Behavioral Modeling
Behavioral modeling is the practice of predicting and explaining customer decisions from recorded behavior: what people bought, when, in what sequence, after what exposure. It exists because stated intentions and recorded behavior routinely disagree. Customers report they would pay more for a feature and then do not, report price as the reason for leaving when the record shows declining usage months earlier, and report brand loyalty while the purchase history shows they buy whatever is discounted. When the two accounts conflict, the recorded behavior is the more reliable one. The size of the gap has been measured outside marketing entirely: a University of Chicago study of location data from more than two million cellphones found that while about 22% of Americans report attending religious services weekly, roughly 5% actually do (Pope, "Religious Worship Attendance in America: Evidence from Cellphone Data," 2024). People misreport even behavior they have no commercial reason to misreport.
How it actually works
The standard version is a propensity score from the analytics platform: likelihood to convert, likelihood to churn, built on whatever events the tracking recorded. The scores get used for targeting, and their inputs go unexamined. The model inherits every gap in the tracking, and the untracked steps in a journey, the store visit, the peer recommendation, the community thread, are exactly where many decisions form.
The quality of a model is capped by the behavioral record underneath it, and the record has to match the decision type. Impulse purchases and deliberate purchases produce different data shapes. An impulse journey is short: exposure, interest, purchase, sometimes inside an hour, and the model needs creative and context signals from that window. A deliberate journey runs weeks or months across research sessions, comparisons, and consultations, most of it invisible to a single session's analytics. Applying an impulse model to a considered purchase, or the reverse, produces confident scores about the wrong process.
Built on a complete record, behavioral models answer questions stated data cannot: which early behaviors precede churn, which sequence of touches precedes purchase rather than merely accompanying it, which segments look identical in a survey and behave differently at the register. The qualitative layer then explains what the model can only detect. The model finds that a cohort defects after the second purchase; the interviews find out why. The economic test sits alongside the statistical one: a propensity model earns its running cost only when the contribution margin per triggered intervention, weighted by the model's precision, exceeds the cost of scoring and acting on each prediction. A model that is right one time in three pays only where the margin on the correct third covers the interventions wasted on the other two.
In practice
At a company where one of our principals worked in-house, the team believed 30% of their customers were repeat buyers, a figure from a 30-day analytics window. The all-time repeat-purchase rate was 8%. The buyers who purchased once were personal and gift shoppers; the 8% who returned were franchise owners buying signage every month, and they represented 80% of the company's sales. Read correctly, the behavioral record reversed the growth strategy: the plan built on the 30% figure had produced 20 product launches that added almost no new revenue that year.
Where we come in
We build behavioral models on the full record instead of a tracking window: actual purchase history, journey shapes matched to impulse and deliberate paths, and qualitative research layered on to explain what the model detects. That approach took one client's customer acquisition cost from $52 to $13.50.
Start a Revenue Health Pre-Assessment →See it in action
Related terms
- Stated vs revealed preference
- What people report they would choose versus what their purchases show they chose. Behavioral modeling is built on the second.
- Propensity model
- A model scoring the likelihood of an action, purchase, churn, upgrade, from past behavior. Its quality is set by the completeness of the behavioral record it trains on.
- Impulse vs deliberate journey
- Two decision processes with different lengths, inputs, and data shapes. A model tuned for one misreads the other.
- Cohort analysis
- Tracking groups by when they joined, so behavior changes over time are visible instead of averaged away.
- Event tracking
- The instrumented record of user actions that behavioral models train on. What was never instrumented cannot be modeled, which is the method's standing blind spot.
- Survivorship in the record
- The behavioral record only contains people the tracking captured. The US Census Bureau's quarterly e-commerce report measures e-commerce as a share of total US retail sales, and the complement of that share, over 83%, is fulfilled in physical stores. Buyers who research and purchase offline never enter the record, so the model reads them as if they do not exist. We cover this in [offline customer journeys](/insights/marketing-measurement#offline-customer-journeys).