MAGEO: A Multi-Agent Optimization Framework for Generative Engines, Shifting Content from 'Being Retrieved' to 'Being Cited'

This article analyzes the MAGEO framework accepted by ACL 2026, discusses the paradigm shift of Generative Engine Optimization (GEO) from ranking competition to answer influence competition, introduces its multi-agent collaboration architecture, hierarchical memory mechanism, and DSV-CF evaluation system, and provides an intelligent automated path for content optimization.

Traditional SEO focuses on ranking, while generative engines (such as ChatGPT, Perplexity) directly generate comprehensive answers, and content needs to be selected and integrated to be visible. This brings four major challenges: opaque presentation methods, ambiguous optimization goals (tension between multi-dimensional goals), unclear optimization paths (complex causal chains), and differences in engine preferences. The MAGEO framework emerged as the times require, realizing the evolution from manual trial-and-error to intelligent automation.

MAGEO adopts a closed-loop optimization process and coordinates five major agents: Preference Agent (constructs engine preference profiles), Planning Agent (formulates revision blueprints), Editing Agent (generates multiple candidate versions), Evaluation Agent (DSV-CF scoring), and Fidelity Gatekeeper (ensures accuracy). The hierarchical memory system includes step-level (traces of successful single optimization) and creator-level (reusable patterns across scenarios), realizing the evolution from experience to skills.

The DSV-CF evaluation system includes Source Selection Visibility (SSV: lexical overlap, etc.) and Integrated Semantic Influence (ISI: attribution accuracy, etc.), and the comprehensive score takes safety into account (penalty for attribution errors). The Twin-Branch protocol requires that the baseline and optimized versions be evaluated under the same conditions to eliminate external variable interference and ensure reliable results.

The codebase focuses on core workflows and is modularly organized for easy expansion: Preference Agent (agent/preference_agent.py), Planning Agent (agent/planner_agent.py), Editing Agent (agent/editor_agent.py), Evaluation Agent (agent/evaluation_agent.py), main entry (pipeline/geo_optimizer.py), etc. Researchers can replace or expand components.

1. Quality remains fundamental: engines directly evaluate content value, and keyword stuffing is ineffective; 2. Structured expression is important: clear architecture and argument-evidence relationships help engines extract information; 3. Attribution and accuracy are the bottom line: avoid factual errors or ambiguous attribution; 4. Adapt to the multi-engine ecosystem: understand the preferences of different engines and adjust strategies.

Limitations: Only focuses on text optimization and does not involve multimedia; assumes engine preferences are stable, and adaptation to rapidly evolving models needs to be verified. Future directions: Multimodal GEO, dynamic preference adaptation, cross-language optimization, fine-grained user intent perception, to continuously promote the development of the GEO field.

MAGEO redefines content optimization as a multi-dimensional balance of visibility, influence, and reliability, and realizes skill accumulation through multi-agent collaboration and hierarchical memory. It provides a platform for researchers, reveals new strategies for practitioners, and raises a core question: In the era of machine-generated answers, how can human content maintain its unique value? The answer is to enable high-quality content to gain due visibility and influence through intelligent optimization.