Propensity Modelling Replaces Rule-Based Segmentation

Hand-crafted segments ('users who viewed 3+ products') are a standard starting point, and propensity models are the stronger next step. Business intuition remains critical for feature selection, but predictive models find patterns that are hard to identify manually.

Train a gradient-boosted classifier on historical conversion data. Features: session count, recency, page category distribution, scroll depth percentiles, time-of-day patterns, referral source, device characteristics. Output: a probability score per user per session.

Segment on probability thresholds rather than arbitrary behavioural rules. Your model discovers interaction effects (mobile + evening + returning + product page = high propensity) that no analyst would manually encode.

BigQuery ML makes this accessible without separate infrastructure: CREATE MODEL with BOOSTED_TREE_CLASSIFIER trains directly on your warehouse data. Score new sessions with ML.PREDICT. The probabilities can feed Looker Studio dashboards, trigger re-engagement campaigns via audience export, or drive real-time personalisation.

Retrain monthly. Monitor feature drift with the Population Stability Index (PSI) and track AUC over time with ML.EVALUATE. When AUC degrades below 0.65, your model is no longer useful — retrain with fresh data or revisit your feature engineering.