Customer Journey Mapping from Raw Clickstream

TL;DR: Most customer-journey maps are designer-led workshop artifacts: a handful of personas, a row of emoji moods, a smoothed path from awareness to purchase. They flatter the team that produced them and tell almost nothing about what customers actually do. The journey that lives in clickstream data is messier, less linear, and far more useful. The methodology to recover it has three pillars: sequence mining (which event orderings occur together, frequently enough to matter), Markov modelling (which transitions are likely, where the absorbing states sit), and bottleneck identification (which transitions throttle the population, and how the throttling shifts under intervention). This essay walks through each pillar, the academic lineage behind it, the partner-data patterns we have seen, and the tooling decisions that determine whether the resulting map is a map or merely a poster.

A note on the named companies. Amplitude, Mixpanel, and Heap appear throughout as well-known examples of three distinct event-analytics architectures. Quantitative figures (path conversion rates, bottleneck severity, time-between-step distributions) come from advisory work with anonymized partner operators across DTC retail, B2B SaaS, and marketplace archetypes, not from those tooling vendors themselves.

---

The Journey Map Hanging on the Wall Is Not the Journey

A typical journey-mapping exercise produces a wall-sized poster with five to eight columns. The columns are stages: awareness, consideration, purchase, onboarding, advocacy, sometimes with the cute additions of a "delight" or "wow" moment. Each column has a row for the customer's emotional state (usually represented by an emoji), a row for the channels they encounter, a row for pain points, and a row for "opportunities" identified by the workshop. The poster ends up framed in the design studio. It is referenced occasionally during stakeholder presentations.

The poster is not lying, exactly. The stages it identifies are real categories of activity. The pain points it lists are real pain points. The customers in question do experience awareness, then consideration, then purchase, and that ordering is not arbitrary. What the poster cannot show, by construction, is what the customer actually did. It cannot show that 40 percent of purchases happened without any awareness-stage touchpoint at all, because the customer came in through a referral. It cannot show that the consideration stage was three months long for half the population and forty-five seconds for the other half. It cannot show that the modal path between two stages is a recursion (the customer goes back to consideration twice before purchase), or that one specific page accounts for two-thirds of the abandonment in the purchase stage.

These are facts about the customer, not facts about the customer's emotional state. They live in clickstream data. The methodology to recover them is technical, requires actual engineering work, and is unglamorous in a way that workshop exercises are not. It is also the only methodology that produces a map you can act on.

---

The Three Pillars of Empirical Journey Reconstruction

The methodology has three pillars, and each one corresponds to an academic literature with decades of formal results behind it.

Pillar 1: Sequence mining. Given a corpus of event sequences (one per user), which sub-sequences occur frequently enough to matter? This is the question that the Agrawal and Srikant (1995) sequential pattern mining literature was built to answer. The output is a set of frequent event orderings, with their support (the fraction of users who exhibit the pattern) and their confidence (the conditional probability that one sub-sequence implies another). Tools like Amplitude Pathfinder and Mixpanel Flows are productised versions of this idea, though they typically expose only a narrow slice of the mining state space.

Pillar 2: Markov modelling. Treat the event sequence as a stochastic process: at each event, the probability of the next event depends on the current event (or on the last k events, for a k-th order Markov model). The estimated transition matrix is a journey map, except every cell is a probability and the whole structure is something you can simulate, run interventions on, and compare across cohorts. The lineage here is Markov, Ross's stochastic processes textbook, and the more recent literature on Markov decision processes for customer attribution.

Pillar 3: Bottleneck identification. Once you have a transition graph (from Markov modelling) and a set of recurring sub-sequences (from sequence mining), the operational question is which transitions throttle population flow. A bottleneck is an edge where a large fraction of users converge but only a small fraction proceed. The bottleneck has two parameters: its severity (the conversion gap) and its position (how upstream it sits in the path-to-purchase). The intervention question is which bottlenecks are responsive to product changes and which are stable features of the population. The methodological lineage is van der Aalst's workflow-mining school, originally developed for business-process discovery and now applied increasingly to consumer journey analysis.

The three pillars compose. Sequence mining narrows the analysis to the orderings that actually occur in volume; Markov modelling estimates the transitions and absorbing states; bottleneck identification picks the highest-leverage edges to intervene on. A journey map produced by composing all three is a structure, not a poster. It can be queried, simulated, A/B tested, and recomputed monthly.

---

Pillar One: Sequence Mining and the Frequent-Pattern Problem

Sequence mining is the question: given a collection of event sequences, which ordered sub-sequences occur in at least s percent of the sequences? The parameter s is the minimum support, and it is the single most important tuning choice in the entire methodology. Set it too low and you drown in noise patterns that one user emits and no one else does. Set it too high and you find only the most obvious patterns, which you already know about from common sense.

A reasonable starting point for consumer products is s between 1 and 5 percent. For a property with a million monthly active users, a 1 percent support floor means a pattern must be exhibited by at least ten thousand users to register. That floor surfaces patterns that are operationally meaningful (large enough to be worth changing the product for) without surfacing patterns that are artefacts of individual user idiosyncrasies.

The classical algorithm is Apriori-based: build candidate sub-sequences of length one, count their support, prune anything below the floor, extend the survivors to length two, repeat. The Han, Pei, Yin (2000) PrefixSpan algorithm and its descendants improved on this dramatically by avoiding the candidate-generation step and projecting the database directly onto prefix patterns. PrefixSpan is what most modern sequence-mining libraries use under the hood, though the user usually does not see this.

What the user does see is a list of patterns and their support and confidence. The interpretation rule is straightforward: support tells you how common the pattern is in the population; confidence tells you, given the prefix, how likely the suffix is to occur. A pattern [product-page] -> [add-to-cart] -> [checkout] with 12 percent support and 47 percent confidence means 12 percent of users emit this whole sequence, and 47 percent of users who reach the add-to-cart from the product page proceed to checkout. The first number is a sizing question; the second is the bottleneck question (and we will return to it under Pillar Three).

The mining output for a real consumer product is typically much more interesting than the workshop poster admits. We have seen, in partner data, that the modal purchase sequence is rarely the linear one. A representative composite from DTC retail partner analyses:

The first row is the canonical workshop pattern. It accounts for roughly a fifth of the volume. The other rows describe four-fifths of the actual sequence space, and most of them are invisible on the wall poster. The recursion patterns (rows two and four) and the hesitation loops (row seven) are the patterns that bottleneck interventions can move; the canonical funnel (row one) is the pattern the team is already optimising for. Sequence mining surfaces the unattended four-fifths.

---

Pillar Two: Markov Chains over Event Sequences

Sequence mining tells you which orderings are common. Markov modelling tells you which transitions are likely. The two are complementary: the first identifies patterns, the second estimates the dynamics.

A first-order Markov chain over events is the simplest version: estimate, for each pair of events (A, B), the probability that B immediately follows A in a user's sequence. The estimate is the count of A -> B transitions divided by the count of A events. The resulting transition matrix is the structure people usually call a "journey graph" when they want to sound technical, though most of the journey-graph visualisations in BI tools are first-order Markov estimates with the math hidden.

A second-order chain conditions on the last two events: P(C | A, B). The conditioning is more accurate (consumer behaviour is rarely truly memoryless) but the parameter space grows quadratically and the estimation requires substantially more data. For most consumer journeys, second-order is the practical sweet spot; third-order or higher tends to be over-fit and rarely outperforms second-order in held-out comparison. The Markov-order question is empirical: estimate the model at orders one through three, evaluate the likelihood on a held-out sample, and pick the order at which improvement stops.

The Markov view exposes structure that pure funnel analysis cannot. The most important structural feature is the absorbing state: an event from which no user transitions back into the active population, except in a small fraction of cases that are recoverable. Common absorbing states in consumer journeys: completed checkout (a positive absorbing state), session abandonment after long idle (a negative absorbing state), and unsubscribe (a brutal absorbing state). The probability of reaching each absorbing state from any starting event is a finite-state computation; the expected time to absorption from any starting event is the same. Together they give you a quantified version of the journey map that the wall poster could never produce.

The diagram represents a simplified first-order transition graph from composite partner data. Each arrow corresponds to a transition with non-trivial probability mass. The two absorbing states are Purchase (positive) and Abandon (negative). The structural feature worth noting is that Abandon is reachable from every non-absorbing state, which is what makes journey reconstruction harder than the typical funnel narrative admits. Customers do not flow linearly through stages; they leak out at every stage, and the leakage rate is the conversion problem.

The next step is to estimate the transition probabilities and then ask which transitions are responsive to product changes. The transition probabilities for the same diagram, in approximate composite form:

The chart makes the operational reality visible. The single highest-leverage transition (in the sense of having both high volume and the lowest conversion) is Product to Add-to-Cart at 18 percent. The single most-throttling stage is Product to Abandon at 66 percent. The team's intuition might be that Checkout to Purchase is the conversion-critical stage, but the data say Checkout to Purchase has 71 percent conversion (the highest in the chain) while Product to Add-to-Cart has 18 percent. The Product page is where the population thins out, and any product-change intervention is likely to have leverage there.

---

Pillar Three: Bottleneck Identification and the Intervention Question

A bottleneck in journey reconstruction is a transition where the population converges (high entry volume) but proceeds at low rate (low conversion). The product of the two quantities is the operational opportunity: improving conversion on a high-volume low-conversion edge moves more users than improving conversion on a low-volume high-conversion edge. This is mechanical but worth saying because the workshop poster usually identifies bottlenecks by intuition rather than by volume-weighted conversion gap.

The classical bottleneck-identification methodology comes from the workflow-mining school, which developed it for business-process discovery (van der Aalst and colleagues at Eindhoven, mid-2000s onward). The clinical version: for each edge in the transition graph, compute the volume (number of users entering the source state per period), the conversion rate (fraction of those users who proceed to the target state), and the bottleneck severity score (volume times one-minus-conversion, which is the user-count that fails to proceed). Rank edges by severity. The top severities are the operational priorities.

The methodology produces a ranked list of intervention candidates, which is more useful than the workshop's "biggest pain points." It is also empirically grounded: the severity score is computed from the data, not from the team's intuition about which pain points feel biggest.

The ranking surfaces the operational pattern: the top two severities are upstream of Add-to-Cart, not at Checkout. The instinct to optimise checkout (because it is the closest stage to revenue) is widespread, and in most partner properties it has the worst leverage of any optimisation target, because the checkout conversion is already high and the volume entering it is small. A team that reorders its optimisation backlog by severity score typically finds the priority list inverted from what intuition suggested.

The intervention question is the second half of bottleneck identification. Not every bottleneck is responsive to product changes. Some bottlenecks reflect product-market fit problems that are not fixable by tweaking page elements (the product is genuinely not interesting to two-thirds of the visitors). Some bottlenecks reflect channel-mix problems (the upstream channels are pulling in unqualified traffic, and the bottleneck is the system filtering them out, which is a feature, not a bug). The intervention question requires an A/B test, or a quasi-experimental design, to determine whether the conversion gap is responsive to the proposed change.

The empirical rule we use in advisory work: a bottleneck is "treatable" if a representative product-page intervention (copy change, social proof addition, image change) moves conversion by at least 10 percent of its baseline in a single test. If multiple representative interventions fail to move conversion, the bottleneck is structural (driven by product-market fit, audience composition, or category-level pricing) and the team should redirect effort elsewhere. The rule is heuristic, but it is the most consistent rule we have found for distinguishing responsive bottlenecks from structural ones.

---

A Worked Example: Reconstructing a Marketplace Onboarding Journey

To make the methodology concrete, consider a worked example from advisory work with a marketplace partner. The product was a two-sided B2C marketplace; the question was why supplier onboarding completion was around 38 percent when the team's intuition (and the workshop poster) suggested it should be closer to 60 percent. The team had spent three months iterating on the supplier signup form. The clickstream-derived analysis told a different story.

The first step was sequence mining over the supplier-onboarding events, with a minimum support floor of 2 percent. The output surfaced eleven distinct sub-sequences with non-trivial support. The most-trafficked pattern was the expected one: signup form, identity verification, listing creation, listing publish. It accounted for 31 percent of supplier sessions. The other ten patterns accounted for the remaining 69 percent, and most of them involved a recursion: suppliers were leaving the listing-creation step, going back to a help article, returning to the form, and frequently abandoning at that point. The recursion patterns were invisible on the funnel chart, which only showed the forward direction.

The second step was estimating a first-order Markov model over the same events. The transition matrix gave the volume-weighted absorbing probabilities: from the signup state, the probability of reaching Publish was 38 percent (matching the team's headline number); the probability of reaching Abandon-after-help-article was 27 percent (a state that the team had never tracked as terminal because it looked recoverable). The expected time from signup to publish was 18 hours for the 38 percent who completed, with a median of 47 minutes and a long right tail. The expected time from signup to the help-article abandon state was 52 minutes, which meant the help-article visitors were giving up faster than the completers were finishing.

The third step was bottleneck identification. The top severity score was on the listing-creation step (37 percent of suppliers entered, 51 percent proceeded), which the team had already been investigating. The second highest severity was a transition the team had not been tracking at all: from listing-creation to help-article, suppliers had a 22 percent transition rate, and of those, only 11 percent returned and completed. The help-article visit was where the population was hemorrhaging, not the form itself.

The intervention was a redesign of the listing-creation step that inlined the contextual help previously living in the help article. The forecasted change (a 35 percent reduction in help-article transitions, a 12 percent improvement in listing-creation completion) was registered before the test. The realised change was a 41 percent reduction in help-article transitions and a 9 percent improvement in listing-creation completion. The composite supplier-onboarding completion went from 38 to 47 percent, which moved the marketplace's supply growth rate by approximately the same amount.

The worked example illustrates two things. First, the bottleneck was not where intuition placed it; the team had spent three months on the form, and the substantive throttle was a transition out of the form to a help article that the workshop had never identified. Second, the forecasted-versus-realised comparison gave the team a calibration data point: the directional forecast was right, the magnitude was off by about 25 percent, and that magnitude error was lower than for most prior interventions. Calibration improves with practice; the discipline is to practice.

---

The Cohort Question and the Composition Confound

Everything above is computed over the full population. A first-order Markov model estimated on the whole user base implicitly assumes the population is homogeneous, which it is not. Different cohorts (new visitors versus returning, paid-channel versus organic, mobile versus desktop) have different transition probabilities, and aggregating them produces a journey map that describes no single cohort accurately.

The cohort-stratification question is its own analytical step. The clinical version: stratify the clickstream by cohort, estimate a separate transition matrix per cohort, and look for transitions where the cross-cohort variance is large. Transitions with low variance are population-level features (everyone behaves similarly); transitions with high variance are composition-sensitive, meaning the apparent rate depends on the cohort mix and will shift as the mix shifts.

We have observed, across partner properties, that the highest-variance transitions are typically the early-funnel ones (Landing-to-Category, Landing-to-Product) and the lowest-variance transitions are the late-funnel ones (Checkout-to-Purchase). The pattern is intuitive in retrospect: early-funnel transitions reflect channel and intent composition, which differ wildly by cohort; late-funnel transitions reflect platform mechanics (payment processing, address validation), which are roughly constant across cohorts. The composition confound is therefore concentrated at the top of the funnel, which is exactly where most teams aggregate aggressively.

The table reveals the diagnostic: the early-funnel transitions span ranges of 30 to 50 percentage points across cohorts; the late-funnel transitions span ranges of under 10 percentage points. A journey map that does not stratify by cohort will report the population mean and miss the bimodality. The operational implication is that the same product-page change can produce a 5 percent lift in the population-mean conversion and a 15 percent lift in one cohort and a negative lift in another; the population-mean is an artefact of the composition.

The cohort-stratification practice we use in advisory work: stratify on the two highest-variance dimensions (typically channel and device-type), estimate transition matrices per stratum, and report both the population aggregate and the cohort-level matrices. The intervention forecasts are made at the cohort level; the population-level realised effect is then computed by composition. This is more work than a single transition matrix, and it is the difference between a journey map that describes the population and one that describes no one.

---

The Tooling Question and the Anti-Pattern of Path Explorers

Most analytics tools (Amplitude, Mixpanel, Heap, Posthog) ship with a path-exploration feature. The feature draws a Sankey diagram showing the most-trafficked paths through a starting event, with thickness proportional to volume. It is the productised version of pillar one, sequence mining, with a fixed minimum-support floor and a fixed display order.

The path-exploration features are useful as a starting point, and they are typically misused in two specific ways. First, the path explorer answers a sub-set of the sequence-mining question (the modal patterns through a specific starting event) and is often interpreted as if it answered the whole question (which sub-sequences are common in the population). The two questions are not the same, and the explorer's narrowness causes teams to miss recursion patterns and reverse-direction flows.

Second, the path explorer hides the Markov estimate behind a visualisation. The Sankey diagram is a transition graph rendered as a flow chart, but the user does not see the transition probabilities directly, cannot compute absorption probabilities, and cannot simulate counterfactual interventions. For the bottleneck question, the user needs the transitions and the volumes, which the visualisation aggregates away.

The operational pattern that produces useful journey maps is to use the path explorer for orientation (which events are reachable from which starting points, roughly), then drop into SQL or Python to compute the sequence-mining output and the transition matrices directly. The SQL pattern over a clickstream table with user_id, event_name, and timestamp columns is a window function that lags the event name and groups transitions; the Python equivalent uses libraries like prefixspan for sequence mining and pomegranate or scikit-learn for Markov estimation. Neither is glamorous, but both are how the people who get useful answers actually work.

---

The Time Dimension and Conditional Transitions

Everything above treats the clickstream as a sequence of events with the timestamps reduced to ordering. That reduction is convenient but lossy. The time between events is informative: a one-second transition from Product to Add-to-Cart is a different signal than a forty-five-minute transition. The first looks like decisiveness (or arbitrage); the second looks like deliberation.

A complete journey reconstruction conditions on time-between-events. Three time-conditioning patterns recur in partner data. First, the same event sequence can have a bimodal time distribution: some users transition in seconds (impulse buyers, returning customers with purchase intent), others transition in days (research-phase customers). The two sub-populations behave differently after the transition and should be modelled separately. Second, the time-to-conversion has a survival-analysis structure (right-censoring is everywhere, because most observation windows end before the customer either converts or abandons), and the proportional-hazards literature (Cox 1972) provides the formal apparatus. Third, time-between-sessions is the signal that distinguishes returning customers from departed ones, and it underlies the entire churn-prediction sub-field.

The time dimension is the part of journey reconstruction that most teams under-invest in. The reason is that adding time to the analysis doubles the analytical surface area: every sequence pattern now has a time distribution, every transition has a duration, every bottleneck has a "how long before users give up" companion question. The under-investment is rational in the short term and expensive in the long term, because the time-conditioning patterns are typically where the highest-leverage interventions live.

The chart shows a composite cumulative-conversion curve for a fourteen-day window. The first six hours capture roughly a third of the eventual conversions; the first forty-eight hours capture about three-quarters. The long tail (hours 120 through 336) adds the final tenth of conversions slowly. The operational implication is that the time-to-conversion distribution is heavily concentrated in the first two days, and any intervention that operates outside this window (a retargeting campaign that fires day five, an email that arrives day seven) is interacting with a much smaller residual population than the team usually assumes.

---

What the Journey Map Should Look Like

The journey map that results from this methodology is not a poster. It is a structure with the following components, which can be revisited monthly without redoing the workshop.

A list of frequent sub-sequences, with their support and confidence, ranked by support. The list typically has ten to thirty patterns at the 1 to 5 percent support floor; the team that has not seen this list before is usually surprised by half of them.

A first or second-order Markov transition graph with estimated probabilities. The graph can be rendered as a visualisation, but the underlying matrix should be available for queries (what is the probability of reaching purchase from Landing? what is the expected time to absorption from Add-to-Cart?). These are not visualisation questions; they are matrix operations.

A ranked list of bottlenecks by severity score (volume times conversion gap). The list typically has six to twelve transitions in the actionable range; the priorities should match the operational backlog of the conversion-rate-optimisation team.

A time-conditioning analysis on the top three bottlenecks, distinguishing fast-converters from slow-converters and identifying the inflection point in the time-to-conversion curve. The intervention strategy differs between the two sub-populations.

A counterfactual simulation capability: given a proposed intervention (a 10 percent improvement in the Product to Add-to-Cart transition), the framework can simulate the downstream effect on total conversion. This is what the Markov estimate enables that the wall poster cannot.

The structure is meant to live in a dashboard, a notebook, or a documented analysis. It is updated monthly as new clickstream accumulates. It is the empirical referent for any conversation about "the customer journey," and it is also the artefact that every workshop poster is implicitly trying to gesture at without ever being specific enough to be wrong.

The journey that lives in the wall poster is a model that cannot be falsified, because it never quite says enough to be wrong. The journey that lives in the clickstream is a model with parameters, and parameters can be measured, contested, and re-estimated.

---

The Operating Discipline

There is a quiet operating discipline that distinguishes teams that use journey maps from teams that produce them. The discipline has three habits.

First, the journey map is treated as an output, not an input. The team does not start with a hypothesised journey and then look for data to confirm it; they start with the clickstream and let the patterns emerge. The output is then compared with the workshop hypothesis, and divergences are documented, not explained away. The divergences are typically where the operational leverage lives.

Second, the map is recomputed on a schedule. A journey map produced in January and never updated is a journey map about January; if the product, the channel mix, or the customer composition has shifted, the map is stale by April. Most partner properties we have worked with recompute the map quarterly, with monthly checks on the top bottlenecks. The recomputation is cheap (the analytical machinery is automatable) and the freshness is essential.

Third, every product change that targets the journey is paired with a forecasted change in the map. The product manager proposing a redesign of the product page states the expected change in the Product-to-Add-to-Cart conversion (say, plus 15 percent of baseline) before the experiment runs. The forecasted change is then compared with the realised change, and the discrepancy is the team's calibration signal. Teams that practice this consistently develop, over several months, a much better sense of which changes are likely to move which transitions, which is the practical version of "knowing your customer."

---

Key Takeaways

1. The wall poster is a model that cannot be falsified. The clickstream-derived journey is a model with parameters that can be measured, contested, and recomputed.

2. Three pillars support the methodology: sequence mining (which orderings are common), Markov modelling (which transitions are likely), and bottleneck identification (which transitions throttle the population). Each has a decades-long academic literature with rigorous results.

3. The Agrawal-Srikant sequence-mining framework and its PrefixSpan descendants are the foundation; modern tools like Amplitude Pathfinder and Mixpanel Flows are productised, narrowed versions of the underlying machinery.

4. First-order Markov chains over events are the practical sweet spot for journey reconstruction. Second-order improves accuracy when data is plentiful; higher orders rarely justify the parameter inflation.

5. Bottleneck severity (volume times conversion gap) is a better intervention prioritiser than intuition about pain points. Teams that re-order their optimisation backlog by severity score typically find priorities inverted from the workshop's recommendations.

6. The path-exploration features in commercial tools answer a subset of the sequence-mining question and are often misinterpreted as if they answered the whole question. They are useful for orientation; SQL and Python are how the substantive analysis actually runs.

7. The time dimension (time-between-events, time-to-conversion) typically holds the highest-leverage interventions and is the part of journey reconstruction that most teams under-invest in.

8. The journey map should be an output, not an input; recomputed on a monthly or quarterly cadence; paired with forecasted changes before any product intervention runs. The forecasted-versus-realised gap is the team's calibration signal.