How a funnel works

A funnel counts how many people (or stores, or orders) reach each step on the way to a goal. Two rates come out of it. The step rate is the share of the previous step that reaches this one. The end-to-end rate is the share of everyone who started that reaches the end. The end-to-end rate is all the step rates multiplied together.

Here is the Store app's adoption funnel across Mangrove's 320 stores in the first month of the receiving rollout. "Fully adopted" means a store is still receiving deliveries in the app four weeks after it started.

Two questions decide where to work. Where is the biggest drop? In stores, step 3: 97 stores signed in and never received a delivery in the app. Where would the same effort add the most? Raise any step rate by 10 points and the end-to-end result is multiplied by (new rate ÷ old rate). That ratio is largest where the rate is lowest. Ten points at step 3 (60% to 70%) multiplies the end by 1.17 and adds about 13 fully adopted stores. Ten points at step 2 (80% to 90%) multiplies it by 1.125 and adds about 10.

Fully adopted now: 320 x 0.95 x 0.80 x 0.60 x 0.70 x 0.75 = 76.6 stores

+10 points at step 3 (60 -> 70%):  76.6 x 70/60 = 89.4   (+12.8)
+10 points at step 4 (70 -> 80%):  76.6 x 80/70 = 87.5   (+10.9)
+10 points at step 5 (75 -> 85%):  76.6 x 85/75 = 86.8   (+10.2)
+10 points at step 2 (80 -> 90%):  76.6 x 90/80 = 86.2   (+9.6)
 +5 points at step 1 (95 -> 100%): 76.6 x 100/95 = 80.6  (+4.0)

Points are not equally easy to win at every step, so treat this as the first cut. And a funnel tells you where people stop, never why. For the why, go and look: visit five stores that signed in and never received in the app, and watch a delivery arrive.

Define the funnel before you read it

Two people can build "the adoption funnel" from the same data and get different answers, because a funnel has five settings. Write them down before anyone reads a number.

• Unit. Count stores, staff or deliveries, whichever matches the decision. Rollout waves and training visits are decided per store; a training video is judged per person. The simulator below lets you switch between stores and staff.
• Steps. Each step is one specific event with a precise definition, such as "first supplier delivery received in the app by a person at the store". Events and their definitions are the subject of Instrumentation.
• Order. Must the steps happen in order? Mixpanel's funnels, for example, require the steps in order by default and offer an "any order" option.
• Conversion window. How long a user has to get from the first step to the last. Mixpanel's default is 7 days. For a store rollout, four weeks from go-live is more realistic.
• Cohort. Group people by when they started. A store that went live yesterday can't have reached "still using at week 4" yet.

The cohort setting causes the most common funnel mistake. If you mix stores that went live four weeks ago with stores that went live this week, the later steps look terrible, because most stores haven't had time to get there. Count only units old enough to have completed the window, or show one funnel per start week, the same way you read cohorts in Retention and cohorts.

And when a definition changes, say so on the chart. Changing the window from 7 to 30 days can lift a funnel's end-to-end rate overnight with no change in anyone's behaviour.

Activation and time to value

Activation is the moment a new user first gets the core value. A good activation event is early, specific and predictive: people who reach it are much more likely to stay than people who don't. Amplitude's write-up of north star metrics notes, for example, that in its early days Facebook tracked how many new users added seven friends in their first ten days: an early, specific action tied to the value of the product.

To find yours:

1. List early actions that might matter: first order, first on-time delivery, first repeat order, adding an address, saving a list.
2. Compare people who were still active at week 4 with people who weren't. Which early action separates them most?
3. Prefer an action the product can help people reach. "Placed a second order within 14 days" is useful; "lives near a store" is not something onboarding can change.
4. Treat it as a hypothesis. Heavy users do everything, so many actions correlate with staying. The test is causal: help more people reach the event (better onboarding, a nudge) and check whether retention follows (Experiments).

Time to value is how long activation takes, usually as a median. For the Mangrove app it is the time from install to the first on-time delivery, the "30 minutes, really?" moment. Every hour of that time is a chance to drop out: an empty first screen, a slot that's full, an item cut from the order. Onboarding that shortens it is the subject of pm-32.

For the Store app, value arrives with the first delivery received in the app instead of on paper. If a store goes live on Monday and its first app receipt is on Thursday, staff have spent three days on paper, and the old habit has had three more days to win. The funnel's step 3 is the activation step, and the time from go-live to that first receipt is the time to value. That is why the rollout plan in Launching puts a trainer in the store for the first delivery.

Adoption of internal tools

Store staff can't uninstall the Store app, but they can keep the paper on the side. So "staff who logged in" says little; head office can make it 100% with one email. Measure adoption as the share of the work done in the tool, in four dimensions:

• Breadth: stores where receiving in the app is switched on and staff are signed in.
• Depth: the share of each store's deliveries received in the app.
• Frequency: every delivery, every shift, including the 6 am delivery when the trained person isn't on.
• Persistence: still at that depth at week 4 and week 12, after the trainer has left.

Then measure the workaround that should die, as in pm-11: paper goods received note pads ordered from the print room, photos of delivery notes in chat groups. If the app says 90% and the print room still sends 400 pads a month, believe the print room.

Two definitions decide whether the adoption number is honest. First, who did the work. Count a delivery as received in the app only when a person scanned it at the store. Some systems close an unreceived stock transfer automatically after a day or two; an auto-closed receipt is the opposite of adoption, and it must not be counted as one. Second, which deliveries are in the denominator, which is where stock transfers come in.

Each Mangrove store gets about 4 supplier deliveries a day and about 2 stock transfers from its DC. A transfer arrives with an electronic transfer note, so receiving it in the app is one barcode scan, and nearly every one (about 98%) is received in the app. A supplier delivery is the hard case: the driver's paperwork, line-by-line checks, quantities that don't match the purchase order. If one headline number mixes the two, it rises and falls with the mix of deliveries as well as with what staff do. Statisticians call the extreme version of this Simpson's paradox.

The fix is in the reporting, not the arithmetic. Report adoption by delivery type, lead with supplier deliveries, keep transfers as a separate line, and show the counts next to the rates so a change in mix is visible. If you need one headline for a slide, use the supplier rate and say so.