

An onboarding funnel's step 3 shows a 98% drop-off that appeared overnight, with no release and no change in support tickets.

What's the most likely cause?

Scenario 7 medium

A checkout funnel converts at 62% on desktop and 24% on mobile, and the blended number of 41% has been flat all year.

What should the team do with the blended number?

Scenario 8 medium

A team improves a landing page and pushes 40% more users into step 2, but final purchases stay flat.

What most likely happened?

Scenario 9 medium

A funnel is defined so users must complete steps in strict order, but many real users open pricing before the product tour and still convert.

What does the strict-order definition do to the measured rate?

Scenario 10 medium

A funnel shows exactly where users drop, and the PM is asked to explain why they leave at that step.

What should the PM do next?

Scenario 11 hard

Users who abandon a checkout often return two days later via a reminder email and complete the purchase, but the funnel is windowed to a single session.

What's the effect on the reported result?

Scenario 12 hard

A PM is told the industry benchmark for e-commerce checkout conversion is 3%, and the team's is 2.4%, so the team declares a crisis.

What's the flaw in this reasoning?

Scenario 13 hard

A growth team reports the funnel improved from 5.0% to 5.4% after a redesign, but the traffic mix shifted heavily toward branded search in the same period.

What's the risk in attributing the gain to the redesign?

Scenario 14 hard

A team's checkout funnel shows a modest 8% drop at the address step, but users who pass it take an average of 4 minutes on that screen.

What does the time data add?

Scenario 15 hard

A subscription product wants to understand growth, and a PM proposes modeling the whole business as a single linear funnel from ad click to renewal.

What does this model miss?

Drilled it. Now apply it to a real situation.

Put it to work
From lesson 1

One optional form field, twelve million dollars

Expedia

Expedia's checkout asked for a billing address, and above it an optional field labelled Company. Harmless. Most people left it blank.

Enough people did not. Some read it as "bank name" and typed that, then entered their bank's address underneath. The billing address no longer matched the card, the payment failed, and they left. Removing the field is reported to have been worth around $12 million a year in profit.

What did a funnel view make visible that a page-by-page review would not?

From lesson 2

Two teams, two funnels, same product

A bank's mobile app

Onboarding reports a 61% completion rate. Risk and compliance report 38% for what they also call onboarding. Both pull from the same warehouse. The exec asks which is right.

The two definitions, side by side
Onboarding teamRisk team
Starts atTapped 'Open an account'App installed
Ends atIdentity submittedFirst deposit cleared
WindowSame session30 days
CountsSessionsPeople
ExcludesNothingApplicants under 18

Which of these are true?

Select all that apply — there are 3 to find.

From lesson 3

The steepest percentage is not the biggest prize

An online pharmacy

Monthly funnel, 100,000 visitors at the top. The team wants to know where to spend next quarter.

Last month
StepReachedDrop from previous
Visited100,000
Searched a medicine42,00058%
Added to basket31,00026%
Started checkout9,30070%
Uploaded prescription5,60040%
Ordered5,1009%

Rank these steps by the size of the prize — how many extra orders a realistic improvement would actually produce.

Drag the rows, or use the arrows, then check.

  1. Basket to checkout — 21,700 people lost at a 70% drop First. Both the steepest rate and the largest absolute loss, and it is late enough that the people are already qualified. Ten points here is roughly 3,100 more into checkout.
  2. Visit to search — 58,000 people lost Second on volume and much weaker on quality: a large share of that 58,000 never intended to buy anything. Improving it moves a lot of people a very short distance.
  3. Checkout to prescription upload — 3,700 lost at 40% Third. A painful rate on genuinely committed people, but the pool is small, so even a big proportional win is a modest number of orders.
  4. Prescription to order — 500 lost at 9% Last. The healthiest step in the funnel, and squeezing it is the most work for the least return.
From lesson 4

A drop-off that was not a problem

A B2B procurement tool

A new funnel report shows 82% of users abandoning at "Compare suppliers". It is by far the steepest drop in the product, and it lands on the quarterly plan as the top priority.

Before starting, someone watches ten recordings of people leaving that step.

Work out what is actually happening.

  1. Step 1 of 3

    The recordings show people reading the comparison, then closing the tab. Most return two or three days later and place an order. What does that make the 82%?

    So the step is not leaking. The funnel's window is too short to contain the behaviour it is measuring.

  2. Step 2 of 3

    What would you change about the measurement?

    Re-run per person over 14 days, the step drops 31%. The 82% was the same people, counted as gone because they closed a tab.

  3. Step 3 of 3

    Where does that leave the quarterly plan?

    The cost of the trap would have been most of a quarter of engineering aimed at the wrong step.

Notification