OpenNovelty.org

What Generative Search Engines Like and How to Optimize Web Content Cooperatively

ICLR 2026 Conference SubmissionAnonymous Authors
OpenReview Score: 6.0Download Report PDF
generative engine optimizationgenerative enginespreference rule discoveryreinforcement learning
Abstract:

By employing large language models (LLMs) to retrieve documents and generate natural language responses, Generative Engines, such as Google AI overview and ChatGPT, provide significantly enhanced user experiences and have rapidly become the new form of search. Their rapid adoption also drives the needs of Generative Engine Optimization (GEO), as content providers are eager to gain more traction from them. In this paper, we introduce AutoGEO, a framework to automatically learn generative engine preferences when using retrieved contents for response generation, and rewrite web contents for more such traction. AutoGEO first prompts frontier LLMs to explain generative engine preferences and extract meaningful preference rules from these explanations. Then it uses preference rules as context engineering for AutoGEO_API\_\text{API}, a prompt-based GEO system, and as rule-based rewards to train AutoGEO_Mini\_\text{Mini}, a cost-effective GEO model. Experiments on the standard GEO-Bench and two newly constructed benchmarks using real user queries demonstrate the effectiveness of AutoGEO in enhancing content traction while preserving search utility. Analyses confirmed the learned rules' robustness and abilities to capture unique preferences in variant domains, and AutoGEO systems' ability to embed them in content optimization. The learned preference rules, our models, and codes will be open-sourced.

Disclaimer
This report is AI-GENERATED using Large Language Models and WisPaper (A scholar search engine). It analyzes academic papers' tasks and contributions against retrieved prior work. While this system identifies POTENTIAL overlaps and novel directions, ITS COVERAGE IS NOT EXHAUSTIVE AND JUDGMENTS ARE APPROXIMATE. These results are intended to assist human reviewers and SHOULD NOT be relied upon as a definitive verdict on novelty.
NOTE that some papers exist in multiple, slightly different versions (e.g., with different titles or URLs). The system may retrieve several versions of the same underlying work. The current automated pipeline does not reliably align or distinguish these cases, so human reviewers will need to disambiguate them manually.
If you have any questions, please contact: mingzhang23@m.fudan.edu.cn

Overview

Overall Novelty Assessment

The paper introduces AutoGEO, a framework for automatically learning generative engine preferences and rewriting web content to increase visibility in AI-powered search systems. Within the taxonomy, it occupies a unique position as the sole paper in the 'Generative Engine Optimization and Content Adaptation' leaf under 'Meta-Level Preference Analysis and Optimization'. This isolation suggests the work addresses an emerging research direction with minimal direct prior work in the examined literature. The leaf's focus on optimizing content for generative search engines distinguishes it from the broader preference learning foundations that dominate other branches.

The taxonomy reveals substantial activity in adjacent areas but limited overlap with AutoGEO's specific focus. Neighboring leaves address personalization (e.g., user-specific adaptation with four papers), recommender systems (generative retrieval and end-to-end recommendation with three papers), and conversational systems (three papers). These directions emphasize individual user modeling or traditional retrieval optimization, whereas AutoGEO targets engine-level preference extraction for content adaptation. The 'Preference Learning Foundations' branch contains dense clusters on offline optimization methods and RLHF, but these focus on model alignment rather than content rewriting strategies for search visibility.

Across all three contributions—the AutoGEO framework, AutoGEO_API, and AutoGEO_Mini—the analysis examined ten candidates each from a total pool of thirty papers, with zero refutable candidates identified for any contribution. This suggests that among the limited semantic search results and citation expansion, no prior work directly addresses automated preference rule extraction for generative engine optimization or the specific prompt-based and RL-based content rewriting approaches proposed. The absence of refutable candidates across thirty examined papers indicates the work explores a relatively uncharted application domain within the broader preference learning landscape.

Given the limited search scope of thirty candidates and the paper's placement in a singleton taxonomy leaf, the work appears to pioneer a distinct application area. However, the analysis cannot confirm absolute novelty beyond the examined literature. The taxonomy structure shows rich activity in related preference learning methods and personalization techniques, suggesting potential conceptual connections that may not have surfaced in the semantic search. The contribution's novelty likely stems from its specific application context rather than foundational algorithmic innovation.

Taxonomy

Core-task Taxonomy Papers
50
3
Claimed Contributions
30
Contribution Candidate Papers Compared
0
Refutable Paper

Research Landscape Overview

Core task: Generative engine optimization through automated preference rule learning. The field structure reflects a broad effort to understand and leverage human preferences across diverse generative systems. The taxonomy organizes work into several main branches: Preference Learning Foundations and Algorithms, which develops core methods for extracting preference signals from comparisons and feedback (e.g., Generalized Preference Optimization[2], Iterative Preference Learning from[4]); Visual and Multi-Modal Preference Learning, addressing image and video generation alignment (e.g., ImageReward[1], VisionReward[11], Diffusion-RPO[6]); Personalization and Individual Preference Modeling, focusing on user-specific adaptation in recommendation and content generation (e.g., Onerec[3], Diverse preference learning for[5]); Domain-Specific Preference Learning Applications, applying preference techniques to specialized tasks like design optimization and robotics; and Meta-Level Preference Analysis and Optimization, which examines higher-order questions about preference aggregation, multi-objective trade-offs (e.g., Multi-Objective Preference Optimization[7]), and how generative engines can be steered toward desired outcomes. Several active lines of work reveal key tensions and open questions. One strand explores how to scale preference learning beyond simple pairwise comparisons, incorporating multi-dimensional human judgments (Learning Multi-Dimensional Human Preference[10]) and aligning to thousands of diverse criteria (Aligning to Thousands of[22]). Another investigates the interplay between preference modeling and generative retrieval or recommendation systems (Generative retrieval with preference[12], Generative Recommender with End-to-End[43]), where learned preferences directly shape content generation. What Generative Search Engines[0] sits within the Meta-Level Preference Analysis and Optimization branch, specifically addressing how automated rule learning can optimize content for generative search platforms. Its emphasis on adapting content to engine-level preferences contrasts with works like Modeling and Optimizing User[14] or Improving Generative AI Student[19], which focus more on individual user modeling, and complements broader optimization frameworks such as Automatic design optimization of[8] by targeting the unique dynamics of generative engines rather than traditional design or recommendation contexts.

Claimed Contributions

AutoGEO framework for automatic preference rule extraction

AutoGEO automatically learns generative engine preferences by prompting frontier LLMs to explain preferences, extract meaningful rules from explanations, merge insights into candidate rules, and filter them into actionable preference rules. This transforms tens of thousands of preference observations into a systematic set of rules capturing how generative engines prioritize content.

10 retrieved papers
AutoGEOAPI: plug-and-play prompt-based GEO model

AutoGEOAPI directly embeds extracted preference rules into instruction templates for powerful LLMs, creating a GEO model that requires no additional training and can be readily applied across different generative engines to rewrite documents for improved visibility.

10 retrieved papers
AutoGEOMini: cost-efficient reinforcement learning-based GEO model

AutoGEOMini is a compact model trained via reinforcement learning using extracted preference rules as reward signals. It employs supervised fine-tuning for cold start followed by group relative policy optimization (GRPO), achieving cost-efficient GEO performance at approximately 0.0071 times the cost of the API-based approach.

10 retrieved papers

Core Task Comparisons

Comparisons with papers in the same taxonomy category

Within the taxonomy built over the current TopK core-task papers, the original paper is assigned to a leaf with no direct siblings and no cousin branches under the same grandparent topic. In this retrieved landscape, it appears structurally isolated, which is one partial signal of novelty, but still constrained by search coverage and taxonomy granularity.

Contribution Analysis

Detailed comparisons for each claimed contribution

Contribution

AutoGEO framework for automatic preference rule extraction

AutoGEO automatically learns generative engine preferences by prompting frontier LLMs to explain preferences, extract meaningful rules from explanations, merge insights into candidate rules, and filter them into actionable preference rules. This transforms tens of thousands of preference observations into a systematic set of rules capturing how generative engines prioritize content.

Contribution

AutoGEOAPI: plug-and-play prompt-based GEO model

AutoGEOAPI directly embeds extracted preference rules into instruction templates for powerful LLMs, creating a GEO model that requires no additional training and can be readily applied across different generative engines to rewrite documents for improved visibility.

Contribution

AutoGEOMini: cost-efficient reinforcement learning-based GEO model

AutoGEOMini is a compact model trained via reinforcement learning using extracted preference rules as reward signals. It employs supervised fine-tuning for cold start followed by group relative policy optimization (GRPO), achieving cost-efficient GEO performance at approximately 0.0071 times the cost of the API-based approach.

Table of Contents