Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Online Advertisements with LLMs: Opportunities and Challenges

Soheil Feizi, MohammadTaghi Hajiaghayi, Keivan Rezaei, Suho Shin

TL;DR

This paper addresses the viability of online advertising within large language model (LLM) systems by proposing a unified LLMA framework that couples four modules—modification, bidding, prediction, and auction—to embed advertisements into generated outputs. It offers design choices for each module, analyzes desiderata such as output quality, latency, and reliability, and situates the framework within recent mechanism-design developments for LLMs. The work highlights key challenges across privacy, latency, and user experience, and relates LLMA approaches to token auctions, RAG-based relevance, and prominence/RLHF-inspired auctions. Overall, the paper lays a foundational research agenda for sustainable LLMA-based advertising, including dynamic creative optimization and autobidding, with practical implications for revenue and user satisfaction in AI-assisted information systems.

Abstract

This paper explores the potential for leveraging Large Language Models (LLM) in the realm of online advertising systems. We introduce a general framework for LLM advertisement, consisting of modification, bidding, prediction, and auction modules. Different design considerations for each module are presented. These design choices are evaluated and discussed based on essential desiderata required to maintain a sustainable system. Further fundamental questions regarding practicality, efficiency, and implementation challenges are raised for future research. Finally, we exposit how recent approaches on mechanism design for LLM can be framed in our unified perspective.

Online Advertisements with LLMs: Opportunities and Challenges

TL;DR

This paper addresses the viability of online advertising within large language model (LLM) systems by proposing a unified LLMA framework that couples four modules—modification, bidding, prediction, and auction—to embed advertisements into generated outputs. It offers design choices for each module, analyzes desiderata such as output quality, latency, and reliability, and situates the framework within recent mechanism-design developments for LLMs. The work highlights key challenges across privacy, latency, and user experience, and relates LLMA approaches to token auctions, RAG-based relevance, and prominence/RLHF-inspired auctions. Overall, the paper lays a foundational research agenda for sustainable LLMA-based advertising, including dynamic creative optimization and autobidding, with practical implications for revenue and user satisfaction in AI-assisted information systems.

Abstract

This paper explores the potential for leveraging Large Language Models (LLM) in the realm of online advertising systems. We introduce a general framework for LLM advertisement, consisting of modification, bidding, prediction, and auction modules. Different design considerations for each module are presented. These design choices are evaluated and discussed based on essential desiderata required to maintain a sustainable system. Further fundamental questions regarding practicality, efficiency, and implementation challenges are raised for future research. Finally, we exposit how recent approaches on mechanism design for LLM can be framed in our unified perspective.
Paper Structure (50 sections, 1 equation, 5 figures)

This paper contains 50 sections, 1 equation, 5 figures.

Table of Contents

  1. Introduction
  2. Search advertising
  3. Motivating example
  4. SA versus LLMA
  5. Our contribution
  6. Related works
  7. Large Language Models
  8. Online Advertisement
  9. Mechanism design with LLM agents
  10. Framework for LLMA
  11. Modification module
  12. Comparison to SA
  13. Bidding module
  14. Comparison to SA
  15. Prediction Module
  16. ...and 35 more sections

Figures (5)

  • Figure 1: An example of providing unstructured advertisement in the LLM. Left column refers to the case where we ask above question from ChatGPT 4, and to incorporate the ads we use queries specified in the right column.
  • Figure 2: Overall framework of LLMA.
  • Figure 3: An example of providing structured advertisement in the LLM. Left column refers to the case where we ask above question from ChatGPT 3.5, and to incorporate the ads we use queries specified in the right column.
  • Figure 4: An example of ChatGPT 4 providing different output tailored to the user context. We keep the context provided in Figure \ref{['fig:intro']}, but asks again to advertise again to the specified user segment.
  • Figure 5: Integrating responsive advertisement in the images generated by ChatGPT 4. In the left panel, a user requests car recommendations within the price range of $40,000 to $60,000, and the chat bot, when capable of advertising, generates a "BMW 3 Series" image (left). Personalized ads modify this image for a 60-year-old woman (middle) or man (right) based on user context like age and gender. Note that it includes an image of woman for man in the output, and vice versa. In the right panel, a user seeks machine learning conferences in Florida. The chat bot advertises "SunShine Travel" with its logo on a generic Florida image (left). With personalized ads, the model tailors promotions, offering flight tickets from the user's location and showing more relevant Florida images. For example, for a young unmarried person from DC, the output advertises DC to Florida flights (middle) and features an image of beach there, while for a married person in California, it promotes California to Florida flights with images of amusement parks. This enhances user experience with targeted and appealing content.