The Impact of Generative AI and LLMs on Publishers, Advertising, Productivity and Market Equilibrium
Paper Session
Monday, Jan. 5, 2026 10:15 AM - 12:15 PM (EST)
-
Chairs:
H. Tai Lam, University of California-Los Angeles - Sam Goldberg, Stanford University
Open(?) AI
Abstract
I provide an economic paradigm of ``open'' AI models based on four features. Open AI models are (1) free to use; (2) free to inspect; and (3) free to modify; and (4) all derivative applications should also have these features. Next, I study a model of AI development that captures the topology of general-purpose models and niche applications that may be close to or far from other niche applications. A ``founder'' builds a general-purpose foundation model. In the open paradigm, a community of developers each decides whether to develop specialized applications for their own purposes or to use another developer's application. In the closed paradigm, the founder centrally builds the applications and sells access. I compare both to the social planner's optimum in terms of the number of applications, the degree of specialization of these applications, and the investment in the overall quality of the foundation model. Notably, the open paradigm features countervailing distortions---a ``reverse Spence'' distortion and a ``free riding'' distortion---that may lead to too little or too much specialization and too little or too much quality. The closed model yields the optimal quality investment but too many applications that are not specialized enough.Exploration of Embedded Ads in LLM-enabled Knowledge Search as a New Advertising Format
Abstract
Large language models such as ChatGPT have the potential to massively disrupt consumer knowledge search, threatening to replace search rankings and websites by providing unique, highly contextual answers to user questions and raising questions of how advertisers will reach consumers in this new paradigm. This research explores one avenue for monetizing the attention captured by AI knowledge search: embedded advertisements in the AI-generated responses. We present a series of randomized studies assessing its effectiveness, mechanism, and moderating effects of advertisement disclosure and availability of targeting information. Our results show that embedded advertisements outperform traditional display advertising in motivating purchase intention, through a mechanism of increased customer awareness rather than shifting brand sentiment. However, inserted embedded advertising come at the cost of lower user satisfaction with the AI knowledge search experience, particularly when the advertising is disclosed. These findings highlight the persuasive potential of these new technologies and may rationalize why embedded advertising has not been widely adopted during the current period of intense competition between LLM providers.The Impact of LLM Adoption on Online User Behavior
Abstract
The adoption of AI tools, and especially Large Language Models (LLMs), has the potential to significantly transform how users engage with information online, potentially serving as substitutes or complements to existing digital resources. We use detailed clickstream data from 2022 and 2023 to examine users’ online behavior following the adoption of large language models. We document a significant decrease in online search activity, a typical entry point to content consumption. Online searches drop slowly, suggesting a period during which users learn to use LLMs, but eventually adopters’ level of online search in traditional search engines is more than 20% below the pre-adoption period, though there is heterogeneity across types of queries. We then turn to the effect of LLM adoption on website traffic. We document that while frequently visited websites are not affected, smaller websites suffer a significant drop in visits. In line with these results, we then report a significant drop in display ad exposures, especially to consumers with high levels of retail activity, though we do not find a reduction in search ad exposures. Last, we study two distinct categories of websites: education-related websites and user-generated content platforms. We document a significant drop in visits to education-related websites and heterogeneity across user-generated content platforms with a pronounced negative effect on Stack Overflow but no significant effect on Wikipedia, Reddit, and social media. We discuss implications for online content creators, for GenAI firms, and for public policy.Discussant(s)
Rachel Xiao
,
Fordham University
Yihao Yuan
,
University of California-Los Angeles
Daniel Martin
,
University of California-Santa Barbara
Daniel Bjorkegren
,
Columbia University
JEL Classifications
- L1 - Market Structure, Firm Strategy, and Market Performance
- D0 - General