Complementor multihoming is often viewed as weakening platforms by reducing differentiation and increasing consumer substitution. We revisit this perspective in the context of content creator multihoming. Using a unique seven-year panel of thousands of creators on Chinese TikTok and RedNote, we first document that multihoming is widespread and strategic, and that creators strategically cross-post selected content. Exploiting staggered platform adoption in a difference-in-differences design, we find that, contrary to traditional concerns that multihoming diverts creators and users away from the focal platform, adopting a second platform raises creators' focal-platform followers by 33\% and their engagement per post by 26\%. Creators also increase rather than divert content production on the focal platform. Mechanism analyses that exploit content appearing only on the new platform show that attention flows back to the focal platform. Engagement gains are much larger for cross-posted posts than for other posts, and video quality does not change. Gains peak when creators cross-post about half of their content, more than the average adopter does, and are larger for smaller creators and for those whose content the new platform lacks. By highlighting this cross-platform attention spillover channel, our paper helps clarify when complementor multihoming can benefit both complementors and focal platforms, with implications for creator strategy and platform governance.

The rapid advancement of generative AI has enabled the automated generation of answers, introducing new dynamics in user behavior related to question-asking on digital platforms. However, the impact of users’ experience with generative AI for answer generation on their questioning behavior remains unclear. On the one hand, the availability of generative AI tools may lead users to rely on these tools for formulating questions, potentially resulting in fewer but higher-quality inquiries. On the other hand, using generative AI may motivate users to ask more questions to enhance their knowledge and build social reputation within the community. We investigate how the use of generative AI, specifically ChatGPT, to generate answers affects users’ questioning behavior on Stack Overflow, one of the largest Q&A platforms. Leveraging a natural experiment involving the introduction and subsequent ban of ChatGPT on the platform, we apply a difference-in-differences (DID) estimation approach to assess the effects of generative AI usage on the quantity and quality of user-generated questions. We measure these outcomes through metrics such as the number of questions posted per week, question length, novelty, and the number of received upvotes. Our findings reveal that the use of generative AI to generate answers is associated with an increase in the number of questions, which feature longer, more novel content and receive more upvotes. We examine two mechanisms: complexity shift, where ChatGPT handles routine queries while more advanced needs remain for the community, and skill enhancement, where users learn from ChatGPT to ask clearer questions. We also test two moderators: answer volume (proxy for user experience) and platform reputation score (proxy for user reputation). Results show weaker effects among experienced answerers but stronger effects among high reputation users. These insights highlight how generative AI shapes engagement and can guide platform strategies for AI integration.

As Amazon evolves from a traditional retailer into a multi-sided platform, it has launched a wholesale program that enables third-party sellers to sell their products directly to the company, making Amazon both their distribution partner and their competitor on the same listings. This study examines how participation in Amazon’s wholesale program affects third-party sellers’ sales volume and revenue. Using proprietary data from an established outdoor recreation seller whose products entered the program at staggered times, we employ a difference-in-differences design and find that participation is associated with a significant reduction in both sales volume and revenue. We propose that this adverse effect stems from an anchoring mechanism: Amazon’s wholesale listing becomes the reference point against which consumers evaluate the seller’s own offer, systematically diminishing its perceived value. The findings contribute to the literature on platform co-opetition and caution sellers to weigh the visibility benefits of wholesaling against its competitive costs, with implications for managers and policymakers concerned with fair competition in digital marketplaces.

Low completion and engagement are persistent challenges in online education. Partnering with a large Indian platform that offers live information technology courses, this study examines short introductory sessions that give prospective students information about a course before they commit. Such sessions improve the match between students and courses, but scheduling them before the main course lengthens the wait time that committed students face, which can dampen engagement. We develop a multiperiod analytical model in which heterogeneous students decide whether to enroll, whether to drop out, and how much effort to invest, while the platform chooses prices, information provision, and teaching support. The model characterizes how these sessions reshape enrollment, completion, engagement, and total skill production. We calibrate the model with rich enrollment and engagement data from the platform and exploit the platform’s removal of these sessions in January 2025 as a natural experiment to test the model’s predictions. The study offers guidance on when experiential screening justifies its operational cost.

Major online platforms such as Amazon, Google Shopping, and Booking.com increasingly rely on large language models (LLMs) to condense large volumes of product reviews into short summaries that, for most shoppers making quick decisions, substitute for the reviews themselves. We ask whether these summaries faithfully convey the collective sentiment of prior buyers. Constructing review corpora with known sentiment distributions and scoring both the reviews and the LLM-generated summaries with a BERT-based rating model and lexicon-based sentiment analysis, we document two systematic distortions. Centralization pulls summary ratings toward the midpoint of the scale regardless of the underlying review distribution, obscuring the distinctive signal of polarized products and suppressing strong praise more substantially than strong criticism. Sentiment dampening attenuates the lexical strength of sentiment words even when the balance of positive and negative content is preserved, compressing the emotional extremes of the review distribution. A pre-registered survey experiment on simulated product pages links these distortions to consumer behavior: AI-generated summaries alter purchase intentions across the entire quality spectrum. As platforms increasingly interpose LLMs between user-generated content and consumers, the findings highlight faithfulness as a central design criterion for AI-mediated information.

Modern digital infrastructure relies heavily on open source software, where a single vulnerable repository can expose every project that depends on it. This project examines how public disclosure of vulnerabilities changes development activity, both in the affected repositories and in their downstream dependents. We draw on the GitHub Advisory Database, dependency graphs, and commit histories covering around 36,000 repositories from 2019-2026, and leverage the staggered timing of disclosures in a difference-in-differences framework. The project sheds light on who bears the maintenance burden after disclosure, with implications for coordinated disclosure policy.