Marketing generates enormous quantities of behavioral data — purchase histories, click streams, ad exposures, survey responses — but the business questions it must answer are fundamentally about unobserved quantities: future customer value, the causal effect of advertising, and the latent preferences that drive purchase decisions. Bayesian methods excel in this domain because they handle missing data gracefully, produce predictions with calibrated uncertainty, and discover latent structure in high-dimensional behavioral data.
Bayesian Customer Lifetime Value
Customer lifetime value (CLV) — the present value of all future profits from a customer — is the central metric of customer-centric marketing. The BG/NBD (Beta-Geometric/Negative Binomial Distribution) model and its Bayesian extensions model two latent processes: how often a customer transacts while active, and when they permanently defect. Individual-level posterior distributions of transaction rate and defection probability produce personalized CLV predictions with uncertainty bounds.
λᵢ ~ Gamma(r, α) [heterogeneous transaction rates]
Defection after each transaction: P(defect) = pᵢ
pᵢ ~ Beta(a, b) [heterogeneous defection probabilities]
P(active at T | purchase history) computed via Bayesian updating
The posterior probability that a customer is still active — given their purchase history of frequency, recency, and tenure — enables managers to distinguish between a lapsed customer who will return and one who has permanently churned. Bayesian estimation handles the heterogeneity in customer behavior through the hierarchical prior, and full posterior inference provides prediction intervals rather than point forecasts.
Marketing Mix Models
Marketing mix models (MMMs) estimate the causal effect of advertising, pricing, promotions, and distribution on sales. Bayesian MMMs have become industry standard because they incorporate prior knowledge about advertising effects (e.g., diminishing returns, carryover effects), handle collinearity among marketing variables through regularizing priors, and produce posterior distributions of return on investment for each marketing channel.
Google's Bayesian MMM framework (Meridian) and Meta's Robyn use informative priors on advertising elasticities derived from meta-analyses of hundreds of marketing studies. These priors express the empirical finding that advertising elasticities are typically small and positive (median around 0.1), preventing the model from producing implausible estimates when data alone are insufficient to identify all effects. The Bayesian framework makes this borrowing of information explicit and transparent.
Customer Segmentation
Bayesian mixture models and latent class models discover customer segments from behavioral data. Unlike deterministic clustering (k-means), Bayesian mixture models estimate the posterior probability that each customer belongs to each segment, capturing the uncertainty in segment assignment. Bayesian nonparametric methods like the Dirichlet process mixture model automatically infer the number of segments from the data, avoiding the need to pre-specify k.
"The most valuable customer is not the one who spent the most last year — it is the one whose posterior expected future value is highest. Bayesian models make that distinction possible." — Peter Fader, Wharton School, co-developer of the BG/NBD model
Choice Modeling and Conjoint Analysis
Bayesian hierarchical choice models estimate individual-level preferences from discrete choice experiments (conjoint analysis). Each respondent's part-worth utilities are drawn from a population distribution, and the hierarchical Bayesian estimation borrows strength across respondents to produce stable individual-level estimates even from limited choice data. These models underpin product design, pricing strategy, and market simulation across industries from consumer packaged goods to automotive.
A/B Testing and Experimentation
Bayesian A/B testing has gained rapid adoption in digital marketing because it answers the question marketers actually ask: "What is the probability that variant B is better than variant A?" rather than the frequentist inversion. Bayesian sequential testing allows continuous monitoring of experiments without inflating false-positive rates, enabling faster decisions on website design, email subject lines, and ad creative. Thompson sampling — a Bayesian bandit algorithm — simultaneously tests and optimizes, directing more traffic to winning variants while still exploring alternatives.