AgenticGEO: A Self-Evolving Agentic System for Generative Engine Optimization

A research team led by Jiaqi Yuan has introduced AgenticGEO, a novel framework that reimagines content optimization for the era of AI-powered search. Unlike traditional SEO focused on ranking algorithms, Generative Engine Optimization (GEO) aims to ensure content gets included and properly attributed in the summarized outputs of systems like ChatGPT or Perplexity. Current GEO methods struggle with static strategies that can't adapt to diverse content types or the unpredictable behavior of black-box LLMs, often requiring impractical amounts of direct engine feedback for optimization.

AgenticGEO addresses these limitations by formulating GEO as a content-conditioned control problem. The system employs a MAP-Elites evolutionary algorithm to develop diverse, compositional optimization strategies that evolve over time. To reduce the costly need for constant interaction with actual search engines, the researchers created a lightweight Co-Evolving Critic—a surrogate model that approximates engine feedback for strategy selection and refinement. This approach allows for efficient evolutionary search and inference-time planning.

Through extensive testing on two representative generative engines across three datasets, AgenticGEO demonstrated state-of-the-art performance, outperforming 14 existing baselines. The system showed particular strength in cross-domain transferability, meaning strategies developed for one type of content or engine could effectively adapt to others. This represents a significant advancement toward making content optimization for AI search engines more adaptive, efficient, and robust against the opaque nature of modern LLM-based systems.