Haohan Wang › Research › Visibility in LLM-Based Search
Visibility in LLM-Based Search
As people shift from search results pages to AI assistants that recommend a few options directly, who gets recommended becomes a question of fairness and trust. Haohan Wang's group at the University of Illinois Urbana-Champaign studies how LLM-based search systems (generative engines) rank and recommend content, how those rankings can be influenced, and what that means for small businesses, independent creators, and the reliability of AI recommendations. This work is part of the group's broader research on trustworthy AI.
Key questions
What is a generative engine?
A generative engine, or LLM-based search system, answers a query by synthesizing a direct answer or a short ranked set of recommendations with a large language model, instead of returning a list of links for the user to explore.
Why does ranking in LLM-based search matter?
Generative engines' recommendations are strongly influenced by the initial retrieval order, which can disadvantage small businesses and independent creators by limiting their visibility. A product can be retrieved by the search engine yet still not appear prominently in the LLM's final recommendation.
Can output rankings be changed without access to the model?
Yes. The group's CORE method treats the LLM and its search interactions as a black box and optimizes only the content the search engine returns. Across GPT-4o, Gemini-2.5, Claude-4, and Grok-3, it promoted target products into the top 5 recommendations 91.4% of the time on average.
Selected projects
Controlling Output Rankings in Generative Engines for LLM-based Search (CORE)
Jin H, Chen R, Zhang P, Luo Y, Zeng H, Luo M, Wang H · arXiv preprint, 2026
CORE appends strategically designed optimization content, of three types (string-based, reasoning-based, and review-based), to the content a search engine retrieves, in order to steer how the LLM ranks items. The paper also introduces ProductBench, a benchmark of 15 product categories with 200 products each, paired with top-10 recommendations from Amazon's search interface.
- Average Promotion Success Rate of 91.4% at top-5, 86.6% at top-3, and 80.3% at top-1 across 15 product categories
- Evaluated on four LLMs with search: GPT-4o, Gemini-2.5, Claude-4, and Grok-3
- Outperforms existing ranking-manipulation methods while preserving the fluency of the optimized content
[Paper] [Project page] [Explainer]
Further reading from the DREAM Lab
Invite a talk or collaborate
Haohan Wang gives talks on how LLM-based search ranks and recommends content and on the trustworthiness of AI recommendations, and welcomes collaborations on measuring and improving fairness and reliability in generative engines. Contact Haohan Wang at haohanw at illinois.edu.
Other research areas: Agentic AI for Genomics · LLM Security · Gene Regulation & Disease Genetics · All publications