Preference-Based Advertising: A Dynamic Model Preference-Driven Framework for Advertisement Matching in Generative AI Systems

The emergence of generative artificial intelligence (AI) introduces a fundamentally different environment for digital advertising. Unlike conventional advertising systems, which primarily rely on user profiles, historical interactions, demographic characteristics, and contextual signals, generative AI systems exhibit identifiable behavioral preferences when interpreting tasks, evaluating alternatives, selecting tools, and constructing solutions. These model-side preferences may provide an additional source of information for advertising decisions that is distinct from conventional user preference modeling. Nevertheless, the systematic use of dynamically emerging model-side preferences as an independent signal for advertisement matching has not yet been sufficiently formalized. This paper proposes Preference-Based Advertising (PBA), a new advertising decision framework designed to utilize dynamically emerging preferences of generative AI models for advertisement matching, selection, ranking, and adaptation during human–AI interactions. To the best of the author's knowledge, PBA is the first proposed framework to explicitly establish dynamically emerging model-side preferences as an independent advertising decision signal and integrate them with user preferences and conversational context within a unified advertising architecture. The framework introduces five interconnected components: User Interaction Input, Dynamic Model Preference Extraction, Preference Representation, Preference–Advertisement Matching, and Preference-Adaptive Advertisement Selection. A Dynamic Model Preference Vector is developed to represent the model's observable tendencies toward particular tools, solution categories, technical approaches, and interaction characteristics. The framework further introduces a joint matching function that considers model-side preference alignment, user preference alignment, contextual relevance, and advertisement intrusion cost. Unlike approaches that treat AI-generated advertisements as merely personalized content, PBA positions the model's task-dependent behavioral preferences as an explicit component of the advertising decision process. Five experimental conditions are proposed to examine the independent and combined contributions of model-side and user-side preferences. Particular attention is given to incremental matching value, preference stability, model–user preference disagreement, and the possibility of advertising-induced preference distortion. The proposed framework establishes a conceptual and methodological foundation for investigating a new direction in generative AI monetization while maintaining transparency, user autonomy, and separation between commercial incentives and model-generated judgments.