Problem
A funnel shows that conversion drops between two steps, but not why. To find the cause you need to see what users do between those steps. Product logs make this hard: the number of unique events is large, the same action is tracked differently on iOS, Android and web, and renamed events are rarely backfilled. Even after cleaning, an extra event between A and B, a duplicated B or a swapped pair leave it unclear which paths are "the same".
Example from the article: conversion falls between "basket" and "shipping details", and the funnel gives no clue about the reason.
Solution in brief
Build a transition graph from raw events: nodes are events, edges are transition probabilities. Unlike a funnel, it shows loops and cycles. To keep it readable, compare groups in a diff graph, compress sessions into cluster types, define composite states, or infer hidden states with a hidden Markov model.
Questions this case answers
- How to find out why users drop off between two steps of a funnel when the event log has hundreds of unique events and inconsistent event names across platforms
- How to compare the behavior of two user groups on raw event sequences by a transition graph that shows loops and paths between any two events
- How to find out why users drop off between two funnel steps
- How to read a transition graph when the product has hundreds of unique events
Tools and methods
- Transition graph (Markov chain) — nodes are unique events, edge weights are transition probabilities
- Diff transition graph — one group's graph subtracted from another's
- Session clustering with a cluster-by-event heatmap — turns sessions into a small set of types
- Composite states with history — mutually exclusive states built from past activity
- Hidden Markov model — infers unobserved states such as intent
- Retentioneering — open source Python library for user behavior analysis
- Hopscotch — the speaker's UI platform built on similar ideas
Google or a one-time email code. No password.
