ICLR 2026 | OpenAI打广告后，如何成为爆款？CMU提出AutoGEO解密流量密码

Core Insights - The article discusses how AI search engines are gradually replacing traditional search methods, with "asking AI" becoming a daily habit. The introduction of commercial recommendations in ChatGPT by OpenAI is redefining the boundaries between search and content distribution. The ability of content to become a "hit" in AI searches now depends more on AI's citation preferences rather than just titles and traffic [2]. Group 1: Generative Search Engines and Content Optimization - Generative search engines (GE), such as Google AI Overview and ChatGPT, are establishing new traffic rules. The process involves "retrieving + synthesizing + generating," where the visibility of content is determined by how often and in what manner it is included in answers, rather than just ranking [7]. - The concept of Generative Engine Optimization (GEO) has emerged, focusing on optimizing web content to increase the likelihood of being included in AI-generated answers. Current GEO practices rely heavily on intuition and experience, which may not always align with AI preferences [7][9]. - The article emphasizes a cooperative approach to optimization, where enhancing visibility should not compromise the utility of generative engines [7]. Group 2: AutoGEO Framework - The core contribution of the paper is the introduction of AutoGEO, which extracts generative engine preference rules from a large dataset of visibility-differentiated evidence. These rules are then used to rewrite web pages while assessing their impact on generative engine utility (GEU) [9][10]. - AutoGEO operates in a four-step process: Explainer, Extractor, Merger, and Filter, which compresses vast comparative samples into executable rules [12]. - Two deployment routes for AutoGEO are proposed: AutoGEO API, which integrates rules into prompts for LLMs, and AutoGEO Mini, a low-cost model fine-tuned for rewriting [13][14]. Group 3: Evaluation and Performance Metrics - The evaluation of AutoGEO goes beyond visibility metrics to explicitly assess whether the utility of the generative engine is compromised. Experiments were conducted across three datasets: GEO-Bench, Researchy-GEO, and E-commerce, testing AutoGEO's performance in various contexts [16][18]. - Results indicate that both AutoGEO API and AutoGEO Mini significantly improve visibility metrics, with AutoGEO API showing a remarkable increase of 50.99% over the strongest baseline model [18]. Group 4: Rules and Content Ecosystem Implications - The paper reveals that while there are commonalities in preference rules across different LLM engines, there are also unique rules specific to each engine. This suggests a potential fragmentation in content strategies across different domains [24][25]. - The findings imply that there may be a future need for multiple versions of the same page tailored to different engines or domain intents, indicating a shift in content creation strategies [28]. Group 5: Future Considerations - AutoGEO opens new avenues for content creators, allowing for the extraction of preferences, updating of rules, and training of rewrites at minimal costs. The reality of the generative search era suggests that content must first pass through the "digestive system" of answer machines [29]. - The ongoing competition will not only be among content creators but also between engines and ecosystem governance, raising questions about the ability of engines to distinguish between informative content and content designed for citation [30].