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

Analysis

Why This Paper Matters

Generative AI is reshaping how consumers discover and engage with brands. Unlike traditional search, where clicks and impressions are trackable, generative engines produce answers that may mention a firm's name without any direct click or measurable impression. This creates a blind spot for marketers: standard marketing data do not record how often users see and notice a firm's name in generated answers. As a result, firms lack a rigorous way to measure the return on investments in Generative Engine Optimization (GEO) and Generative Engine Marketing (GEM).

This paper introduces Generative Marketing Mix Modeling (GMMM), a causal inference framework designed to fill that gap. By explicitly modeling notice probabilities and combining them with generated answers, question counts, and system usage shares, GMMM aims to estimate the causal effects of GEO and GEM on business outcomes. The work is timely because generative AI platforms are becoming major gateways for product discovery, and without such measurement, marketing budgets may be misallocated.

The paper also contributes to the broader methodological challenge of causal inference in settings where treatments are not directly observed. By establishing sufficient conditions for identification, it provides a foundation for future empirical work in generative marketing.

Technical Contributions

• GMMM framework: A causal inference approach that compares expected business responses under alternative treatment sequences, tailored to generative marketing contexts.
• GEO effect estimation: Combines repeated generated answers with question counts, shares of use across generative systems, and notice probabilities to estimate the causal impact of GEO.
• GEM effect estimation: Integrates records of sponsored placements with notice probabilities to estimate the causal impact of GEM.
• Identification conditions: Provides sufficient conditions under which the causal effects of GEO and GEM can be identified from observational data.
• Simulation study: Evaluates the method using simulated answers to product recommendation in English and Japanese, demonstrating feasibility.

Results

The abstract does not report concrete metrics or comparisons. The authors state that they investigate the empirical performance of GMMM using simulated answers to product recommendation in English and Japanese. No quantitative results, such as bias reduction or coverage rates, are provided in the abstract. Therefore, the practical performance of the method remains to be seen in the full paper.

Significance

GMMM addresses a critical measurement gap in the emerging field of generative marketing. As generative engines become more prevalent, the ability to causally attribute business outcomes to GEO and GEM will be essential for firms to optimize their strategies. The framework's emphasis on notice probabilities acknowledges that mere presence in generated answers is not enough; users must notice the brand. This nuanced view could lead to more accurate marketing mix models.

Moreover, the paper's methodological contributions—particularly the identification conditions—may inspire further research on causal inference with unobserved exposures. While the current validation is limited to simulations, the approach could be extended to real-world data as generative platforms begin to share more detailed logs. Overall, GMMM represents a step toward rigorous, data-driven decision-making in the age of generative AI.