A new Generative Marketing Mix Modeling framework translates AI-generated answers and sponsored placements into measurable advertising inputs so firms can estimate the total business impact of Generative Engine Optimization and Marketing within standard marketing-mix models; the approach is well-specified and simulation-validated but not yet tested on real-world sales or conversion outcomes.

Main Finding

The paper introduces Generative Marketing Mix Modeling (GMMM), a causal-inference framework that augments standard marketing mix models (MMM) to measure and estimate the business impact of (i) Generative Engine Optimization (GEO) — changes to source material that alter nonsponsored content in generative-model answers — and (ii) Generative Engine Marketing (GEM) — sponsored placements in generated answers. GMMM constructs expected noticed exposures from generated answers and sponsored placements, applies standard carryover and saturation (Hill) transformations, and gives identification conditions and estimation strategies (plug-in, joint, modular/cut-Bayesian). In simulations using GPT-5.6 Luna and GPT-4o answer collections, accurate estimation of carryover and saturation substantially improves coefficient recovery; allowing response data to update measurement parameters (joint vs cut posterior) has no uniform advantage for estimating the GEO treatment effect.

Key Points

• Problem addressed
  • Standard MMM inputs (impressions, clicks) do not capture nonsponsored mentions of a firm’s name inside generative-model answers, nor whether users notice sponsored placements. Without constructing these inputs, GEO and GEM cannot be analyzed credibly in MMM.
• Constructed inputs
  • GEO input EG_it(z): expected number of generated answers in unit i, period t, under source state z ∈ {0,1}, where a specified property (e.g., firm name cited) both occurs and is noticed: EG_it(z) = sum_q N_iqt * sum_p ω_ipt * π_qpt(z) * λ_qpt
  • N_iqt: question counts by cluster q
  • ω_ipt: share handled by generative system p
  • π_qpt(z): occurrence probability of the property in answers (depends on source state z)
  • λ_qpt: notice probability conditional on occurrence (assumed independent of z)
  • GEM input EP_it: expected number of noticed sponsored placements: EP_it = μP_it * ρP_it
  • μP_it modeled from spending via a Poisson link (log μP_it = ϕ0 + ϕS log S + …), ϕS>0
  • ρP_it: notice probability for placements
• Modeling and transformation
  • Both constructed counts enter MMM after the same carryover (normalized geometric lags) and Hill saturation transforms used for established channels.
  • The full response model includes conventional channels, constructed GEO/GEM channels, a direct source-state term βD Z_it (to capture off-answer effects like search/landing-page changes), and interaction between GEO and GEM.
• Measurement models
  • π_qpt(z) estimated via hierarchical logistic model (repeated answers per question/system/source state) with question-cluster random effects.
  • λ_qpt estimated via a binomial model from user-notice studies, also hierarchical.
  • Sponsored-placement delivery modeled with Poisson conditional on spending; notice probability estimated from user studies or observed delivery × noticed counts.
• Treatment effect definition and evaluation
  • Defines V(aG,aP) = sum_i sum_{t∈TE} E[Y_it(aG,aP)] for complete treatment sequences (carryover-aware).
  • GEO effect ΔG = V(1,1) − V(0,1); analogous ΔP for GEM.
  • Emphasizes comparison of entire sequences because carryover makes responses depend on past inputs.
• Identification & modular inference
  • Characterizes sufficient conditions under which treatment effects are identifiable even if coefficients on the source-state indicator and the constructed GEO input cannot be separated after media transformations are fixed.
  • Discusses when replacing expected market counts by occurrence probabilities reduces to a scale change and when market composition breaks that reduction.
  • Explores estimation strategies:
  • Plug-in: estimate measurement params from measurement data, then fix them and estimate response parameters.
  • Joint: estimate measurement and response parameters jointly.
  • Cut (modular) Bayesian: allow measurement data to determine measurement-parameter posterior and prevent response data from updating that posterior; still permit response data to inform transformation parameters if desired.
  • Uses Laplace approximation for measurement-module posterior and importance weights in joint computations.
• Empirical / simulation evidence
  • Data: 2,240 answers (20 repeats × 56 questions) from GPT-5.6 Luna and GPT-4o; target name occurred in 33.8% (GPT-5.6) and 27.8% (GPT-4o) of answers (English/Japanese split noted).
  • Simulations use these observed occurrence rates as baselines and generate treatment effects; results:
  • Estimating carryover and saturation (vs fixing them in a plug-in) explains most of the improvement in coefficient recovery.
  • Allowing response data to update measurement parameters (joint vs cut) does not uniformly improve estimation of the GEO treatment effect.

Data & Methods

• Data required to construct GEO/GEM inputs (DM)
  • Repeated generated answers across question clusters, generative systems, and source states (to estimate π_qpt(z)).
  • Counts of relevant questions by market and period (N_iqt).
  • Shares of usage across generative systems by market and period (ω_ipt).
  • User-notice measurements for generated-answer features and sponsored placements (λ_qpt and ρP_it), from lab or field user studies collecting notice indicators.
  • For GEM: spending records and, if available, observed delivery counts LP_it; otherwise model delivery using Poisson conditional on spending.
• Response data (DY)
  • Standard MMM data: aggregate responses Y_it (sales, conversions) by unit and period, baseline controls W_it, and observed established-media inputs Em_it.
• Statistical models
  • Occurrence: X_qptr ∼ Bernoulli(π_qpt), logit(π_qpt) = A_qpt^T α + u_q (hierarchical zero-sum random effects).
  • Notice conditional on occurrence: C_qpt ∼ Binomial(n_qpt, λ_qpt), logit(λ_qpt) = D_qpt^T ξ + v_q (hierarchical).
  • GEM delivery: LP_it | S_it > 0 ∼ Poisson(μP_it), log μP_it = ϕ0 + ϕS log S_it + ϕ_D D_it + ϕ_R R_it.
  • Carryover: normalized geometric weights over L lags (parameters α_m); saturation: Hill function H = A^κ /(A^κ + θ^κ).
  • Response: g_Y(μ_it) = b(W_it; γ) + sum_m β_m H_mit + β_D Z_it + β_G H_Git + β_P H_Pit + β_GP H_Git H_Pit (+ interactions among established media optionally).
• Estimation & computation
  • General criterion J = JM(DM|ηM) + JY(DY|ηM,ηT,ϑ) + penalties; covers ML, penalized, and Bayesian formulations.
  • Bayesian implementation:
  • Fit measurement module and approximate its posterior by Laplace qL(ηM|DM); sample η = (ηM,ηT) from qL(ηM|DM) p(ηT).
  • Joint posterior weights samples by approximate marginal likelihood p(DY|η).
  • Cut posterior keeps qL(ηM|DM) fixed and conditions response estimation on each ηM without letting DY update ηM.
  • Diagnostics: simulated recovery studies and comparisons between plug-in, joint, and cut approaches.
• Practical assumptions and limits
  • Assumes notice probability λ does not depend on source state z.
  • Requires data variation in source state to identify the change in occurrence probabilities π(z).
  • Aggregation unit must be common to media construction and response measurement.
  • When delivery counts are missing, requires spending → expected placements model to be valid.

Implications for AI Economics

• Measurement of generative-AI exposure is essential
  • Generated answers produce nonsponsored brand exposures not captured by conventional ad logs; firms and analysts must instrument for these exposures (question volumes, system shares, occurrence and notice probabilities).
• Causal inference for GEO/GEM requires thinking about sequences
  • Carryover effects mean treatment comparisons must replace entire media-input sequences; naive coefficient interpretation (e.g., coefficient on G-input) can misstate the total causal effect.
• Design and data-collection priorities
  • Collect repeated-answer datasets across source states and systems and invest in notice studies (lab or large-scale panels) to estimate λ and ρ reliably.
  • Track question/query volumes and system shares by market and period to convert per-answer occurrence probabilities to market-level expected noticed counts.
  • Randomized or quasi-experimental evidence remains valuable: where available, it can be incorporated into the likelihood/posterior for stronger identification of treatment effects.
• Attribution and channel interactions
  • GEO can affect outcomes both through constructed GEO inputs and off-answer channels (search traffic, landing pages); models must allow an explicit source-state coefficient βD to capture that.
  • GEM and GEO can interact in their effects (β_GP), and both should be modeled jointly in MMM to avoid misattribution.
• Methodological guidance for practitioners
  • Estimating transformation parameters (carryover, saturation) from response data matters materially; plug-in fixes of these transformations degrade recovery.
  • Modular (cut) Bayesian procedures are defensible when measurement-module data are of different provenance or when one wants to avoid response data updating measurement parameters; joint updating may not always yield better causal estimates for treatment effects.
• Broader economic research implications
  • For market-level welfare and policy analyses (e.g., competition, ad regulation), GMMM provides an explicit way to map algorithmic ranking/answer changes into expected consumer exposures and hence into economic outcomes.
  • The framework clarifies what data regulators or platforms would need to provide to support credible measurement of firm-level effects of generative-AI rankings and sponsored placements.
• Caveats and open issues
  • The approach relies on assumptions (e.g., notice independent of source state) and needs careful empirical validation.
  • Market composition changes, correlated unobservables, or misspecified delivery models can threaten identification; randomized interventions remain the gold standard where feasible.
  • Extending to richer user attention models (time to read, position bias within answers, multi-exposure effects) is a natural next step.

Overall, GMMM operationalizes how to translate model-generated content and sponsored placements into MMM-ready exposure measures, provides a causal-comparison framework that respects carryover, and offers practical estimation strategies (including modular Bayesian inference) for firms and researchers studying the economic impact of generative-AI channels.