LLMO Guy

What if you could make subtle changes to content to successfully influence AI search results in your favour?

Research submitted to Arxiv on 25 May 2025 looked into a technique for doing just that.

Here's what you need to know about it.

Key Findings

The researchers showed that, through an automated optimization process, they could tweak content about an item in way that (a) was intentionally hard-to-detect and (b) maintained the fluency of the text, yet (c) caused the item to rank significantly higher in the output of LLM-driven ranking systems.

They named this new technique "StealthRank Prompt" (SRP).

This built on previous research that had shown that more blatant, easier-to-detect, text adjustments could be used to influence LLM-driven ranking.

Although the StealthRank Prompt optimizations were harder to detect than those of other approaches, they did still result in weird text in some cases.

Potential Weaknesses of the Study

The study looked at the effects of text optimization on simple LLM-driven ranking systems powered by various open source LLMs. It did not look at whether such techniques would be equally effective against the most-used LLM-driven ranking systems such as the ones increasingly playing a part in Google searches.

Practical Takeaways

Sophisticated marketers with sufficient technical resources could look at using the techniques described in this paper (and/or the accompanying source code) to optimize their content and, potentially, get improved visibility as a result.

Any such effort would need to address the likelihood of peculiar text being generated in some cases and, presumably, have a good system for filtering that out.

Authors

The research was conducted by Yiming Tang, Yi Fan, Chenxiao Yu, Tiankai Yang, and Yue Zhao from the University of Southern California, along with Xiyang Hu from Arizona State University.

Paper

Here's a link to the paper: arxiv:2504.05804