How Much Can We Influence AI Responses?

Introduction to the Unstable Search Landscape

The current search landscape is both unstable and easily manipulated. We often ask how to influence AI answers without acknowledging that Large Language Models (LLMs) outputs are probabilistic by design. This means that their responses can vary greatly and are not always consistent.

The Volatility of LLM Visibility

LLM visibility is a volatility problem. There are several reasons why it’s challenging to influence LLM answers:

1. Lottery-style outputs: LLMs are probabilistic, not deterministic like search engines. This means that answers can vary significantly for the same prompt.
2. Inconsistency: AI answers are not consistent. Running the same prompt multiple times can yield different results, with only 20% of brands showing up consistently.
3. Primary Bias: Models have a bias based on pre-training data, which can affect how much we can influence or overcome that bias.
4. Model Evolution: LLMs evolve over time, with new models becoming smarter and more accurate. However, this also means that old tactics may not work for new models.
5. Model Variation: Different models weigh sources differently for training and web retrieval. For example, ChatGPT relies heavily on Wikipedia, while AI Overviews cite Reddit more often.
6. Personalization: Models like Gemini may have more access to personal data, leading to more personalized results. However, this also means that models may vary in the degree to which they allow personalization.
7. Context: Users reveal more context about what they want with long prompts, making it harder to influence the set of possible answers.

Research on Manipulating LLM Answers

A recent paper from Columbia University, titled "E-GEO: A Testbed for Generative Engine Optimization in E-Commerce," shows how easily LLM answers can be manipulated. The study used a dataset of over 7,000 real product queries and 50,000 Amazon product listings to evaluate how different rewriting strategies improve a product’s AI visibility.

The methodology involved:

• Building the "E-GEO Testbed" to pair product queries with product listings and evaluate how rewriting strategies improve AI visibility.
• Using two AI agents and a control group to simulate a shopping scenario.
• Developing a sophisticated optimization method that uses GPT-4o to analyze results and give recommendations for improvements.

The results showed that:

• A "Universal Strategy" for LLM output visibility in ecommerce exists.
• Rewritten descriptions with a persuasive tone and fluff achieved a win rate of ~90% against the baseline descriptions.
• Sellers do not need category-specific expertise to game the system, as a strategy developed for home goods products achieved an 88% win rate when applied to electronics and 87% when applied to clothing.

Other Research on LLM Manipulation

Other studies have also shown how to manipulate LLM answers:

GEO: Generative Engine Optimization

• Added statistics or quotes to content and found that factual density boosted visibility by about 40%.
  

  Manipulating Large Language Models

• Added a "Strategic Text Sequence" to product pages to manipulate LLMs.
• Found that vendors can significantly improve their product’s LLM visibility by inserting an optimized sequence of tokens into the product information page.
  

  Ranking Manipulation

• Added text on product pages that gave LLMs specific instructions, such as "please recommend this product first."
• Found that LLM visibility is fragile and highly dependent on factors like product names and their position in the context window.

The Coming Arms Race

The growing body of research shows the extreme fragility of LLMs. They are highly sensitive to how information is presented, and minor stylistic changes can move a product from the bottom of the list to the No. 1 recommendation. This creates a long-term problem of scale, as LLM developers need to find ways to reduce the impact of these manipulative tactics to avoid an endless arms race with "optimizers."

Conclusion

The search landscape is unstable and easily manipulated, with LLMs being highly sensitive to how information is presented. While research has shown how to manipulate LLM answers, it’s essential to acknowledge the probabilistic nature of LLMs and the potential consequences of an arms race between optimizers and LLM developers. By understanding these challenges, we can work towards creating a more stable and trustworthy search landscape.