Manipulating Large Language Models to Increase Product Visibility

TL;DR

The paper investigates whether large language models (LLMs) used in search and e-commerce can be manipulated to increase product visibility by inserting a strategically crafted text sequence (STS) into a product's information page. It proposes a framework that optimizes the STS via the Greedy Coordinate Gradient (GCG) algorithm within a retrieval-augmented generation (RAG) style setup, and tests robustness to variations in the ordering of retrieved product data using open-source LLMs. Experimental results on a catalog of fictitious coffee machines show that STS can significantly boost a target product's ranking, often elevating it to the top recommendation under both fixed and permuted input orders, with robustness enhanced when optimizing across randomized permutations. The findings highlight potential competitive advantages for vendors and raise important considerations about fairness and safeguards in AI-driven search systems, pointing to ethical guidelines and countermeasures as future work.

Abstract

Large language models (LLMs) are increasingly being integrated into search engines to provide natural language responses tailored to user queries. Customers and end-users are also becoming more dependent on these models for quick and easy purchase decisions. In this work, we investigate whether recommendations from LLMs can be manipulated to enhance a product's visibility. We demonstrate that adding a strategic text sequence (STS) -- a carefully crafted message -- to a product's information page can significantly increase its likelihood of being listed as the LLM's top recommendation. To understand the impact of STS, we use a catalog of fictitious coffee machines and analyze its effect on two target products: one that seldom appears in the LLM's recommendations and another that usually ranks second. We observe that the strategic text sequence significantly enhances the visibility of both products by increasing their chances of appearing as the top recommendation. This ability to manipulate LLM-generated search responses provides vendors with a considerable competitive advantage and has the potential to disrupt fair market competition. Just as search engine optimization (SEO) revolutionized how webpages are customized to rank higher in search engine results, influencing LLM recommendations could profoundly impact content optimization for AI-driven search services. Code for our experiments is available at https://github.com/aounon/llm-rank-optimizer.

Figures (7)

• Figure 1: Bing Copilot's response for the search phrase "coffee machines".
• Figure 2: Rank distribution of the target product before (not recommended) and after (top recommendation) adding the strategic text sequence.
• Figure 3: LLM Search: Given a user query, it extracts relevant product information from the internet and passes it to the LLM along with the query. The LLM uses the retrieved information to generate a response tailored to the user's query. The circled numbers indicate the order of the steps. STS: The strategic text sequence is added to the target product's information page to increase its chances of being recommended to the user.
• Figure 4: The target product ColdBrew Master goes from not being recommended to the top recommended product in 100 iterations of the GCG algorithm. The optimized text significantly increases the chances of the target product being listed as the top recommendation.
• Figure 5: Evaluating the advantage from the STS for the target product ColdBrew Master under variations of the product ordering in the LLM's input prompt. Figure (a) plots the advantage of optimizing with a fixed product order. Figure (b) shows that this advantage can be significantly improved by optimizing with random permutations of the product list.