From Experience to Skill: Multi-Agent Generative Engine Optimization via Reusable Strategy Learning
Summary: arXiv:2604.19516v1 Announce Type: new
Abstract: Generative engines (GEs) are reshaping information access by replacing ranked links with citation-grounded answers, yet current Generative Engine Optimization (GEO) methods optimize each instance in isolation, unable to accumulate or transfer effective strategies across tasks and engines. We reframe GEO as a strategy learning problem and propose MAGEO, a multi-agent framework in which coordinated planning, editing, and fidelity-aware evaluation serve as the execution layer, while validated editing patterns are progressively distilled into reusable, engine-specific optimization skills. To enable controlled assessment, we introduce a Twin Branch Evaluation Protocol for causal attribution of content edits and DSV-CF, a dual-axis metric that unifies semantic visibility with attribution accuracy. We further release MSME-GEO-Bench, a multi-scenario, multi-engine benchmark grounded in real-world queries. Experiments on three mainstream engines show that MAGEO substantially outperforms heuristic baselines in both visibility and citation fidelity, with ablations confirming that engine-specific preference modeling and strategy reuse are central to these gains, suggesting a scalable learning-driven paradigm for trustworthy GEO. Code is available at https://github.com/Wu-beining/MAGEO.
Introduction
The landscape of information retrieval is undergoing a significant transformation with the advent of Generative Engines (GEs). These engines have the potential to revolutionize how users access information by providing citation-grounded answers rather than conventional ranked links. However, despite their promise, existing optimization techniques for GEs, known as Generative Engine Optimization (GEO), often tackle optimization on a per-instance basis, which limits their ability to learn from past experiences or transfer strategies across different tasks and engines.
Proposed Solution: MAGEO Framework
In response to these challenges, we introduce MAGEO, a novel multi-agent framework that reinterprets GEO as a strategy learning problem. This framework employs a collaborative approach where agents are tasked with coordinated planning, editing, and evaluating content fidelity. The unique aspect of MAGEO is its focus on distilling validated editing patterns into reusable optimization skills that are specific to individual engines.
Evaluation Methods
To ensure a rigorous assessment of the MAGEO framework, we developed the Twin Branch Evaluation Protocol. This protocol allows for a causal attribution analysis of content edits, while the DSV-CF metric integrates semantic visibility with attribution accuracy, providing a comprehensive evaluation of optimization strategies.
Benchmarking and Results
In conjunction with the MAGEO framework, we also released MSME-GEO-Bench, a benchmark designed to evaluate multiple scenarios and engines using real-world queries. Experimental results demonstrate that MAGEO significantly outperforms traditional heuristic baselines in both visibility and citation fidelity. Key findings reveal that engine-specific preference modeling and the reuse of strategies are crucial for achieving these performance improvements.
Conclusion
The development of MAGEO suggests a scalable, learning-driven approach to Generative Engine Optimization, enabling a more trustworthy and effective means of information retrieval. By leveraging reusable strategies, MAGEO sets the stage for future advancements in the field, making it a vital contribution to the ongoing evolution of generative technologies.