AI vs. Human Search

The increasing use of AI agents is changing how customers search and make decisions among options. This raises a fundamental question: how does AI search compare to human search? We investigate a search process in which each option has a quality component and a random preference shock, where quality is linear in observable characteristics. Under human search, a decision maker incurs search costs to evaluate several options and selects the one with the highest realized utility. Under AI search, an AI agent scores numerous options using a linear function of characteristics with a misspecified preference vector and recommends the option with the highest value. We characterize the expected utility and quality of the options selected under AI and human search. Preference shocks can improve utility while reducing quality for human search (the serendipity effect), whereas misspecification error weakly degrades both under AI search. Despite misspecification, AI search can benefit from larger search scales. When the search scale is sufficiently large, AI search outperforms human search, while human search could perform better under significant AI misspecification. Using an environment calibrated with real-world data from JD.com and LLM-based simulations of agents and consumers, we show that LLM-induced preference misspecification generates performance patterns consistent with our theoretical predictions.