Predictive Customer Journeys: How to Use Path Analysis and Propensity Models to Grow Revenue

Predictive customer journey analytics gets sold as a magic box that tells you what every customer will do next. It's more useful, and more modest, than that. Done well, it tells you which paths lead to revenue, which customers are likely to buy, leave, or upgrade, and which action is worth taking with each of them. Done badly, it produces confident charts that point your budget at people who were going to buy anyway.

Here is the short answer. Start with clean, connected data on who did what and when. Use path analysis to see which sequences of touchpoints come before conversion and where people drop off. Use propensity models to score customers on the outcome you care about. Then use next-best-action rules that weigh the value of each action against its cost, and prove the gain with a holdout group. The last step is easy to skip, and it's the only one that tells you whether the model made you money.

1. What is predictive customer journey analytics, in plain terms?

It's three related tools. Path analysis looks backward: it maps the sequences of touchpoints customers went through and shows which sequences tend to end in a purchase, a booked call, or a cancellation. Propensity modeling looks forward: it uses past behavior to estimate each customer's probability of doing something specific in a set window, such as buying again in the next 60 days or canceling in the next 90. Next-best-action is the decision layer: for each customer, given their scores, which message, offer, or channel is worth using, and which is worth skipping.

None of this replaces strategy. If you are hiring a statistician or modelers, make sure they understand marketing strategy as well. You can make data say anything you want it to say, so models must be set up with a solid strategic mindset. The model answers the question you give it, so the question has to be the right one.

2. What data do you need before you start?

You need four things, and the first one is the hardest. First, a consistent customer ID that connects web and app behavior, CRM records, purchases or contracts, and support interactions. Without it, you're modeling fragments of people. Second, event data with timestamps: page and product views, form starts and submits, email opens and clicks, calls booked, purchases, renewals, cancellations. Third, outcome data you trust, meaning purchases and revenue that tie back to your billing or ecommerce system, not just what an ad platform reports. Fourth, enough history to cover a few full buying cycles, so the model sees seasonality and the full path from first touch to purchase.

Check the boring things before any modeling. Do GA4, your CRM, and your billing system agree on how many customers bought last month? Are UTMs consistent? Are offline touches like sales calls and events logged anywhere? If you sell in healthcare or another regulated category, decide with your compliance team up front which data can be used for modeling and what must never be passed to ad platforms. A model built on broken tracking will be precise and wrong. If your leads are already falling through the cracks before anyone models anything, my post on why your leads aren't converting is a better place to start.

3. How does path analysis work, and what can it actually tell you?

Path analysis groups customers by the sequence of steps they took and compares outcomes across those sequences. It's very good at finding drop-off points and at showing which content or steps tend to appear in converting journeys.

Example (illustrative numbers): in one quarter, 10,000 new visitors arrive and 280 become customers, a 2.8% conversion rate. Of those visitors, 800 viewed the pricing page and then a case study, and 96 of them converted (96 ÷ 800 = 12%). The other 9,200 visitors produced 184 conversions (184 ÷ 9,200 = 2%). It's tempting to conclude that pushing everyone to pricing and then a case study would multiply conversions by six.

It won't, and this is a classic trap in journey analysis. People who seek out pricing and case studies were already more serious buyers. The path partly reflects intent that was already there. What path analysis gives you is a hypothesis: maybe surfacing a case study near pricing helps. You then test it by showing the case study to a random half of pricing-page visitors and comparing conversion. Path analysis tells you where to look. Experiments tell you what works.

4. How do propensity models work, and how do you judge them?

A propensity model estimates, for each customer, the probability of a specific outcome in a specific window. Logistic regression is often the right starting point because it's transparent and you can explain what drives the score. Tree-based methods such as gradient boosting can squeeze out more accuracy on larger datasets, at some cost to explainability. Clustering is a different tool, useful for finding segments, but it doesn't predict an outcome on its own.

Judge a model on held-out data it never saw in training, and judge it in business terms. The most useful view is a lift table by decile. Example (illustrative numbers): you score 20,000 customers on their likelihood to buy again in the next 60 days. The base rate is 5%, so you expect about 1,000 repeat buyers. When you check the model's top-scored 10% (2,000 customers), 300 of them bought, a 15% rate (300 ÷ 2,000). That's a lift of 3.0 (15% ÷ 5%), and that decile captured 30% of all repeat buyers (300 ÷ 1,000) while being 10% of the list. That's a model that ranks customers well.

Repeat these analyses. This is not a set it and forget it system. Customer behavior, your offer, and your channels all shift, and a model trained a year ago can quietly stop ranking well. Track the top-decile lift every month or quarter, and retrain when it slips.

5. Why isn't a high propensity score the same as a good target?

This is where predictive journey work can quietly lose money. A propensity model tells you who is likely to buy. It doesn't tell you who is likely to buy because you contacted them. Those are different groups, and the second one is the one you pay for.

Example (illustrative numbers, continuing from above): you send a $10 incentive to the top decile and hold out a random group for comparison, 1,000 customers each. The treated group buys at 17% (170 orders), and the holdout buys at 15% (150 orders). The campaign produced 20 incremental orders, not 170. At $80 of margin per order, that's $1,600 of incremental margin against $10,000 in incentive cost (1,000 × $10), a net loss of $8,400. And a 2-point difference with 1,000 people per group isn't even statistically reliable: the standard error of the difference is about 1.6 points, so the z-score is about 1.2, well short of the 1.96 you'd want.

The fix is to measure uplift, the difference between treated and held-out customers, and to target the people whose behavior actually changes. They may well sit in the middle deciles. The same logic applies to churn: the customers most likely to leave aren't always the ones a save offer can keep. My guide to calculating customer churn covers how to measure the baseline before you build a churn model on top of it.

6. How does next-best-action work?

Next-best-action compares the expected value of each possible action for a customer, including doing nothing, and picks the best one. The math is simple. For each action, the expected value is the probability the customer responds times the margin, minus the cost of the action. You then compare that to the do-nothing baseline.

Example (illustrative numbers): a customer has a 3% chance of buying in the next month with no contact. At $80 of margin, the do-nothing value is $2.40 (0.03 × $80). A reminder email lifts the probability to 4% and costs $0.10, for an expected value of $3.10 (0.04 × $80 − $0.10), which is $0.70 better than doing nothing. A $15 discount lifts it to 9%, but every buyer gets the discount, so the value is $5.85 (0.09 × ($80 − $15)), which is $3.45 better than doing nothing. For this customer, the discount wins. For a customer whose baseline is already 12%, the same discount gives away $15 on many orders that would have happened anyway, and the email may come out ahead. Those response probabilities have to come from holdout tests, not guesses, or the whole exercise is decoration.

7. Where does predictive journey work usually go wrong?

Watch for six failure points. Data leakage, where the model uses information that's only available after the outcome (like a cancellation reason field), which makes it look brilliant in testing and useless in practice. Selection bias, as in the path example, where you mistake a sign of intent for a cause. No holdout, so you can't separate incremental revenue from revenue you'd have had anyway. Optimizing a proxy, like email clicks, instead of margin or retention. Drift, where nobody checks whether the model still ranks well. And tracking gaps from consent rules, browser limits, and offline sales that never reach the data, which quietly bias which journeys you can see at all.

8. How does this fit with the rest of your measurement?

Journey and propensity models work at the customer level. They're good at deciding who gets which message and when. They're weak at telling you how much to spend in each channel, because they only see the touchpoints you can track. For budget allocation across channels, you need a top-down view like marketing mix modeling, which uses spend and outcomes over time and doesn't depend on tracking every click. The two work best together: the mix model sets the channel budgets, and the journey models decide how to use them with each customer. That combination is what my marketing mix modeling and analytics work is built around.

If you have the data but aren't sure whether a predictive model would pay for itself, or you have a model and aren't sure it's making you money, I can help you sort that out. You can book a free 30-minute introductory call here.

Dr. Linda Orr, PhD, is the founder of Orr Consulting, a fractional CMO and marketing analytics practice serving brands from $5M to $500M in revenue. She built Upwork's first enterprise marketing mix model, covering an $88M annual media portfolio, and spent about 20 years as a marketing professor before moving fully into consulting.