Working Paper
AI Adoption and Consumer Demand [Job Market Paper]
with Se Yan, Zachary Zhong, Wenyu Zhou, and N. Mehta
Abstract. We study how the adoption of a platform-embedded AI assistant affects consumer search and demand. We run a large field experiment involving 15.9 million users on Ctrip, a major Chinese online travel platform. The treatment makes the chat icon more salient, increasing AI chat requests by 12.66% and creating experimental variation in AI chat usage. Using the treatment-induced increase in chat requests to estimate the AI-usage channel, we find that AI chat usage increases all activities along the purchase funnel—total search queries, browsing, clicks, and hotel orders—while leaving click-through and conversion rates largely unchanged. We identify geographic demand activation as one mechanism: chat helps users translate broad and vague travel intent into specific locations and hotel options, expanding search toward distant provinces and niche scenic or leisure-oriented destinations. Our results show that shopping AI assistants can create value by activating latent demand at the top of the purchase funnel, even when click-through and conversion rates remain unchanged.
with Se Yan, Zachary Zhong, Wenyu Zhou, and N. Mehta
Abstract. We study how the adoption of a platform-embedded AI assistant affects consumer search and demand. We run a large field experiment involving 15.9 million users on Ctrip, a major Chinese online travel platform. The treatment makes the chat icon more salient, increasing AI chat requests by 12.66% and creating experimental variation in AI chat usage. Using the treatment-induced increase in chat requests to estimate the AI-usage channel, we find that AI chat usage increases all activities along the purchase funnel—total search queries, browsing, clicks, and hotel orders—while leaving click-through and conversion rates largely unchanged. We identify geographic demand activation as one mechanism: chat helps users translate broad and vague travel intent into specific locations and hotel options, expanding search toward distant provinces and niche scenic or leisure-oriented destinations. Our results show that shopping AI assistants can create value by activating latent demand at the top of the purchase funnel, even when click-through and conversion rates remain unchanged.
Reasoning AI, Consumer Search, and Purchase: Evidence from an Online Platform Field Experiment [SSRN] with Se Yan, Zachary Zhong, Wenyu Zhou, and N. Mehta
Major Revision at Marketing Science
Abstract. We study how more deliberative conversational assistance reshapes consumer search and purchase behavior on digital platforms. We examine this question in the context of reasoning AI, using a large-scale field experiment on a major online travel platform, where more than 510,000 users were randomly assigned to an AI assistant powered by either a reasoning model (DeepSeek-R1) or a comparable non-reasoning model (DeepSeek-V3). Access to the reasoning assistant reduced hotel bookings by 2.5%. The effect operates through the search funnel: users initiated fewer search queries, browsed fewer hotel options, and clicked fewer hotels for deeper inspection. Additional evidence suggests that this narrowing is more consistent with the assistant’s more informative replies reducing consumers’ perceived need for further search than with a time-costbased explanation. At the same time, the reasoning assistant increased subsequent engagement with the chat feature. Our findings show that more deliberative conversational assistance can narrow consumer search, creating a trade-off between assistant engagement and search-driven purchases.
Major Revision at Marketing Science
Abstract. We study how more deliberative conversational assistance reshapes consumer search and purchase behavior on digital platforms. We examine this question in the context of reasoning AI, using a large-scale field experiment on a major online travel platform, where more than 510,000 users were randomly assigned to an AI assistant powered by either a reasoning model (DeepSeek-R1) or a comparable non-reasoning model (DeepSeek-V3). Access to the reasoning assistant reduced hotel bookings by 2.5%. The effect operates through the search funnel: users initiated fewer search queries, browsed fewer hotel options, and clicked fewer hotels for deeper inspection. Additional evidence suggests that this narrowing is more consistent with the assistant’s more informative replies reducing consumers’ perceived need for further search than with a time-costbased explanation. At the same time, the reasoning assistant increased subsequent engagement with the chat feature. Our findings show that more deliberative conversational assistance can narrow consumer search, creating a trade-off between assistant engagement and search-driven purchases.
Shopping with a Platform AI Assistant: Who Adopts, When in the Journey, What For [SSRN]
with with Se Yan, Zachary Zhong, and Wenyu Zhou
Major Revision at Marketing Science
Abstract. This paper provides some of the first large-scale descriptive evidence on how consumers adopt and use platform-embedded shopping AI in e-commerce. Using data on 31 million users of Ctrip, China's largest online travel platform, we study "Wendao," an LLM-based AI assistant integrated into the platform. We document three empirical regularities. First, adoption is highest among older consumers, female users, and highly engaged existing users, reversing the younger, male-dominated profile commonly documented for general-purpose AI tools. Second, AI chat appears in the same broad phase of the purchase journey as traditional search and well before order placement; among journeys containing both chat and search, the most common pattern is interleaving, with users moving back and forth between the two modalities. Third, consumers disproportionately use the assistant for exploratory, hard-to-keyword tasks: attraction queries account for 42% of observed chat requests, and chat intent varies systematically with both the timing of chat relative to search and the category of products later purchased within the same journey. These findings suggest that embedded shopping AI functions less as a substitute for conventional search than as a complementary interface for exploratory product discovery in e-commerce.
with with Se Yan, Zachary Zhong, and Wenyu Zhou
Major Revision at Marketing Science
Abstract. This paper provides some of the first large-scale descriptive evidence on how consumers adopt and use platform-embedded shopping AI in e-commerce. Using data on 31 million users of Ctrip, China's largest online travel platform, we study "Wendao," an LLM-based AI assistant integrated into the platform. We document three empirical regularities. First, adoption is highest among older consumers, female users, and highly engaged existing users, reversing the younger, male-dominated profile commonly documented for general-purpose AI tools. Second, AI chat appears in the same broad phase of the purchase journey as traditional search and well before order placement; among journeys containing both chat and search, the most common pattern is interleaving, with users moving back and forth between the two modalities. Third, consumers disproportionately use the assistant for exploratory, hard-to-keyword tasks: attraction queries account for 42% of observed chat requests, and chat intent varies systematically with both the timing of chat relative to search and the category of products later purchased within the same journey. These findings suggest that embedded shopping AI functions less as a substitute for conventional search than as a complementary interface for exploratory product discovery in e-commerce.
Gender Inequality and Household Purchase Decisions: The Case of Automobiles in China [SSRN]
with Zachary Zhong, Nan Chen
Working Paper
Abstract. How do major social shifts like rising gender equality alter high-stakes household consumption? As women’s economic power grows, their influence over joint purchases like automobiles increases, yet causal evidence on how this shapes market outcomes remains scarce. This paper investigates how a fundamental dimension of gender equality—the education gap—affects household car purchases, a major financial decision in the Chinese auto market. Using administrative data on 15 million new vehicle registrations, we first document that female ownership is significantly higher in more gender-equal regions. To establish causality, we leverage the staggered rollout of China’s 1986 Compulsory Schooling Law as a natural experiment, which exogenously reduced the gender education gap with varying intensity across regions and birth cohorts. Our instrumental variable analysis provides novel causal estimates: closing the gender education gap by one year increases the probability of a car being registered to a woman by a substantial 9.6 percentage points. This effect is strongest for higher-priced, foreign-brand sedans, suggesting a genuine shift in control over the primary family vehicle, not just secondary ones. Corroborating this shift in influence, we also find that a smaller education gap increases the market share of female-preferred vehicle attributes (e.g., color), even for cars ultimately registered to men. Our findings offer key insights for firms, demonstrating that rising gender equality is not just a social trend but a fundamental driver of market evolution that reshapes demand for high-value goods.
with Zachary Zhong, Nan Chen
Working Paper
Abstract. How do major social shifts like rising gender equality alter high-stakes household consumption? As women’s economic power grows, their influence over joint purchases like automobiles increases, yet causal evidence on how this shapes market outcomes remains scarce. This paper investigates how a fundamental dimension of gender equality—the education gap—affects household car purchases, a major financial decision in the Chinese auto market. Using administrative data on 15 million new vehicle registrations, we first document that female ownership is significantly higher in more gender-equal regions. To establish causality, we leverage the staggered rollout of China’s 1986 Compulsory Schooling Law as a natural experiment, which exogenously reduced the gender education gap with varying intensity across regions and birth cohorts. Our instrumental variable analysis provides novel causal estimates: closing the gender education gap by one year increases the probability of a car being registered to a woman by a substantial 9.6 percentage points. This effect is strongest for higher-priced, foreign-brand sedans, suggesting a genuine shift in control over the primary family vehicle, not just secondary ones. Corroborating this shift in influence, we also find that a smaller education gap increases the market share of female-preferred vehicle attributes (e.g., color), even for cars ultimately registered to men. Our findings offer key insights for firms, demonstrating that rising gender equality is not just a social trend but a fundamental driver of market evolution that reshapes demand for high-value goods.
Publications
Comparing Human-Only, AI-Assisted, and AI-Led Teams on Assessing Research Reproducibility in Quantitative Social Science social science crowd analysis
Forthcoming, Proceedings of the National Academy of Sciences (PNAS)
Abstract. Large Language Models (LLMs) such as ChatGPT are transforming how scientists conduct and validate research. LLMs are thus seen as promising tools to improve scientific reproducibility. We experimentally test how collaboration between researchers and LLM assistants influences the reproduction of quantitative social science findings. Study I (2024) assigned 288 researchers to 103 teams working in three groups: human-only, AI-assisted, and AI-led. In the AI-led group, the LLM conducted reproducibility checks with minimal human oversight. Study II (2025) replicated the design with 95 researchers in 34 teams. Human-only and AI-assisted teams reproduced published results at comparable rates, and both outperformed AI-led teams. Human-only teams also identified more major errors than AI-assisted and AI-led teams. Finally, both human-only and AI-assisted teams outperformed AI-led approaches in both proposing and implementing robustness checks. In an exploratory analysis, we observe that the gap in most outcomes between AI-led and the other two groups began to narrow by the final event of 2024 and was further reduced in 2025. Despite rapid model advances, expert human judgment currently remains indispensable for reliable empirical verification.
Forthcoming, Proceedings of the National Academy of Sciences (PNAS)
Abstract. Large Language Models (LLMs) such as ChatGPT are transforming how scientists conduct and validate research. LLMs are thus seen as promising tools to improve scientific reproducibility. We experimentally test how collaboration between researchers and LLM assistants influences the reproduction of quantitative social science findings. Study I (2024) assigned 288 researchers to 103 teams working in three groups: human-only, AI-assisted, and AI-led. In the AI-led group, the LLM conducted reproducibility checks with minimal human oversight. Study II (2025) replicated the design with 95 researchers in 34 teams. Human-only and AI-assisted teams reproduced published results at comparable rates, and both outperformed AI-led teams. Human-only teams also identified more major errors than AI-assisted and AI-led teams. Finally, both human-only and AI-assisted teams outperformed AI-led approaches in both proposing and implementing robustness checks. In an exploratory analysis, we observe that the gap in most outcomes between AI-led and the other two groups began to narrow by the final event of 2024 and was further reduced in 2025. Despite rapid model advances, expert human judgment currently remains indispensable for reliable empirical verification.
Selected Work in progress
Business Responses to Supply-Chain Shock: Evidence from the 2018-19 African Swine Fever Outbreaks in China with Z. Zhong
Gender Difference in Consumer Boycott with Z. Zhong
