Your onboarding funnel tells you where, never why

You have the funnel. You know that most new accounts clear step three and something like a third of them never clear step four, and that the gap has been there for three quarters. You have known the number long enough to have put it on a slide, twice.

What you cannot do is finish the sentence. Accounts drop at step four because — and then nothing.

That is not a tooling problem, and no amount of reconfiguring fixes it. The answer was never in the data you are looking at.

Detection and diagnosis are different jobs

A thermometer is an excellent instrument: precise, repeatable, right often enough to act on. It also cannot tell you which infection you have, and nobody holds that against it. It is not a broken diagnostic. It is a very good detector.

Your funnel is doing the same work, well. It found the leak, told you which step, when it got worse, and which segment it is worst in. Every one of those facts is real and none is available any other way. What it cannot price is what the undiagnosed stall is costing you — every account sitting in that bar is one you have already paid full freight to acquire.

The mistake is only in the next question. You point the detector at why and it returns the answer it always had: here is where, in more detail.

What a clickstream can and cannot contain

Work backwards from what an event actually is: a record of an action, taken on a surface you own, by a person who was logged in.

Three boundaries in that sentence, and each one excludes a whole category of cause.

It happens on your surface. The approval that never came, the security review sitting in a queue, the CSV somebody repaired by hand in Excel, the Slack thread where a manager said "park this until Q3" — none of that touches your product, so none of it emits an event. Onboarding runs partly inside the customer's own organisation, and your instrumentation stops at the property line.

It is a logged-in person. The people who decide whether your project moves often do not have accounts — the IT lead who will not whitelist your domain, the CFO who wants one more look at the contract, the two colleagues who were meant to be brought in and never were. Your analytics is scoped to users. The decision is made by the account.

It is an action, not an intention or a belief. The clickstream records that someone opened the integrations page and left. It cannot record that they read "connector" and thought it meant something else, or assumed this was IT's job, or were waiting on the one person who has the API key to come back from leave.

The information that separates one cause from another is almost always in that third category, and that is not a gap in your setup. Beliefs, intentions and other people's politics are not events, so no configuration makes them arrive.

Absence has no properties

Here is the sharper version, and the reason more instrumentation does not help. The event that matters most is the one that did not fire. When a customer completes step four you get a payload: user, timestamp, duration, path, device, referrer. When they do not complete it you get nothing — and nothing has the same shape regardless of what caused it.

Take six accounts that all failed to connect their payroll system in week two.

The first read the setup copy, decided "connector" meant something they did not have, and went looking in the wrong menu. The second needs an integration you do not build, found that out on the page, and moved on. The third is the wrong person in the seat: an ops coordinator handed the task by the VP who bought the product, without the credentials to complete it. The fourth was stopped by an internal security review nobody told you was happening. The fifth exported the data, did the whole thing in a spreadsheet in twenty minutes, and does not feel blocked at all. The sixth did connect it — in a second workspace they created in week one and abandoned.

Six causes, five owners. The first is a copy fix. The second is a roadmap input, or a sales-qualification signal. The third is a handoff failure between sales and onboarding. The fourth is a process gap you close with one question on a kickoff form. The fifth is not a problem at all, and the sixth is a data artefact quietly corrupting your denominator.

In the funnel, all six are one bar.

More resolution does not produce a different kind of answer

The natural response is to look harder: session replay, finer-grained events, a heatmap, a churn model over the behavioural features. Some of that genuinely helps. Replay adds real information — the hesitation, the field filled in and cleared, the three returns to the pricing page. When a drop-off is caused by confusing copy, replay will often show you exactly that, and you should use it.

But notice what improved. You now have a higher-fidelity recording of the same boundary. Under replay, the ops coordinator without credentials and the account blocked by a security review look identical: a person opens a page and does not act. So does the person who is perfectly happy doing it in a spreadsheet. The thing that separates them was never on the screen.

Adding pixels to a photograph does not turn it into a sentence.

Segmentation and predictive models have a related limit: they rank suspects and cannot interrogate one. A churn model relates variables you already record, so it surfaces only causes someone thought to instrument — the same structural problem that makes onboarding surveys unreliable, arriving by a different route. A closed instrument returns only answers you imagined in advance.

So ask them. The obvious way of asking also fails.

Asking is the right move. The standard version of it is another detector. A survey fired at the stalled cohort samples the customers who are still opening your email, measures an attitude rather than an event, and offers a set of reasons you wrote down before you knew the answer. Bolting a form onto a dashboard leaves you with two instruments that both tell you where.

The property that makes something diagnostic is narrow and specific: its next question depends on the last answer. A funnel's next question was fixed when you defined the events. A survey's was fixed when you wrote it. A conversation's is not fixed at all, which is why it can find a cause nobody had hypothesised.

Someone says the integration was fine. You ask what they connected. They did not, in the end — they just exported it. You ask how long that takes. Half an hour, every fortnight, forever. None of that was reachable from question one, and none of it is in the event stream, because from the product's point of view nothing happened.

What a diagnostic loop looks like in practice

Analytics is half of this workflow; the two instruments are complementary, not rivals.

Pick one step and one cohort. Not "onboarding." A single transition in a defined population — accounts that signed in the last two quarters and did not connect a payroll system within thirty days. Diagnosis works on a specific symptom.

Let the funnel build the sample. This is where your existing stack earns its licence: analytics produces an exact list of the accounts that did not complete the step. No conversational method can generate that list, and it beats the self-selected group who answer a survey: it is the population in question, not a filtered slice of it.

Ask within days, not quarters, and ask about the last time rather than about usually: what somebody did at their last month-end is retrievable; how they "usually" do it is a tidy description of a process that does not exist.

Ask about the parts you cannot see. What happened outside the product. Who else was involved. What they were waiting on. What they did instead. Each question aims at one of the three boundaries, and none of the answers are in your warehouse.

Tag the causes and count them. Product gap, process gap, expectation gap, mis-sale, blocked-elsewhere, not-actually-a-problem. Counting converts anecdotes into a ranked list — the same discipline you already apply to behaviour, applied to reasons.

The objection is that this does not scale. It does not need to. Abbie Griffin and John Hauser's 1993 Marketing Science paper "The Voice of the Customer" — still the best empirical answer to the question — found that twenty to thirty customer interviews surface ninety to ninety-five per cent of customer needs, and that twenty captured over ninety per cent of what thirty produced. Twenty-five conversations per segment: a fortnight of someone's part-time attention, once, not a permanent research function.

Then give the answer back to the funnel. This is the step that makes it a loop, and the most valuable of the six. A cause you have named is usually a cause you can measure. "Blocked by an internal security review" is invisible until somebody tells you it happens; the moment you know, it becomes a question on the kickoff form and a field on the account, and next quarter your dashboard counts it. "Wrong person in the seat" becomes a check on whether whoever completed setup is whoever signed. Diagnosis converts an unknown into a metric, and watching metrics is what your existing tools are excellent at.

The funnel finds where. Conversations find why. The why goes back into the funnel, so the next occurrence is detected automatically instead of being discovered a year later by accident.

Keep the dashboard

None of this is an argument for buying fewer analytics tools or trusting them less. Detection is genuinely hard and yours is doing it well. You should have a smoke alarm in every room, and yours is working — it has been telling you which room for three quarters. The mistake is standing in the doorway trying to work out what is burning by listening harder to the beeping.

Your funnel has said everything it knows. The rest of the sentence is held by about twenty-five people, and the only way to get it is to ask them.

---

Where Nosie fits

Your funnel already does the hard half: it produces the exact list of accounts that did not complete the step, which no conversational method can generate. What it cannot do is ask them why.

Nosie is the asking half. Hand it the cohort your funnel flagged and it runs an outbound voice study across it, following each answer rather than reading from a fixed form, then returns the causes transcribed, tagged and counted — a ranked list you can set beside the bar chart. And a cause you have named once becomes a field your dashboard can count from then on, which is the loop closing.

Try it on yourself. Your first self-test interview is free — Nosie rings you, so you hear the interview before your customers do.