Back to Explore
GuideVibeFix editorialUpdated Nov 21, 2026

Don't Just Watch Total Visits: Funnel Analysis for Your Vibe Project

Of 640 visits, only 214 came back a second time within 7 days: the funnel tells you which step lost the other 426. Fix the steps before instrumenting, count heads lost before rates, cut by one dimension at a time, and fix one step per week. The walk-through project, people, and numbers in this guide are fictional.

A real photo: a laptop showing a web analytics dashboard with five KPI cards on top — users, new users, sessions, purchases, revenue — and traffic charts below.

Visits are up, but users are leaking out halfway

Ah Zhe built a pomodoro timer app called "Fan Qie Da Ka" — "tomato check-in". One month after launch, visits looked healthy: 640 visits a day. But the more he looked, the more wrong it felt: only 512 people signed up, 398 finished onboarding, 296 actually completed one focus session, and only 214 came back a second time within 7 days. 640 people came in, 214 stayed, and 426 were lost in between — two thirds gone. The visits number kept climbing, but Ah Zhe knew: what was climbing was the number of people walking in; what was leaking was the number of people staying. He did not know which step those 426 people left at, or which step to fix first.

A real photo: a laptop showing a web analytics dashboard with five KPI cards on top — users, new users, sessions, purchases, revenue — and traffic charts below.

This guide is about how to find where those 426 people went: funnel analysis. To be clear up front: the walk-through project, people, and numbers in this guide are fictional; the method itself is general, and you can follow it as written. Funnel analysis is nothing advanced. It means lining up the steps a user walks from entering to staying, counting how many people are left at each step, and seeing where they drop off. If you can count, you can build a funnel.

Fix the five steps first, then instrument

The first step of a funnel is not instrumentation; it is fixing the steps. Get the steps wrong and every number after that is wrong. Take "Fan Qie Da Ka" as the example. The steps are five: visit, sign up, finish onboarding, complete a first focus session, come back a second time within 7 days. They follow the road users actually walk, each step continuing from the one before: someone who never visited cannot sign up, someone who never signed up cannot finish onboarding, someone who never finished onboarding will rarely complete a first focus session, let alone come back within 7 days.

There are three rules for fixing steps. First, every step needs a clear definition of "done" — writable in one sentence and countable by a machine. "Signed up" means the account is created and can log in, not "clicked the sign-up button" — someone who clicked but never verified their email does not count as signed up. Second, steps must neither skip nor overlap: the same person cannot count in two steps at once, and nobody jumps from step one to step four. Third, do not use too many steps; around five is right. Beyond seven, the funnel shatters, each step holds too few people, and you cannot tell which step is the real problem.

Only once the steps are fixed do you instrument. The rule is also one sentence: one event per step, with a clear name. visit, signup_done, onboarding_done, first_focus_done, return_day7. Do not use names like "click1" or "event2" that nobody can read — three months later, you will not be able to read them either. After instrumenting, do not rush to look at the numbers. Walk the whole flow yourself first: visit, sign up, onboard, first focus session, come back the next day, and check that all five events fire, in the right order, with the right counts. An event that does not fire when you walk through it is the same as not instrumented.

Two ways to read the numbers: count heads lost first, then check the rate

The "Fan Qie Da Ka" funnel reads like this: 640 visits, 512 sign-ups, 398 finished onboarding, 296 first focus sessions completed, 214 second visits within 7 days. Step-to-step conversion is 80%, 78%, 74%, and 72%; overall conversion is 33.4%. The chart below is this funnel, with the headcount and conversion at every step.

Funnel chart: where 640 visits go — 640 visits, 512 sign-ups, 398 finished onboarding, 296 first focus sessions completed, 214 came back a second time within 7 days, with step-to-step conversion of 80%, 78%, 74%, and 72%.

There are two ways to read a funnel, and you need both. The first counts heads lost in absolute terms: 128 people lost from visit to sign-up, 114 from sign-up to onboarding, 102 from onboarding to first focus, and 82 from first focus to the second visit. The biggest loss is the first step — 128 people never even finished signing up. The second looks at conversion rates: the last step has the lowest rate, only 72%, which means 28% of the people who completed a first focus session never came back. The two readings point at different fixes: looking only at rates, you would fix the last step first; looking only at headcount, you would fix the first step. The right answer is to use both: the first step loses the most people, so fix it first; the last step has the lowest rate, so note it down for the second round.

The first step's 128 lost people are worth three questions: is the sign-up flow too long? Are there too many fields to fill in? Is the email verification step blocking a lot of people? When Ah Zhe checked, he found sign-up asked for 6 fields plus email verification, and many people walked away halfway through the form. He cut sign-up to email plus password only, and moved verification to after registration — and the first step's conversion went up. The lesson is that a funnel never tells you the answer; it only tells you where to look for it. Numbers point the way; you still have to ask and test for the cause.

Cut it open: look by channel and by device

Once you have read the total funnel, the next move is to cut it open. The same funnel can behave completely differently for people from different channels or devices. People arriving from a search engine might sign up at 85%; people clicking in from social media might sign up at only 60% — the latter just tapped in passing, so lower intent to sign up is normal. Mobile and desktop users can differ too: a sign-up flow that is too long hurts more on a phone, while a desktop screen shows onboarding more clearly and finishes at a higher rate.

Take the sign-up step of "Fan Qie Da Ka" as a worked example: of the 640 visits, 320 came from search engines and 272 of them signed up — 85%; the other 320 came from social media and 240 signed up — 75%. The two sides add up to exactly 512, matching the total funnel. The 10-point gap says one thing: search visitors were already looking for a pomodoro timer and wanted to sign up; social visitors tapped in passing and wanted to look around first. The fix differs too: 85% from search is already fine, leave it alone; on the social side, what needs fixing is not the sign-up flow but the landing page — first make it obvious at a glance what the app does, then talk about signing up.

The rule for cutting is to cut by one dimension at a time. Cut by channel first, read it, then cut by device — do not cut by both at once, because channel times device leaves too few people in each cell, the numbers start to wobble, and no pattern survives. After cutting, look for the "odd cell": whichever channel or device has one step clearly worse than the average is the cell to fix first. If every channel signs up at around 80% except one at 55%, go look at that channel's landing page: does the headline disagree with the ad, is it loading too slowly, is the button unreachable on a phone?

Cutting also separates product problems from traffic problems. If one step is bad across every channel, that step of the product has a problem; if only one channel is bad, the traffic from that channel is probably the wrong audience. Confusing the two is expensive: if the product is fine and you change the product, the change is wasted; if the traffic is wrong and you change the product, the change is wasted too.

Three more numbers to watch alongside the funnel

A funnel tells you where people drop off, but it cannot answer three questions: will the people who left come back, how long does it take to first use the core feature, and how many new users genuinely "activate". Those need three more numbers, and once the funnel is built, you can wire them up alongside it.

The first is retention: how many people are still coming back after 7 days. The funnel's last step — "came back a second time within 7 days" — is the beginning of retention, but counting one return is not enough: of those who came back, how many return after 14 days? After 30 days? Retention has to be read by day: day 1, day 7, and day 30 after sign-up. Looking only at 7-day retention of 33.4% (214 divided by 640) hides a piece of information: of the people who came back, how many truly stayed, and how many came back once and left again. If the retention curve flattens after day 7, the people who stayed are the core users; if it keeps sliding down, the product has not yet found a reason for people to return.

The second is "how long until the core feature is first used". For a pomodoro app, the core feature is completing one focus session. From sign-up to the first completed focus session, with onboarding in between, how long does it take on average? If the average is two days, onboarding is too long — or users signed up and forgot to come back. If the average is 10 minutes, onboarding is smooth. The funnel counts "how many people reached this step" but not "how long it took". Adding the time dimension makes the problem sharper: people are not refusing to come; they are coming too slowly, slowly enough to forget.

The third is the activation rate, which you define yourself: for "Fan Qie Da Ka", activation can mean "completed one focus session within 7 days of signing up". That relates to the funnel's last step, but it is not the same number: the funnel counts 214 out of 640 visits, while the activation rate counts, out of 512 registered users, how many completed a focus session within 7 days. Different denominators, different questions. The funnel answers "how many people were lost at each step"; the activation rate answers "of the people who registered, how many actually started using it". Read together, they separate "not enough people coming in" from "people coming in but not using it".

Fix one step at a time, then look at the numbers again

Once you find the leakiest step, fix it, wait, and look at the numbers again. The order matters. Fix one step at a time, because if you change three things at once and the numbers improve, you will not know which change did it; if they get worse, you will not know which to roll back. When Ah Zhe fixed sign-up, he changed only sign-up: fields cut from 6 to 2, verification moved to after registration. After a week, he looked at the funnel again: sign-up conversion had risen from 80% to 88%, the later steps barely moved, and overall conversion rose with it. Only then could he say the sign-up change worked.

How long to wait depends on how much traffic you have. At 640 visits a day, a week gives you more than 4,000 visits and the numbers are stable enough to judge. At 50 visits a day, a week gives you only 350, the numbers wobble, and 80% this week versus 70% next week may be pure noise — do not rush to conclude. Small products should either wait longer, or go and ask users directly: where exactly does sign-up get stuck? Asking five people beats staring at wobbling numbers and guessing for a month.

After fixing, write it down: what changed, when, what the number was before, and what it is after. Without a note, three months later you will have forgotten how the number got there, and the next time you look at the funnel you will have to guess all over again. Ah Zhe's method is crude: one document, three lines per change — what changed, the number before, the number after. After three months, the document is a history of "how this product got better".

What to look at every week once the funnel is built

A funnel is not a one-off build; it is a table you read every week. Pick a fixed time each week and look at three things. First, did any step change suddenly compared with last week? A sudden change usually means the product changed, not the users: a release last week altered the sign-up flow or added a step to onboarding. Check the release notes before concluding anything. Second, is the leakiest step still the same one? If the first step has been the leakiest for three weeks running, the fix has not worked, or it was aimed the wrong way — try a different angle. Third, did the step you fixed recover? One week after the fix, if the number has not moved, the fix was not aimed at the root cause; if it moved, write it down in the change log.

There is one more rule for the weekly review: do not read it alone. Bring the people who build the product and the people who write the code. With the numbers on the table, nobody can say from feeling that "the sign-up flow is fine" or "users are just lazy". Ah Zhe's team spends 15 minutes every Friday afternoon looking at the funnel, discussing only one question each time: which step do we fix next week. Fifteen minutes is short, but it never slips. Over three months they fixed the sign-up fields, the onboarding length, and the reminder timing for the first focus session, and overall conversion rose from 33.4% to 41% — not from one big redesign, but from fixing one step a week, slowly.

Four traps to avoid

The first trap is double-counting. One person refreshing a page fires the visit event twice; one person retrying a failed sign-up three times counts three sign-up events. When instrumenting, decide: does one step count a person once, or every time it fires? For sign-up, once per person is the reasonable call; for page visits, counting every visit is fine — but write the definition down. If the definitions differ, the numbers cannot be compared.

The second trap is mixing new and returning users. New users walk the "visit — sign up — onboard — first focus" road; returning users go straight to using it. Count returning users in the "visit to sign-up" step and the sign-up rate gets dragged down — returning users never needed to sign up. A funnel has to count "the same group of people": for new users, count only new users; for returning users, build a separate return funnel. Mix them and the number is wrong by construction.

The third trap is looking only at totals and ignoring time. Funnel numbers are computed over a window: 640 visits this week, 700 next week, and conversion moves with the headcount. Compare using the same window — both weeks, for example — and do not compare this week's sign-ups with last month's visits. The steadier method is to compute by cohort: of the people who signed up in the same week, how many came back a second time within 7 days — that way everyone's time window is the same and the numbers stay clean.

The fourth trap is putting private data into events. When instrumenting, record only "what happened", never "who it is, where they live, or their phone number". Use a de-identified user ID, never a raw email or phone number. The cleaner the data, the smaller the risk; if you genuinely need to analyze individual behavior, aggregated data is enough — there is no need to bury private fields in events.

If there is one thing to do this week, do this: draw your funnel and look at how many people the first step loses. Many teams have never drawn one — visits, sign-ups, and retention are each read separately and never lined up. Line them up once, and the problem takes shape for the first time: people are not "slowly leaking away"; they are dropping off in force at one particular step. Find that step, fix it, and look again next week. That is the whole job of funnel analysis.

One last line: the end point of funnel analysis is not a chart; it is the step you fixed. However beautiful the chart, if nobody fixes anything, the drawing was wasted. Read the funnel once a week, fix only the leakiest step each time, and look at the numbers again after a week. After three months, you will find the product did not get better from one big redesign — it got better from fixing one step a week, slowly.

Views 10Comments 0

Comments (0)

ME
0/1000
Loading comments...
Browse projectsPublish your project

Related articles

Real photo: a bank of light gray metal filing cabinets with one drawer pulled open, revealing rows of hanging file folders inside
Guide
A real database inside the browser: IndexedDB local storage, explained plainly

localStorage only stores strings and has no indexes. IndexedDB is a real database in the browser: structured data, far larger capacity, indexes, and transactions. Following MDN's Using IndexedDB, IDBDatabase, and IDBObjectStore: the five-step routine, opening, upgrades, keys, indexes, transactions, add versus put, get, cursors, ranges, and multi-tab upgrades.

Tool TipsData WorkflowsFrontend Engineering
Real photo: a developer pointing at code on a laptop screen, with two more monitors showing code behind it
Guide
When the program crashes, catch the error first: Python exception handling

An exception is not a broken program: it is Python telling you what went wrong, with a name you can catch. Following Chapter 8 of the official Python Tutorial: syntax errors vs exceptions, try/except, matching by inheritance, which errors to catch and which to leave alone, raise, else, finally, with, user-defined exceptions, chaining, and ExceptionGroup — plus a 100-accident breakdown from a fictional bookstore script.

Project BuildingPythonLearning & Career
Real photo: the aisle of an independent bookstore, wooden shelves packed with colorful book spines under warm light
Guide
Turn a book into an object and its rules into a class: Python classes and objects

A class bundles data and its rules together: one Book class gives every book its own state, methods let objects do the work themselves, and inheritance lets an e-book stand on a paper book's shoulders. Following Chapter 9 of the official Python Tutorial: class and __init__, self, methods, class vs instance variables, attribute lookup, inheritance and overriding, super(), the underscore-private convention, dataclasses, __iter__ and __next__ — plus a 100-accident chart.

Project BuildingPythonLearning & Career