Learning Dynamics of Strategic Publishers in Generative AI Ecosystems

Generative AI (GenAI) search systems are fundamentally changing how information is accessed and how content creators (publishers) are rewarded. Unlike traditional search engines that rank documents, GenAI systems synthesize responses and attribute information to sources. This paper introduces a game-theoretic model to analyze how publishers behave strategically in this environment to maximize their exposure through citations. The authors model the ecosystem as a game where publishers choose content embeddings, and the platform uses a generation function (based on centroids of retrieved documents) and an attribution mechanism to reward them.

The authors evaluate the stability of these ecosystems by studying whether better-response dynamics—where publishers iteratively update their content to improve their utility—converge to a Pure Nash Equilibrium (PNE). They find that the widely used winner-takes-all mechanism, which assigns attribution only to the most relevant document, often fails to reach a stable state and may not even possess a PNE. Similarly, the softmax attribution function, while smoother, does not guarantee convergence and can lead to infinite sequences of strategic deviations.

A key contribution of the paper is the identification of a specific family of linear attribution functions that guarantee convergence to a stable equilibrium. Through extensive simulations, the authors show that there is a significant trade-off between ecosystem stability, publisher welfare, and user welfare. Notably, they find that stable mechanisms do not necessarily maximize social welfare, suggesting that platform designers must carefully balance these competing objectives when selecting an attribution strategy.