Source-linked AI summary
Agent2UCB: Agentic System for Generative Engine Optimization
Sheldon Yu, Rui Wang, Tong Yu, Sungchul Kim, Doga Dogan, Junda Wu, Julian McAuley
TL;DR
Generative search gives site operators limited control over page selection and citation, while existing GEO systems either use fixed category-level rewrites or exhaustively test tools. Agent2UCB adaptively evaluates nine GEO strategies per content item with a bandit controller using LLM-informed priors and online visibility rewards, matching or exceeding baselines while reducing exploration to about 15 iterations and preserving SEO quality.
Problem
Existing GEO tooling offers limited control and typically uses fixed category-level strategies or exhaustive evaluation, despite optimization effectiveness varying across content items.
Method
Agent2UCB applies and evaluates nine GEO tools per content item, using a bandit controller that combines LLM-informed priors with online visibility rewards to select methods adaptively.
Results
Agent2UCB matches or exceeds category-level and exhaustive-search baselines on average GEO score, converging within about 15 iterations and roughly 10× fewer LLM calls than exhaustive search.
Takeaways & Limitations
Per-item bandit selection supports scalable GEO deployment while exposing GEO gains and SEO trade-offs through readability, topical coverage, and EEAT evaluations.
Abstract
from arXiv · showhide
Large language model driven search engines such as Google AI Overviews and Perplexity have created new opportunities for Generative Engine Optimization (GEO) the practice of refining content to increase its likelihood of being cited or summarized by generative systems. We demonstrate Agent2UCB, an agentic GEO system that autonomously improves content visibility through customized, feedback-driven optimization. For each content item, the system evaluates nine GEO strategies, identifies the most effective method, and accelerates selection using a bandit-based Agent2UCB policy that integrates LLM priors with online reward signals. To monitor side effects, the system also provides a lightweight, text-only SEO readiness evaluation covering readability, topical coverage, and EEAT-style credibility. Experiments on GEO-Bench show consistent visibility gains while preserving SEO quality. The demo allows users to choose the websites of interest, observe the optimization workflow, and compare GEO/SEO outcomes across methods.
I. INTRODUCTION
Agent2UCB addresses the limits of generic or exhaustive GEO by adapting optimization to each content item. It combines feedback-driven tool selection with GEO and SEO-style analytics for human inspection.
- Motivation: Existing GEO systems often use one category-level strategy, overlook item-level variation, or exhaustively test every tool.These patterns limit generalization and increase computational cost for large-scale or interactive deployment.
- System contribution: Agent2UCB evaluates nine GEO optimization tools per content item and selects the best-performing method through feedback-driven comparison.This provides per-item customization beyond category-level GEO baselines.
- System contribution: The Agent2UCB controller uses LLM-informed priors and online visibility rewards to balance exploration and exploitation.The controller identifies effective tools within about 15 iterations, described as 10× more efficient than brute-force search.
- Human-facing analytics: A text-only SEO readiness pipeline measures readability, topical coverage, and credibility alongside GEO scores in an interactive dashboard.The dashboard lets users inspect how optimization affects generative visibility and SEO-style quality.
- Overall contribution: Together, the contributions form a scalable, adaptive framework linking agentic reasoning, online learning, and human-centric GEO analytics.The system is presented as a practical Lighthouse-style tool for generative search.
II. SYSTEM ARCHITECTURE
Agent2UCB processes candidate pages, diagnoses their content quality, applies selected GEO tools, and repeatedly re-evaluates optimized drafts. Its dual-UCB controller uses priors and observed visibility gains to reduce exploration while tracking SEO readiness.
- Input and GEO scoring: The system accepts a query and multiple website links, extracts page text, and computes generative-visibility scores for each URL.The GEO evaluator simulates generative search by measuring how often and prominently each page is cited.
- Content analysis: Target-page analysis produces EEAT-R diagnostics and natural-language suggestions that guide likely effective optimization tools.Examples include adding citations, clarifying authorship, and simplifying language.
- GEO tool library: The tool library includes authoritative rewriting, evidence injection, quote and citation, readability tuning, and SEO-friendly structuring.Each tool is an LLM prompt template mapping a query-document pair to an optimized draft.
- Bandit controller: Agent2UCB models each tool as a bandit arm whose reward is the observed visibility gain for a fixed query-document pair.The controller uses a UCB-style policy with empirical reward, selection counts, and a tunable exploration parameter.
- Evaluation loop: The workflow re-runs GEO evaluation and SEO readiness analysis after each tool application, updating reward signals and the user-facing dashboard.The resulting loop surfaces optimized content and comparative analytics to the user.
- Bandit controller: LLM-based priors warm-start underexplored tools, while online GEO gains can override poor priors during at most 10–15 iterations per item.The dual-UCB design is intended to reduce LLM calls relative to exhaustive evaluation.
III. EXPERIMENT
Agent2UCB matches or exceeds baseline GEO performance while converging with far fewer LLM calls, and its SEO effects are mixed rather than uniformly positive. The results also show that effective tool choice varies across content items.
- Agent2UCB matches or exceeds category-level and exhaustive-search baselines on average GEO score.
- RQ2: About 15 bandit iterations achieve comparable or higher scores with roughly 10× fewer LLM calls than exhaustive search.
- No single optimization tool dominates across content items, indicating that the most effective strategy varies by item.
- RQ3: Agent2UCB attains the highest Overall SEO Score and raises EEAT credibility by roughly 80% over unedited pages and 95% over the category baseline.
- RQ3: Agent2UCB has a modest readability dip and improves topical coverage over the original while remaining below the category baseline.
- The dashboard compares GEO/SEO outcomes across methods and supports EEAT-R analysis and optimized-draft inspection.
IV. CONCLUSION
The paper frames GEO tool selection as adaptive exploration–exploitation rather than exhaustive search. Agent2UCB combines bandits with visibility rewards, offline predictions, and online feedback to reduce computational cost while maintaining competitive performance.
- Agent2UCB formulates GEO tool selection as an adaptive exploration–exploitation process instead of exhaustive search.
- Integrating multi-armed bandits with verifiable visibility rewards reduces computational cost while maintaining competitive performance.
- The controller combines offline predictions with online feedback to mitigate cold-start issues.
- Agent2UCB converges to the optimal optimization method within a few iterations, supporting scalable deployment in generative search systems.