Proactive Hybrid Advertising (PHA): First Experimental Report — A Core-Mechanism Feasibility Evaluation of AI-Initiated Organic–Sponsored Content Integration in Generative AI

Proactive Hybrid Advertising (PHA) is a conceptual advertising framework for generative artificial intelligence (AI) in which an AI system proactively initiates an advertising opportunity and integrates independently generated or selected organic content with advertiser-sponsored content within the same AI-generated output. Its defining characteristics are proactive AI initiation, hybrid organic–sponsored content integration, and presentation-format independence. Although the conceptual framework establishes a general architecture for this form of generative AI advertising, its operational feasibility requires empirical investigation. This paper presents the first experimental report on PHA, focusing on the preliminary computational evaluation of its core mechanisms. A controlled, scenario-based experiment was conducted using one generative model, three AI-based evaluators, six conversational scenarios, and five experimental conditions. The scenarios represented high-commercial-relevance everyday contexts, a low-commercial-relevance informational context, an explicit advertising request, a sensitive personal-support context, and an explicit no-advertising preference. The experimental system generated proactive advertising decisions using predefined thresholds for commercial relevance, incremental user value, intrusiveness risk, and confidence, together with contextual restrictions. Across six scenarios, the system initiated proactive advertising in three cases. Both scenarios predefined as eligible for proactive advertising triggered successfully, producing an eligible-scenario trigger rate of 100%. However, one of four predefined noneligible scenarios also triggered, resulting in a noneligible-scenario trigger rate of 25%. The exploratory decision metrics were 83.33% accuracy, 66.67% precision, 100% recall, 75.00% specificity, and an F1-score of 0.80, although these values are descriptive and based on researcher-assigned scenario labels rather than independently established ground truth. Five output conditions were evaluated by three AI evaluators, producing 30 condition records and 90 successful evaluation records. In the format-independent PHA condition, all three triggered outputs were recorded as successfully integrated after programmatic repair. However, none achieved native integration before repair, revealing a substantial gap between assisted output construction and the generative model's ability to produce a complete hybrid response independently. Two output formats were observed, but the small number of triggered cases does not establish format independence. The experiment completed 24 successful model calls in 7.29 minutes. Overall, the findings provide preliminary computational evidence that proactive opportunity detection and basic contextual safeguards can be operationalized, while identifying native hybrid integration, false-positive control, and cross-format robustness as unresolved technical challenges. This study constitutes an initial feasibility investigation rather than definitive validation of PHA's effectiveness, user acceptance, or commercial performance.