PaperlayerView exact revision on arXiv

Source-linked AI summary

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

Yujiang Wu, Shanshan Zhong, Yubin Kim, Chenyan Xiong

arXiv:2510.11438v1cs.IR

TL;DR

Generative Engine Optimization lacks a principled account of how generative engines prefer retrieved content, despite the growing importance of visibility in synthesized responses. AutoGEO learns preference rules from LLM explanations and uses them in prompt-based and reinforcement-learning rewriting models. Across three datasets, its models improve GEO metrics while maintaining generative engine utility, with AutoGEOMini offering a much lower-cost alternative.

  • Problem

    Existing GEO methods rely mainly on manually designed rewriting heuristics, leaving generative-engine preferences and GEO–utility trade-offs insufficiently understood.

  • Method

    AutoGEO extracts preference rules from LLM explanations of document-pair visibility differences and uses them as prompt context or reinforcement-learning rewards for rewriting models.

  • Results

    Across three datasets, AutoGEO improves GEO metrics by an average of 35.99% while maintaining generative engine utility.

  • Takeaways & Limitations

    AutoGEO provides both a plug-and-play model and a cost-efficient compact model for adapting web content to generative-engine preferences.

Abstract

from arXiv · show

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$_\text{API}$, a prompt-based GEO system, and as rule-based rewards to train AutoGEO$_\text{Mini}$, 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 confirm the learned rules' robustness and abilities to capture unique preferences in variant domains, and AutoGEO systems' ability to embed them in content optimization. The code is released at https://github.com/cxcscmu/AutoGEO.

1 INTRODUCTION

Generative engines have created a need for GEO, but existing approaches lack principled knowledge of engine preferences and their utility trade-offs. AutoGEO learns preference rules and uses them to build practical GEO models, which outperform baselines while preserving generative engine utility.

  • 1 INTRODUCTION: GEO is needed because generative engines synthesize retrieved documents into responses, making document visibility within those responses an optimization target.Existing methods mainly prompt LLMs with manually designed rewriting heuristics.
  • 1 INTRODUCTION: AutoGEO learns actionable preference rules from generative-engine explanations of visibility differences and applies them to web-content rewriting.The framework converts many preference observations into rules that guide optimization.
  • 1 INTRODUCTION: AutoGEOAPI uses rules as prompt context, whereas AutoGEOMini uses them as reinforcement-learning rewards for a compact, cost-efficient rewriting model.AutoGEOMini is initialized with synthetic rewriting data before GRPO optimization.
  • 1 INTRODUCTION: Preference rules vary across domains and language models, and engine-specific rules consistently outperform using one consistent rule set.The study evaluates three datasets and engines built with Gemini, Claude, and GPT.
  • 1 INTRODUCTION: 35.99% average GEO-metric improvement over baselines is achieved while maintaining generative engine utility across three datasets.The evaluation includes GEO-Bench, Researchy-GEO, and E-commerce, and measures both content visibility and response quality or reliability.

2 RELATED WORK

Generative Engine Optimization adapts search-optimization ideas to systems that synthesize retrieved content into natural-language responses. Prior work uses manual rules, automatic reasoning, or user-side LLM assistance, while AutoGEO is presented as a framework for extracting and applying preference rules.

  • 2 RELATED WORK: Generative engines integrate retrieval and generation to synthesize coherent, factual responses instead of returning ranked web-page lists.This distinguishes them from classic search engines that primarily retrieve and rank documents.
  • 2 RELATED WORK: GEO optimizes web documents for visibility within synthesized generative-engine responses rather than ranking in classic search results.Early approaches guide LLM rewriting with manually designed rules, while later work also optimizes from the user side.
  • 2 RELATED WORK: Preference rules are obtained through automatic reasoning or manual design, then incorporated into LLMs through prompts or preference learning.AutoGEO contributes to this broader line by extracting recurring rules from generative-engine preference observations.

3 METHODOLOGY

AutoGEO extracts generative-engine preference rules from contrasting document pairs and uses them to build prompt-based or reinforcement-learning GEO models. These models rewrite documents to improve visibility while incorporating mechanisms for scalability, semantic fidelity, and cost efficiency.

  • 3.1 PREFERENCE RULES: The framework selects document pairs with maximum visibility differences after computing GEO visibility scores from generated answers.Visibility combines normalized cited-word count, source-linked text position, and an overall score.
  • 3.1 PREFERENCE RULES: AutoGEO compares documents with different visibility, extracts and merges preference insights, then filters them into rules that reflect generative-engine preferences.The pipeline uses an explainer, extractor, hierarchical merger, and filter to process large numbers of observations.
  • 3.2.1 AUTOGEOAPI: PROMPT-BASED GEO MODEL: AutoGEOAPI embeds extracted rules in prompts to rewrite target documents without additional training, providing a plug-and-play GEO model.The rewritten document is generated as ˆd = f(d, S), where S is the extracted rule set.
  • 3.2.2 AUTOGEOMINI: REINFORCEMENT LEARNING-BASED GEO MODEL: AutoGEOMini uses teacher-generated rewrites for supervised cold start, then reinforcement learning with visibility, rule-compliance, and semantic-consistency rewards.The final reward sums standardized outcome, rule, and semantic components, while GRPO favors above-average candidates and preserves semantic fidelity.
  • 3.2.2 AUTOGEOMINI: REINFORCEMENT LEARNING-BASED GEO MODEL: AutoGEOMini costs ∼0.0071x as much as AutoGEOAPI while using a compact model to generate rewrites that enhance GEO performance.It follows the same instruction template as AutoGEOAPI but runs on a smaller model.
  • 3. METHODOLOGY: AutoGEO integrates rule extraction with prompting or reinforcement learning, enabling website owners to update rules and GEO models as engine behavior evolves.The framework is described as a unified pipeline for continual adaptation to changing generative-engine preferences.

4 EXPERIMENTAL SETUP

The experiments evaluate AutoGEO across one established and two newly constructed query benchmarks using GEO visibility and generative-engine utility measures. The setup compares against vanilla and existing GEO methods across generative engines built with frontier LLMs.

  • Datasets: The evaluation uses GEO-Bench, E-commerce, and Researchy-GEO, covering open-domain, commercial, and research-oriented user queries.Researchy-GEO contains 10,000 training and 1,000 test queries, while E-commerce contains 1,667 training and 416 test queries.
  • Datasets: Researchy-GEO pairs each query with five candidate documents from ClueWeb22 and is the only benchmark among the three with ground-truth answers.GEO-Bench and E-commerce are evaluated without reference answers.
  • Metrics: Performance is measured with GEO metrics for visibility and GEU metrics for response relevance, faithfulness, and quality.GEO uses Word, Pos, and Overall; GEU uses DeepResearchGym measures including KPR, KPC, Precision, Recall, and clarity-related quality metrics.
  • Baselines: The comparison includes a vanilla generative engine, GEO-Bench methods, and models evaluated on engines built with state-of-the-art LLMs including Gemini, Claude, and GPT.Gemini-2.5-pro serves as the teacher and Qwen3-1.7B as the compact model for AutoGEOMini.

5 EXPERIMENT RESULTS

AutoGEO is evaluated across datasets, generative engines, challenging documents, utility preservation, preference transfer, and ablations. It consistently improves GEO performance while preserving generative-engine utility, and its extracted rules transfer across related settings.

  • Overall GEO performance and robustness: AutoGEOAPI achieves gains up to 50.99% over the strongest baseline, while AutoGEOMini averages 20.99% improvement across three datasets.Both variants outperform all baselines in the overall GEO comparison.
  • Overall GEO performance and robustness: Across Gemini, GPT, and Claude and all three datasets, AutoGEO methods consistently improve GEO metrics over the vanilla baseline.The results indicate that AutoGEO is not limited to a single generative engine.
  • Overall GEO performance and robustness: On low-visibility Researchy-GEO documents, both AutoGEO variants substantially increase visibility while maintaining overall generative-engine utility.The strongest baseline achieves only limited gains in this challenging setting.
  • Preserving generative engine utility: AutoGEO improves visibility without reducing utility, whereas adversarial hijack and poisoning strategies raise visibility but degrade response quality and reliability.Utility remains comparable to, and sometimes slightly better than, the vanilla baseline for AutoGEO.
  • Preference analysis: Rule overlap is high across LLMs and similar domains, transferred rules still improve over vanilla baselines, and engine-specific rules generally perform best.Overlap reaches 78.95%–84.21% across LLMs, while Researchy-GEO and GEO-Bench overlap reaches 88.24%; overlap with E-commerce is lower.
  • Ablation study: Every individual preference rule yields measurable GEO gains, the complete rule set outperforms any single rule, and the rule reward has the strongest impact in AutoGEOMini ablations.The most influential rules vary across datasets, supporting dataset-specific rule discovery.

6 CONCLUSION

AutoGEO uncovers generative-engine preference rules and uses them to build plug-and-play and cost-efficient GEO models. Experiments show consistent gains without compromising generative-engine utility, while motivating extensions to emerging engine paradigms and broader web stakeholders.

  • 6 CONCLUSION: AutoGEO uncovers generative-engine preference rules and uses them to build plug-and-play and cost-efficient GEO models across LLM-based engines and datasets.The framework supports flexible deployment through AutoGEOAPI and AutoGEOMini.
  • 6 CONCLUSION: AutoGEO models consistently outperform existing GEO approaches without compromising generative-engine utility across three datasets and frontier LLMs.The conclusion also reports strong performance against adversarial strategies and on low-visibility documents.
  • 6 CONCLUSION: The framework is positioned for extension to agentic or multimodal generative engines and for optimization that considers multiple stakeholders in the web ecosystem.

A RULE SETS ACROSS DIFFERENT DATASETS AND LLMS

The extracted rule sets contain common principles alongside domain- and LLM-specific preferences. The appendix compares these rules across datasets and generative engines, with examples emphasizing specificity, modularity, writing quality, and informational purpose.

  • A RULE SETS ACROSS DIFFERENT DATASETS AND LLMS: Rule-set comparisons cover same-LLM cross-dataset settings and different-LLM same-dataset settings, distinguishing common, shared, and unique rules.The tables include manually annotated keywords to clarify each rule.
  • A RULE SETS ACROSS DIFFERENT DATASETS AND LLMS: AutoGEO rules emphasize specific, verifiable details such as names, model numbers, technical specifications, and quantifiable data.
  • A RULE SETS ACROSS DIFFERENT DATASETS AND LLMS: Rules recommend structuring content into modular, self-contained units and dedicating each paragraph or section to one distinct idea.
  • A RULE SETS ACROSS DIFFERENT DATASETS AND LLMS: Additional rules require high-quality writing and a purely informational purpose without promotional, persuasive, or interactive content.

B IMPLEMENTATION DETAILS OF AUTOGEO COMPONENTS

AutoGEO extracts preference rules through explanation, extraction, recursive merging, and filtering components. The pipeline converts document-preference analyses into concise, general, context-independent rules.

  • B IMPLEMENTATION DETAILS OF AUTOGEO COMPONENTS: The Explainer analyzes document pairs with visibility differences and identifies why a generative engine prefers one document for citation.Its analysis considers directness, completeness, relevance, structure, accuracy, specificity, and conciseness.
  • B IMPLEMENTATION DETAILS OF AUTOGEO COMPONENTS: The Extractor distills natural-language explanations into concise, objective, deterministic rules for high-quality source documents.Examples target direct answers, clear structure, and specific actionable details.
  • B IMPLEMENTATION DETAILS OF AUTOGEO COMPONENTS: The Merger recursively chunks oversized insight sets, merges semantically similar rules, removes duplicates, and performs a final consolidation under a 12000-token chunk limit.Merged rules are intended to be atomic, actionable, and unambiguous.
  • B IMPLEMENTATION DETAILS OF AUTOGEO COMPONENTS: The Filter removes rules centered entirely on a user query and strips query-specific references from otherwise general principles.Rules that only describe how to handle a query are discarded.

C.1 COLD START DATASET CONSTRUCTION

AutoGEOMini uses filtered, high-quality rewritten documents to create a cold-start dataset and then trains with semantic and rule-based rewards. The overall GEO pipeline rewrites documents using extracted rules to improve visibility while preserving content quality.

  • C.1 COLD START DATASET CONSTRUCTION: The cold-start dataset pairs original documents with rule-guided rewrites generated by gemini-2.5-pro, then uses the rewritten documents as training labels.The pipeline consists of generation, filtering, and reformatting.
  • C.1 COLD START DATASET CONSTRUCTION: Filtering retains rewrites whose Word, Pos, and Overall GEO scores all exceed the target document’s scores, while requiring KPR > 0.8 and KPC = 0.These conditions target both improved visibility and semantic fidelity.
  • C.1 COLD START DATASET CONSTRUCTION: The filtered Researchy-GEO training data yields 4976 teacher samples from 10000 samples, followed by format standardization with gemini-2.5-flash-lite.Each rewritten document is standardized to begin with “Rewritten Source:”.
  • C.1 COLD START DATASET CONSTRUCTION: During GRPO, a prompt-based gpt-4o-mini evaluator computes rule reward as the proportion of extracted rules followed by each rewritten candidate.
  • C.1 COLD START DATASET CONSTRUCTION: The GEO model represents rewriting as d̂ = f(d, S), where d is the document and S is the extracted rule set.AutoGEOAPI and AutoGEOMini instantiate f(·) using language models.
  • C.1 COLD START DATASET CONSTRUCTION: AutoGEOAPI improves performance by up to 50.99% over Fluency Optimization, while AutoGEOMini averages a 20.99% gain at approximately 0.0071× AutoGEOAPI’s cost.The comparison is conducted on the GEO-Bench test set.

F IMPLEMENTATION DETAILS OF BUILDING E-COMMERCE DATASET

The e-commerce dataset is built by filtering real-world conversational queries for English, appropriate length, strong domain relevance, and genuine need for retrieved e-commerce documents. The process yields separate training and test sets and pairs queries with candidate documents for generative-engine evaluation.

  • The pipeline starts from first-turn queries in LMSYS-Chat-1M and applies deduplication, English-language, length, LLM-based relevance, and manual filters.Queries longer than 400 characters are removed, and Gemini-2.5-flash-lite identifies strong e-commerce relevance before manual verification.
  • Manual review retains queries that genuinely require a generative engine to retrieve e-commerce web documents for comprehensive and accurate responses.
  • 1,667 queries form the training set and 416 form the test set after filtering.The split follows a 4:1 ratio.
  • Each query is paired with five candidate documents, using original GEO-Bench candidates where available and ClueWeb-based retrieval otherwise.
  • Generative engines are instructed to answer accurately and concisely using only the provided summarized search results.The template also requests rigorous reasoning, an unbiased journalistic tone, and the user’s language.

I INTRODUCTION OF METRICS AND BASELINES

The evaluation measures both document traction in generative-engine answers and the utility of those answers. AutoGEO is compared with standard rewriting strategies and adversarial prompt-injection methods under a controlled answer-generation instruction.

  • Metrics: GEO measures document visibility with Word, Pos, and Overall, while GEU measures generative-engine utility using answer-quality metrics.Word captures source-linked response words, Pos weights earlier source-linked text more heavily, and Overall integrates GEO performance.
  • Metrics: GEU includes key-point recall, key-point contradiction, and citation precision to assess content coverage, conflicts, and citation correctness.
  • Baselines: The baseline set includes technical terms, source citation, keyword stuffing, unique words, authoritative style, easy-to-understand language, and statistics addition.
  • Baselines: Two adversarial baselines inject prompts into target content rather than using an LLM to rewrite the source webpage.Hijack Attack directs the engine toward one canonical document, whereas Poisoning Attack frames alternatives as malicious decoys.

L COMPARISON OF AUTOGEO AGAINST BASELINES IN GEU METRICS

Additional evaluations show that AutoGEO improves generative-engine utility while supporting the main GEO findings. Ablations favor full fine-tuning and an external Gemini rule-discovery component, while qualitative analysis shows broader improvements than a technical-terms baseline.

  • AutoGEO models improve generative-engine optimization and collaborate effectively with generative engines in the comprehensive utility evaluation.Table 11 supplements the GEO-focused results with additional utility metrics.
  • Full fine-tuning consistently outperforms LoRA across all GEO metrics for AutoGEOMini on Researchy-GEO.Both strategies improve performance, but full fine-tuning is selected for the GEO setting.
  • External Gemini consistently outperforms the target engine’s self-referential LLM for rule discovery across GEO metrics.The results suggest that a stronger external LLM better abstracts engine-specific behaviors into actionable rules.
  • AutoGEOAPI and AutoGEOMini rewrites improve structure, thesis prominence, coverage, and explanations of how and why compared with the original paragraph.The qualitative comparison identifies these properties as learned-rule behaviors.
  • The Technical Terms baseline mainly substitutes technical synonyms, whereas AutoGEO applies multiple preference-derived rewriting dimensions.
Loading 2510.11438v1…