AgenticGEO: A Self-Evolving Agent Optimization System for Generative Search Engines

AgenticGEO is an open-source Generative Engine Optimization (GEO) system developed by the AIcling team, which transforms traditional static heuristic methods into a content-adaptive agent system. Through the MAP-Elites strategy archive and co-evolutionary evaluator, it achieves efficient content optimization with minimal query feedback and demonstrates excellent transferability in cross-domain experiments. Its core goal is to maximize the visibility and attribution of content in generative engine outputs, addressing challenges such as difficulty in predicting black-box systems, static strategies' inability to adapt to diversity, and high feedback costs.

Search engines are shifting from ranking-based retrieval models to LLM-centric generative synthesis models. The goal of content optimization has changed from pursuing rankings to striving for adoption and citation by generative engines. Generative Engine Optimization (GEO) has emerged as a result, but it faces three major challenges: generative engines are black-box systems, traditional static strategies cannot adapt to content diversity, and optimization requires high costs due to extensive engine interaction feedback.

AgenticGEO defines GEO as a content-conditional control problem, with its core architecture consisting of three components: 1. MAP-Elites Strategy Archive: A quality-diversity strategy memory bank that adaptively selects optimal strategy combinations; 2. Co-Evolutionary Evaluator: A lightweight proxy model that approximates engine feedback, reduces reliance on real queries, and guides multi-round rewriting during inference; 3. Inference-Time Multi-Round Rewriting Mechanism: Iterative content improvement with dynamically adjusted strategies.

The workflow of AgenticGEO is divided into three stages: 1. Offline Evaluator Alignment: Using Qwen2.5-1.5B fine-tuned via LoRA to establish initial strategy evaluation capabilities; 2. Online Strategy-Evaluator Co-Evolution: Evolving the strategy archive under a limited query budget and calibrating the evaluator; 3. Inference-Time Planning and Execution: The evaluator analyzes content features, selects combinations from the top 25 strategies, and performs up to 3 rounds of rewriting iterations.

In the GEO-Bench benchmark test, evaluation is based on the number of attribution words, position-weighted citation order, and comprehensive metrics. The experiments cover in-domain testing, cross-domain testing, and low-feedback scenarios, with downstream engines including Qwen2.5-32B-Instruct and Llama-3.3-70B-Instruct. The results show that AgenticGEO is the best among 14 baselines and maintains strong performance even with an extremely low feedback budget, verifying the effectiveness of the co-evolutionary evaluator in reducing query costs.

AgenticGEO has been fully open-sourced, with the code repository providing reproducibility guidelines. It supports datasets such as GEO-Bench, MSdata, and e-commerce data, and is compatible with local OpenAI-compatible APIs (vLLM, llama.cpp) and cloud APIs. It uses the MIT license to encourage community contributions, supports preloading cache and concurrent processing, and allows configuration of model paths and parameters via environment variables.

AgenticGEO marks the shift of GEO from rule-driven to agent-driven. For creators: optimization strategies are more personalized and dynamic, with no fixed templates; for SEO practitioners: automated adaptive optimization becomes a core competency, and open-source code and pre-trained models provide a foundation for research and practice.

AgenticGEO combines evolutionary algorithms, reinforcement learning, and agent architecture to open up a new path for GEO. Its self-evolving characteristics and low-feedback optimization capabilities give it significant advantages in practical deployment. As generative AI develops, such adaptive optimization systems will play a more important role in the content ecosystem.