Abstract

Retrieval-augmented generation (RAG) is transforming search from ranking links into producing answers, making quality a per-query production decision. We develop a game-theoretic model to analyze how platforms allocate scarce compute between retrieval breadth and reasoning depth in this new paradigm, uncovering three counter-intuitive operational dynamics with profound policy implications. First, we identify an efficiency rebound: contrary to the “Green AI” narrative, improvements in reasoning efficiency incentivize a massive expansion of retrieval breadth, driving total infrastructure consumption up rather than down. Second, we show that hygiene is antitrust: information pollution functions as a verification barrier that disproportionately binds entrants, meaning that governance to clean the web is an effective way to foster market contestability. Finally, we demonstrate that the social value of competition depends on compute externalities: while competition typically benefits consumers, high environmental damages turn the race for quality into a socially wasteful redundancy of energy-intensive verification pipelines. The results highlight that information hygiene and environmental standards are necessary complements to antitrust in the governance of generative search.

Author: Xing Hu