The short version
A funnel chart looks like a picture of your customers. It is really a picture of your tracking. Every bar is built from events your product sends when someone does something, and if those events are missing, doubled up, or fired at the wrong moment, the funnel will tell you a confident story that is not true.
This is a practical guide to how funnel events work, how to design them, and how to check them before you make a decision on them. It comes straight out of real work. In a recent onboarding test, the way the events were built changed the size of the headline result.
The old version was losing more of its real buyers to tracking gaps, so the detailed funnel made the new version look better than it was. Both numbers also added up weeks while the traffic split moved, and both used the browser's copy. Read week by week on server records, there was no clear difference at all.
How a funnel works
Under the hood, a funnel is a query, not a picture. You hand your analytics tool (Mixpanel, Amplitude, PostHog, and the rest all work this way) an ordered list of events. For each person, it finds the first event, then looks for the next one after it, then the next, inside a time limit you set. Everyone who fires step one is the top of the funnel. Each bar after that is the share of people who made it to the next step, in order.
That means a person only counts at a step if the event for that step exists, belongs to the right person, and happened in the right order. Real behaviour that never became an event does not exist as far as the funnel is concerned.
The settings that change the answer
- Order. Most funnels are strict about sequence. A step only counts if it happens after the one before it, so anyone missing a step in the middle falls out at that point.
- Conversion window. How long a person has, from their first step, to finish. Seven days and thirty days give different answers, and people who signed up yesterday have not had their full window yet.
- Counting unit. Unique people or total times. Funnels almost always want unique people, which hides duplicate events in the funnel but not in your other reports.
- Identity. Whether what someone did before signing up is joined to the account they create. If it is not, your pre-signup steps and your post-signup steps look like two different crowds.
- Breakdowns. Splitting the funnel by a property such as experiment variant, platform, or channel. This only works if every event carries that property.
Four people go through the same four-step funnel. Here is what each one sent, and what the funnel reports.
- 1Signed up
- 2Profile completed
- 3Plan selected
- 4Purchased
| Person | 1 | 2 | 3 | 4 | What the funnel reports |
|---|---|---|---|---|---|
| A | CustomerEvery event fired, in order. | ||||
| B | Lost at step 3They paid, but the plan event never fired. | ||||
| C | CustomerCounted once, though step 2 sent two events. | ||||
| D | Lost at step 3Left after their profile. A real drop-off. |
Look at B and D. The funnel reports them identically: both lost at step 3. Only D actually left. B is a paying customer the chart has thrown away, and nothing on the chart tells you which is which.
What makes a good event
Every event is a small record of one thing one person did. Simplified, it looks like this:
{
"event": "Plan Selected",
"user_id": "u_8f3a21",
"timestamp": "2026-10-14T18:22:05Z",
"properties": {
"plan": "monthly",
"onboarding_version": "v6",
"platform": "web"
}
}
Four parts, and each one has a job:
- A name that means one thing forever. Use the object and a past-tense action:
Plan Selected,Checkout Completed. Pick one convention, write it down, and never rename a live event. A rename quietly splits your history into two events. - A trigger at the moment of success. Fire when the thing actually happened, not when someone tried.
Questionnaire Completedwhen the answers are saved beatsSubmit Clickedwhen someone taps a button that might fail. - An identity that follows the person. When someone signs up, tell your analytics tool who they are (most tools call this
identify) so their anonymous browsing and their new account join up as one person. - Properties you will want to slice by later. The version or experiment variant belongs on every single event. Most tools can attach it automatically once it is set; Mixpanel calls these super properties.
Designing the events for a funnel
- Start with the decision. Write down the question the funnel has to answer, such as "does the new onboarding turn more signups into customers?" That fixes the first and last step, and those two matter most.
- Mark the step boundaries, not every click. One event at each point where a person has clearly moved forward. Five to eight steps is plenty for most flows.
- Send each event from the most reliable place. Money and account events (signups, purchases, refunds) should come from your server or your payment provider's webhook, where ad blockers and closed tabs cannot stop them. Interface events can come from the browser.
- Give every version shared milestones. A redesign changes the screens, so old and new versions will not have the same events. Keep a small set of milestones that mean exactly the same thing in both, so you can compare them on equal terms.
- Tag everything with the variant. If an event does not carry the version, you cannot break that step down by version later.
- Write a tracking plan. One table that lists each event, exactly when it fires, its properties, and where it is sent from. It becomes the contract between product, engineering, and whoever reads the numbers.
| Event | Fires when | Key properties | Sent from |
|---|---|---|---|
| Signed Up | The account is created | signup method, onboarding version | Server |
| Profile Completed | The profile saves successfully | onboarding version | Server |
| Plan Selected | The plan choice is saved | plan, onboarding version | Browser |
| Purchase Completed | The payment succeeds | plan, currency, onboarding version | Payment webhook |
Why it matters: six ways a funnel can mislead you
- Missing events make real customers disappear. Person B above. Worse, the gaps are rarely even. In the onboarding test, about 7% of buyers on the old version were missing a step event against about 1% on the new one, which tilted the comparison toward the new version.
- Double fires inflate steps. A screen that re-renders, a retry, or a back button can send the same event twice. Unique-user funnels hide it. Event totals, averages, and anything per-session do not.
- Clicks are not outcomes. Tracking the button instead of the result counts attempts, including every one that failed validation or timed out.
- Broken identity splits one person in two. If the anonymous visitor is never joined to the account they create, landing-to-signup conversion looks terrible and everything after signup looks perfect.
- Browser-only tracking undercounts unevenly. Ad blockers and privacy settings drop a share of browser events, and that share varies by device and audience. If one variant skews toward those people, so does your result.
- Redefined events break history. If
Checkout Startedmeant one thing until March and something else after, every trend line across that date is comparing two different measures.
How to check a funnel before you trust it
- Run the simple version too. Compare the full funnel with a two-step funnel of just the first and last event. If they disagree, and especially if they disagree differently for each variant, an event in the middle is going missing. This is the check that caught the onboarding gap.
- Reconcile with the source of truth. Purchases in your analytics tool should match purchases in your payment system within a few percent.
- Walk the funnel yourself. Open the live event view, go through as a brand-new user on web and on mobile, and check that each event fires once, in order, with the right properties.
- Check the traffic split. In an experiment, confirm each variant got the share of people it was supposed to. A lopsided split usually means assignment or tracking is off somewhere.
- Let the data mature. Only count people who have had their full conversion window, or the most recent signups will drag every version down.
The takeaway
Funnel analysis is only as good as the events underneath it. A few minutes spent naming an event properly, firing it at the right moment, and tagging it with the version saves weeks of arguing about whether a result is real. When the numbers matter, check the plumbing first.