Product & DesignFree
Product Funnel and Retention Analysis
Every step ranked by the number of people it actually loses, not by its conversion rate, so the fix goes where the users are.
River reads the export rather than the dashboard screenshot. Funnel and retention data from Amplitude, Mixpanel or PostHog goes in. Every step comes back with three numbers side by side: the conversion rate, the absolute users lost, and the share of the starting cohort that ever reached it. The step with the worst rate and the step that loses the most people are usually different steps, and the report names both rather than letting the ranking pick for you.
A funnel is not one number, it is a number plus a definition, and the definition lives in a dropdown. Amplitude lets the same three steps be counted in this order, any order, or exact order, which produces three different conversion rates from identical data. PostHog puts the trap more bluntly: think of funnel steps as actions to be performed, not filters to be applied. Every report here states the definition it used, in the header, above the numbers that definition produced.
Built for the product manager staring at a funnel chart with no idea which step to take to engineering, and for the growth team whose retention curve flattened and nobody knows where. Reach for it before you scope the fix, not after the ticket is written. Once you know the step, sizing the test is the experiment design brief. Reading the result afterwards is the experiment readout. Why users behave that way comes from research synthesis.
Why the worst-converting step is often the wrong fix
Conversion rate is a ratio and it hides its denominator. A step converting at 34 percent looks like the emergency, and if only 900 people reach it, fixing it perfectly recovers under 600 users. A step converting at 88 percent that 40,000 people reach loses 4,800. The percentage says one thing and the arithmetic says the other, and the arithmetic is what shows up in revenue. Ranking by absolute loss puts the two in the right order without any judgement being applied.
Retention has the same problem in a different shape. A curve that drops from 100 to 38 percent by day seven and then flattens is two findings, not one: an onboarding problem and a habit problem, with different owners and different fixes. The flattening point matters more than the day-one number, because that is where the users who stay actually are. The report gives the curve by cohort and by segment, and names the day the slope changes.
Then there is the segment that makes the average meaningless. A funnel converting at 61 percent overall can be 84 percent on desktop and 39 percent on mobile, and no amount of work on the shared flow will close that. Splitting by device, plan, acquisition source and first-session behaviour is where the actionable finding usually is, because a gap between two segments points at a specific cause in a way that an overall rate never does.
How it works
Export the funnel
Whatever your analytics tool gives you, including the raw step counts rather than only the percentages.
Pin the definition
Step order, conversion window and counting method, recorded so the number can be reproduced later.
Count the losses
Absolute users lost at every step, alongside the rate and the share who reached it.
Split the segments
Device, plan, source and cohort, with the splits whose gap is larger than the fix.
What you get
- Every step with its conversion rate, its absolute users lost, and its reach
- Steps ranked by people lost, so the biggest leak is the top row
- The funnel definition stated: step order, conversion window and counting method used
- Conversion by device, plan and source, with the segment gaps that beat the average
- Retention curves by cohort, with the day the slope stops falling named
- The plausible causes for the biggest loss, and what would confirm each one
Common questions
Amplitude and Mixpanel give me different numbers for the same funnel. Which is right?
Both, under their own definitions. The step order setting alone changes the answer, and the conversion window changes it again, so two tools configured differently will disagree on identical events. The report states which definition produced each number and recomputes both ways where the gap is large enough to change a decision.
My funnel export is a gzipped NDJSON dump. Can it read that?
Yes, and that shape is often better than the chart CSV because it keeps the per-event rows the chart already collapsed. Group-by limits on funnel tables mean the aggregated export frequently drops the segment you needed. Hand over whichever you have, and the report says which questions the format you chose can answer.
Will it tell me why users drop out?
It gives the plausible causes and what would confirm each, which is as far as behavioural data honestly reaches. A step losing 4,800 people on mobile and 200 on desktop points somewhere specific. Confirming it takes interviews or a session recording review, and the report says which of its candidates is worth that time.
How should I read a retention curve that never flattens?
As a product people finish rather than a product they abandon, or as a cohort too young to have flattened yet. The report checks the second before concluding the first, by comparing cohorts with enough elapsed time. Where no cohort is old enough, it says the curve cannot be read yet rather than reading it.
Does it handle funnels where users legitimately skip a step?
That is what the step order setting is for, and getting it wrong is the most common funnel bug. An optional step counted in strict order drops everyone who skipped it, and the funnel reports a collapse that never happened. The report checks for that pattern and flags any step whose loss looks like a definition artefact.
What do I do with the output?
Take the top row to engineering with the number of users attached, since a step losing 4,800 people gets scheduled and a step converting at 88 percent does not. Then size the fix as a test before building it, which is the experiment design brief. Where the leak is a feature people try once and never repeat, read the feature adoption review instead.
Product Funnel and Retention Analysis
Fill in the form and your workspace opens with the work already underway.