Generative Marketing Mix Modeling: A Causal Inference Framework Linking GEO and GEM to Business Impact

Traditional marketing mix modeling uses channel inputs recorded at the market-period level, while GEO and GEM exposure is not directly captured by conventional advertising logs. The target problem is to represent these generative-media exposures at that scale and evaluate their causal total effects on sales, conversions, and related business responses. The evaluation also concerns answer-based exposure, sponsored-display exposure, treatment sequences, carryover, saturation, measurement uncertainty, and alignment across the analysis unit, target population, and evaluation window.

GMMM inserts a generative-media measurement and counterfactual-sequence layer before standard marketing mix modeling. For GEO, it constructs expected noticed answer-feature occurrences from question counts, generation-system shares, feature appearance probabilities, and attention probabilities. For GEM, it constructs expected noticed sponsored displays from spending, display counts or a model for them, and attention probabilities. These market-period inputs are passed through carryover and Hill saturation transformations and then into a business-response model, together with source states and a complete GEO/GEM treatment sequence. Each sequence is mapped to a response, allowing comparison through V(a_G,a_P) while retaining source-state pathways and GEO-GEM interactions. The GEO total effect includes both pathways through changes in answer inputs and pathways through other variables caused by source-material modification.

Identification is formulated through a longitudinal g-formula under consistency, sequential exchangeability, positivity, identifiable observed-data laws on the target support, and specified aggregation of interference. Measurement data, business responses, and treatment effects must share the same analysis unit, target population, and evaluation window. The contribution also describes plug-in, two-stage, cut, and joint estimation procedures, alongside analyses of carryover, saturation, and measurement-parameter uncertainty. Simulated studies examine feature appearance rates, model and language differences, and numerical behavior under different estimation settings. Applying the framework therefore depends on collecting and aligning the required question, answer, attention, spending, and display information.

A single source state identifies only its appearance probability. When corresponding question, answer, and attention observations are missing, existing referral data cannot construct GMMM inputs, and expected counts are not equivalent to unobserved actual display counts. GEO comparisons fix question counts and system shares and assume that source-material changes do not alter attention conditional on feature appearance; identifying such changes requires observations under both source states and adequate target-population support. GEM display counts may depend on a spending model, and carryover requires initialization history. Behavior and identification beyond the stated generators, populations, treatment designs, interference structures, or unobserved-display settings remain undetermined.