Generative Marketing Mix Modeling: A Causal Inference Framework Linking GEO and GEM to Business Impact
Masahiro Kato, Daiki Honma, Taka Kato
stat.ML · cs.AI · cs.LG · econ.EM · stat.ME
Quick verdict
Proposes GMMM linking AI-answer exposure to MMM with rigorous causal identification, but validation is simulation-only; real-data evidence is indirect.
Summary
The paper introduces Generative Marketing Mix Modeling (GMMM), extending marketing mix modeling to channels whose inputs are not directly logged: Generative Engine Optimization (GEO) and sponsored placements in generated answers (GEM). For GEO, the media input is the expected number of noticed occurrences of a firm's name, built from hierarchical-logistic occurrence probabilities, question counts, shares across generative systems, and notice probabilities; for GEM, spending maps through a Poisson delivery model times a notice probability. Treatment effects are defined over complete treatment sequences and identified via a longitudinal g-formula (Theorem 4.4) and rank conditions (Proposition 4.2) that allow the total effect to be identified even when component coefficients are not. Estimation compares plug-in, cut, and joint Bayesian posteriors via importance sampling with Laplace approximations. Simulations seeded by 2,240 real generated answers (GPT-5.6 Luna and GPT-4o, 56 English/Japanese questions; 'Glasp' appears in 33.8% and 27.8% of answers) show that estimating carryover and saturation explains most accuracy gains (vector RMSE 1.207 to 0.950), while jointly updating measurement parameters with response data gives no uniform benefit. Limitations: no real-panel application of the full model appears in the provided text; notice probabilities are thinly measured and market counts assumed known; relative effect error stays near 17.5%.
Review complete
Assessment at a glance
6
risk flags identified
● high 0
● medium 3
● low 3
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- GEO and GEM media inputs must be constructed as expected noticed occurrences — combining occurrence probabilities, question counts, system shares, and notice probabilities — before standard MMM machinery applies.
- Treatment effects over complete GEO/GEM treatment sequences are identified by a longitudinal g-formula under consistency, sequential exchangeability, positivity, interference, and observed-data-law assumptions (Theorem 4.4).
- The total GEO effect can remain identified even when the source-state and GEO-input coefficients are not separately identified, per the null-space rank condition (Proposition 4.2).
- Simulations seeded by 2,240 real generated answers show estimating carryover and saturation explains most coefficient-recovery gains, while joint updating of measurement parameters with response data has no uniform benefit for ∆G.
- A common scale factor in the GEO input is absorbed by the Hill midpoint (Proposition 4.3), so occurrence-probability sequences differ from market counts only under heterogeneous composition.
Strengths
Well-posed formalization of the measurement problem
Equation (6) decomposes the GEO input into question counts, per-system usage shares, occurrence probabilities, and notice probabilities, and Proposition 4.3 precisely characterizes when a probability sequence differs from a market count only by scale (absorbed by the Hill midpoint, rescaled to θ/c0).
Careful, honest identification analysis
Estimands are complete-sequence contrasts; Theorem 4.4 places the longitudinal g-formula on the sequences under explicit Assumptions 4.1-4.5, and Proposition 4.2's null-space condition distinguishes an identified total effect from unidentified component coefficients — including the correct warning that E[h(M)] ≠ h(E[M]) for expected versus realized counts (Section 4.3).
Reports negative and mixed results transparently
Fully specified, reimplementable estimation pipeline
Real answer collection with informative contrasts
Weaknesses
Simulation-only validation of the full method
Fragile measurement assumptions
Large residual effect error and coverage loss under misspecification
Sign-restricted priors may inflate positive effects
Numerical fragility in some conditions
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
GMMM extends marketing mix modeling (MMM) to two channels whose media inputs are not logged: Generative Engine Optimization (GEO, organic source modifications) and Generative Engine Marketing (GEM, sponsored placements in generated answers). The core construction is Eq. (6): the GEO input is the expected number of noticed occurrences, Σ N_iqt·ω_ipt·π_qpt(z)·λ_qpt, combining question counts, per-system usage shares, hierarchical-logistic occurrence probabilities, and binomial notice probabilities. The GEM input (Eq. 13) maps spending through a Poisson delivery model times a notice probability. Both pass through adstock and Hill saturation into response equation (3), which includes a direct source-state term β_D·Z_it and a GEO×GEM interaction β_GP. Estimands are complete-sequence contrasts, ∆G = V(1,1) − V(0,1). The paper claims identification via a longitudinal g-formula (Theorem 4.4), a rank condition (Proposition 4.2) under which the total effect is identified even when β_D and β_G are not separately identified, a scale-equivalence result (Proposition 4.3: a common scale is absorbed by rescaling the Hill midpoint to θ/c0), and simulation evidence seeded by 2,240 real collected answers.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 6 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is Generative Marketing Mix Modeling (GMMM)?
GMMM extends marketing mix modeling (MMM) to generative-AI channels. Because firms cannot log how often users see their name inside AI answers, GMMM constructs the media input as the expected number of noticed occurrences — combining answer-level occurrence probabilities, question volumes, usage shares across AI systems, and user notice probabilities — then applies standard carryover, saturation, and causal treatment-effect machinery on top.
Does the paper prove that GEO increases sales using real data?
No. The full GMMM pipeline is validated only in simulations seeded by 2,240 real generated answers from two GPT models. A separate referral-traffic case study (Appendix F) reports an immediate ratio near 1.8-2.4 but is sensitive to the assumed treatment date, has a mid-range placebo rank, and is explicitly not a GMMM application.
What is the difference between GEO and GEM?
GEO (Generative Engine Optimization) means modifying source material so that generative engines mention or draw on the firm's content organically; GEM (Generative Engine Marketing) means paying for sponsored placements inside generated answers. The paper models the causal effect of both on business outcomes like sales or conversions.
Is code or data available for this paper?
The provided text contains no code-availability statement. The referral-traffic dataset analyzed in Appendix F comes from a public GitHub repository cited in the paper, but availability of the collected answers and simulation code is not stated in the reviewed excerpt.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- The provided text is a sampled excerpt with several [...sampling gap...] regions (parts of Sections 3-5 and Appendices A-C); judgments about those regions rest only on material visible elsewhere in the excerpt.
- No code-availability statement appeared in the provided excerpt; the paper may state one in unobserved sections.
- Main-body Section 5 results are only partially visible in the excerpt; headline numbers cited here come from Section 1 and Appendices B.3-B.6.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.