Overview
AutoGEO automatically discovers actionable rules to improve content visibility in generative search engines. Unlike traditional methods that rely on expert-designed prompts and trial-and-error, AutoGEO uses large language models to automatically extract preference rules of generative search engines, enabling better adaptability across domains and engines. This paper has been accepted by ICLR 2026.
Key Contributions
🎯 Automated Rule Discovery
Automatically extracts actionable rules from generative engines, eliminating manual prompt engineering.
⚡ Dual Deployment Options
AutoGEOAPI for plug-and-play integration; AutoGEOMini for cost-efficient fine-tuned models.
📈 Significant Improvements
Up to 50.99% improvement over baselines while maintaining generative engine utility.
🔄 Cross-Engine Transferability
Rules transfer effectively across different LLMs (Gemini, GPT, Claude) and domains.
Introduction
(a) Generative Engine Workflow: The engine retrieves relevant information and uses an LLM to generate the final answer.
(b) GEO Model Workflow: The GEO model rewrites relevant information to improve its visibility in the final output.
(c) Existing GEO Models: Traditional methods rely on expert-designed prompts and trial-and-error, with limited visibility improvements and adaptability.
(d) AutoGEO (Ours): AutoGEO automatically discovers actionable rules to flexibly build efficient GEO models. These rules can be applied directly in a prompt-based approach or used to fine-tune a compact model, enabling better adaptability across domains and engines.
Method
Left Panel - Automated Preference Rule Discovery: Takes conversational interactions from the Generative Engine as input. An LLM analyzes these interactions to identify patterns and preferences that correlate with higher visibility. Outputs a structured Rule Set containing explicit, actionable guidelines for rewriting content.
Right Panel - Rule Guided GEO Models:
- AutoGEOAPI (Plug-and-play): The Rule Set is directly fed as an Instruction into a black-box LLM API (e.g., GPT-4, Claude). Designed for easy integration without complex fine-tuning.
- AutoGEOMini (Cost-efficient): Uses a two-stage process: Cold Start with rule-based prompting, followed by GRPO fine-tuning. Reduces inference costs and latency compared to large LLM APIs.
Experimental Results
🚀 Superior Performance
AutoGEOAPI and AutoGEOMini consistently outperform all baselines across three datasets:
- AutoGEOAPI: Up to 50.99% improvement over the strongest baseline
- AutoGEOMini: Average 20.99% gain over baselines
✅ Key Findings
- Cross-Engine Robustness: Improvements are consistent across Gemini, GPT, and Claude engines
- Utility Preservation: Maintains or slightly improves generative engine utility metrics (Precision, Recall, Clarity, Insight)
- Challenging Scenarios: Substantially improves visibility for low-visibility documents
- Rule Transferability: Rules show high overlap across LLMs (78-84%) and transfer effectively across engines
🛡️ Cooperative vs. Adversarial
AutoGEO achieves strong visibility improvements without harming engine utility, unlike adversarial methods (Hijack Attack, Poisoning Attack) that degrade answer quality.
More details can be found in the 📄 Paper. Try it out in the 🚀 AutoGEOMini Demo.
Resources
📚 Paper & Code & Demo
🤖 AutoGEOMini Checkpoints
Citation
If you find AutoGEO useful, please cite:
@article{wu2025generative,
title={What Generative Search Engines Like and How to Optimize Web Content Cooperatively},
author={Wu, Yujiang and Zhong, Shanshan and Kim, Yubin and Xiong, Chenyan},
journal={arXiv preprint arXiv:2510.11438},
year={2025}
}