Optimizing Generative Search Engines with AgenticGEO

Generative search engines represent an evolution from traditional ranking-based systems, shifting the focus towards synthesis based on Large Language Models (LLMs). Generative Engine Optimization (GEO) aims to maximize visibility and attribution in summarized outputs by strategically manipulating source content.

AgenticGEO is a framework that addresses the challenges posed by existing methods, often limited by static heuristics or the difficulty of adapting to the changing behaviors of search engines. The system formulates optimization as a content-conditioned control problem, improving the intrinsic quality of the content to robustly adapt to the unpredictable behaviors of engines.

Architecture and Operation

Unlike fixed strategies, AgenticGEO employs a MAP-Elites archive to evolve diverse, compositional strategies. To reduce interaction costs, it introduces a Co-Evolving Critic, a lightweight surrogate that approximates engine feedback for content-specific strategy selection and refinement, efficiently guiding both evolutionary search and inference-time planning.

Experimental results demonstrate that AgenticGEO achieves superior performance compared to 14 baselines across 3 datasets, showing solid transferability between different domains. The code and model are available on GitHub.