A funnel is a named sequence of steps through a product with a count of how many people reached each one, and a conversion rate is the share of one step's people who went on to reach the next.
Collection that can be trusted is what makes those counts worth reading. A team that has monitored its own events and knows which ones fire reliably can draw a funnel and believe what it shows. Without that monitoring, the chart reports the state of the instrumentation as much as the behaviour of customers.
Given trustworthy events, the shape a team draws from them is one borrowed from selling. Elias St Elmo Lewis set out attention, interest, desire and action as the stages of a sale in 1898, and William Townsend drew those stages as a funnel in Bond Salesmanship in 1924. Product analytics kept the shape and changed what sits inside it. A sales funnel counts prospects a salesperson is working through, and a product funnel counts people the product itself is carrying from one screen to the next.
The reason this reaches a product manager every week is that most requests for work arrive as a claim about a step. Sales reports that the trial is too hard to start, support reports that customers cannot connect their data, and a designer wants to rebuild the second screen. A funnel is the artefact that says how many people each of those claims is actually about.
The sections below cover the three choices that decide what a funnel measures. One funnel is then worked through with figures a reader can check. Finally they set out what the denominator does to the number and what a single drop leaves open.
The three choices that decide what a funnel measures
The first step loses the most people in absolute terms, because most of the visitors counted there never intended to buy anything. Every step multiplies the one above it, so a one tenth improvement at any single step raises the number of paying customers by one tenth.
The steps. A funnel is only as meaningful as the events behind it, and the events are chosen by a person. A team that defines the second step as opening the invite dialogue will read a different story from a team that defines it as an invitation actually being sent. Both are defensible. Neither is comparable with the other, so the definition belongs next to the chart.
The conversion window. A window is how long somebody has to complete the sequence before the funnel stops waiting for them. Seven days and thirty days produce different numbers from the same events, because a share of people finish late. If the funnel above uses a window of seven days, 360 of the 480 customers have paid inside it, and the end to end rate reads 0.9 per cent. The same funnel over thirty days reads 1.2 per cent. Nothing about the product changed between those two figures.
The order. An enforced order counts somebody only if they took the steps in the sequence given. A funnel with no enforced order counts anybody who did all of them, in any sequence, inside the window. Checkout is the case for an enforced order, since paying before choosing a delivery address is not a real path. Onboarding is usually the case against one, because a person who connects their data before inviting a colleague has done both things and is no less onboarded for the sequence.
One funnel worked through with figures
The worked example is a product sold to small finance teams, measured over one month with a conversion window of thirty days. Every figure is invented, and each step conversion is the count on that row divided by the count on the row above it.
| Step | People | Step conversion | Share of the top |
|---|---|---|---|
| Visited the pricing page | 40,000 | 100 per cent | |
| Started a trial | 8,000 | 20 per cent | 20 per cent |
| Invited a colleague | 2,400 | 30 per cent | 6 per cent |
| Connected a data source | 1,200 | 50 per cent | 3 per cent |
| Paid | 480 | 40 per cent | 1.2 per cent |
The four step conversions multiply out to the end to end figure. Twenty per cent of forty thousand is eight thousand, thirty per cent of that is two thousand four hundred, half of that is twelve hundred and forty per cent of that is four hundred and eighty. Written as a product, 0.20 times 0.30 times 0.50 times 0.40 is 0.012, which is the 1.2 per cent in the last row.
That multiplication has a consequence worth sitting with. Lifting the second step from 30 per cent to 33 per cent, which is a one tenth improvement, gives 40,000 times 0.20 times 0.33 times 0.50 times 0.40, or 528 customers. Lifting the first step from 20 per cent to 22 per cent instead gives 40,000 times 0.22 times 0.30 times 0.50 times 0.40, which is also 528. A one tenth gain is worth forty eight customers wherever in the sequence it happens, so the choice between the two is a question about which one the team can actually move and at what cost.
The denominator decides the number
Four hundred and eighty customers can be quoted honestly as 1.2 per cent conversion or as 6 per cent conversion. The first divides by the forty thousand people who saw the pricing page. The second divides by the eight thousand who started a trial. Five times separates them, and the only difference is which population the team decided to hold itself responsible for.
Marketing usually prefers the larger denominator because it includes the traffic they bought, and the product team usually prefers the smaller one because it starts where the product starts. Both preferences are reasonable and neither is a fact about the product. The practice that ends the argument is writing the denominator into the name of the measure, so the chart is labelled trial to paid conversion and nobody has to ask.
The same discipline catches a quieter failure. A team that reports conversion rising while the count of customers falls has usually shrunk its denominator by turning off a traffic source, so the rate improved because the weakest visitors stopped arriving. Reporting the count beside the rate makes that visible in one glance.
What a single drop leaves open
A stated denominator makes a rate comparable and leaves the more interesting question untouched. A drop between two steps records how many people stopped and gives no reason for it. Eight thousand people started a trial and five thousand six hundred of them never invited anybody, which is a fact. Whether those people found the invite flow confusing, had nobody to invite, were evaluating the product alone on purpose or were never going to buy is not in the funnel at all.
Nothing in a single funnel separates those explanations, because the funnel counts arrivals at a step and records no reason for a departure. Session replays, support conversations and a short survey to people who stopped are what fill that gap, and the funnel's job is to say which stop is worth the research time.
A funnel measured once also cannot say whether 30 per cent is good. The figure has no comparison attached to it, so a team looking at it alone will read improvement into normal variation and read a decline into a week with a holiday in it. The comparison that answers this groups people by when they arrived and follows each group forward, which turns one number into a series that can be read.