Generative search engines represent a transition from traditional ranking-based retrieval to Large Language Model (LLM)-based synthesis, transforming optimization goals from ranking prominence towards content inclusion. Generative Engine Optimization (GEO), specifically, aims to maximize visibility and attribution in black-box summarized outputs by strategically manipulating source content. However, existing methods rely on static heuristics, single-prompt optimization, or engine preference rule distillation that is prone to overfitting. They cannot flexibly adapt to diverse content or the changing behaviors of generative engines. Moreover, effectively optimizing these strategies requires an impractical amount of interaction feedback from the engines. To address these challenges, we propose AgenticGEO, a self-evolving agentic framework formulating optimization as a content-conditioned control problem, which enhances intrinsic content quality to robustly adapt to the unpredictable behaviors of black-box engines. Unlike fixed-strategy methods, AgenticGEO employs a MAP-Elites archive to evolve diverse, compositional strategies. To mitigate interaction costs, we introduce a Co-Evolving Critic, a lightweight surrogate that approximates engine feedback for content-specific strategy selection and refinement, efficiently guiding both evolutionary search and inference-time planning. Through extensive in-domain and cross-domain experiments on two representative engines, AgenticGEO achieves state-of-the-art performance and demonstrates robust transferability, outperforming 14 baselines across 3 datasets. Our code and model are available at: https://github.com/AIcling/agentic_geo.
Figures & tables
Figure 1. Characterization of the GEO result on GEO-Bench instances. y -axis reports maximum performance among 9 rewriting strategies, and x -axis reports performance variance among strategies. (i) Optimization success varies greatly by strategy and content. (ii) Existing strategies fail to optimize nearly half of instances (points in gray and red areas), indicating static strategy pool is not enough and needs evolving. Details can be found in Appendix A.1.1 .
Figure 2. GEO v.s. AgenticGEO. Static GEO methods apply fixed rewriting heuristics, whereas AgenticGEO maintains an evolving strategy archive and a critic to adaptively retrieve high-scoring strategies for iterative rewriting.
Figure 3. Overview of the AgenticGEO framework with two-stage training. Offline Alignment warm-starts a surrogate critic using offline preference pairs from the initial archive M0 , calibrating it for fast, content-conditioned strategy scoring. Online Co-Evolution then interacts with the black-box Generative Engine (GE) to iteratively, simultaneously evolve a MAP-Elites quality-diversity archive and the critic. Parent strategies are mutated by a learned evolver trained with sibling-aware AWR, the critic screens candidates to reduce GE calls, and newly collected GE feedback is stored in a replay buffer to continually recalibrate the critic and update the archive through a value-novelty gate. At inference time, the evolved archive and critic enable agentic multi-turn rewriting by selecting and executing a content-adaptive plan of strategies.
Table 1. Structure of the evolving strategy representation. Each dimension is mutated independently. It can be rendered into a compact summary for critic scoring and a full strategy prompt for rewriting (see Appendix A.1.5 ).
Methods
GEO-Bench
Qwen2.5-32B-Instruct
Llama3.3-70B-Instruct
word
pos
overall
word
pos
overall
No optimization
20.05
20.26
20.21
19.19
19.33
19.20
Keyword Stuffing
20.73
20.86
20.69
19.99
20.16
20.02
Unique Words
17.59
17.94
17.78
16.78
16.66
16.56
Easy-To-Understand
20.10
20.19
20.05
18.72
18.93
18.85
Table 2. Overall Performance on the in-domain setting. Average results on 5 independent runs are reported. ∗ indicates the statistically significant improvements over the best baseline, with p -value smaller than 0.001 .
Qwen2.5-32B-Instruct
Llama3.3-70B-Instruct
Methods
MS MARCO
E-Commerce
MS MARCO
E-Commerce
word
pos
overall
word
pos
overall
word
pos
overall
word
pos
overall
No optimization
19.99
20.15
20.05
18.30
18.09
18.01
19.45
19.64
19.67
19.70
19.45
19.68
Keyword Stuffing
23.26
22.63
22.75
18.94
18.52
18.65
22.02
21.77
21.96
20.24
19.96
20.16
Unique Words
17.95
18.24
18.23
18.12
18.09
18.01
18.94
18.88
18.76
19.16
19.04
19.15
Easy-To-Understand
20.48
20.46
20.46
20.46
20.45
20.29
19.75
19.84
19.84
20.28
19.91
20.18
Table 3. Overall Performance on the cross-domain setting. Average results on 5 independent runs are reported. ∗ indicates the statistically significant improvements over the best baseline, with p -value smaller than 0.001 .
Figure 4. Ablation study of AgenticGEO on three datasets. We compare AgenticGEO with four variants: (a) an offline-only critic trained without online co-evolution, (b) removing the evolved strategy archive, (c) replacing the critic with random rewrite planning and the evolver directly generates the strategy, (d) maintaining the archive by performance only without diversity.
Figure 5. Hyper-parameter sensitivity analysis on GEO-Bench. We report overall impression score when varying the multi-turn rewriting steps (left) and the archive size (right).
Figure 9
Figure 7. Semantic consistency and optimization effectiveness. Semantic Similarity is measured by BERTScore-F1 ( Zhang et al., 2019 ) with roberta-large ( Liu et al., 2019 ) , computed between the original content and the rewritten version.
Appendix figures & tables3 assets
Supplementary material from the paper’s appendix.
Appendix
Dataset
#Queries
#Docs
Avg. Content Tokens
GEO-Bench
1000
5,000
980.61
MS-Marco
1000
5,000
91.91
E-commerce
416
2,180
1,459.69
Appendix
Table 5. Dataset statistics after preprocessing. Following the GEO-Bench protocol, each query is paired with 5 documents.
Module
Param.
Description
Value
Offline Critic Alignment
Loss weight
λ
Weight in Ltotal=Lpair+λLreg
0.2
Warm-up steps
Sfreeze
Epochs with frozen backbone before unfreezing
1
Online Co-Evolution (Selection & Budget)
Iterations
T
Total online evolution iterations
100
Exploit size
Ktop
Top- Ktop selected by critic
4
Appendix
Table 6. Key hyperparameters of AgenticGEO (values left blank).
Figure 8. Qualitative case studies of AgenticGEO. For each query-content pair, we show the original content, the optimized rewrite, and the activated strategy sequence selected by critic-guided planning.
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.
Sheldon Yu, Rui Wang, Tong Yu +4
UC San Diego La Jolla, CA, USA · Adobe Research San Jose, CA, USA
Generative engines (GEs) are reshaping information access by replacing ranked links with citation-grounded answers, yet current Generative Engine Optimization (GEO) methods optimize each instance in isolation, unable to accumulate or transfer effective strategies across tasks and engines. We reframe GEO as a strategy learning problem and propose MAGEO, a multi-agent framework in which coordinated planning, editing, and fidelity-aware evaluation serve as the execution layer, while validated editing patterns are progressively distilled into reusable, engine-specific optimization skills. To enable controlled assessment, we introduce a Twin Branch Evaluation Protocol for causal attribution of content edits and DSV-CF, a dual-axis metric that unifies semantic visibility with attribution accuracy. We further release MSME-GEO-Bench, a multi-scenario, multi-engine benchmark grounded in real-world queries. Experiments on three mainstream engines show that MAGEO substantially outperforms heuristic baselines in both visibility and citation fidelity, with ablations confirming that engine-specific preference modeling and strategy reuse are central to these gains, suggesting a scalable learning-driven paradigm for trustworthy GEO. Code is available at https://github.com/Wu-beining/MAGEO
Beining Wu, Fuyou Mao, Jiong Lin +7
Hangzhou Dianzi University · Central South University · Shanghai Artificial Intelligence Laboratory +3
The landscape of search has changed drastically with how people look for information online. Traditional search engines are being replaced by Generative Search Engines (GSEs), which use Large Language Models (LLMs) to generate natural language responses to user queries. For content creators, visibility is no longer solely determined by ranking in search results but by being cited within generated responses. But Generative Search Engines are black-boxes, leading to the emergence of Generative Engine Optimization (GEO), a set of techniques aimed at improving content visibility in generative search settings. Most existing approaches rely on the explicit queries or query derived signals to align content to better suit user needs. We propose Query Implied Generative Engine Optimization (QI-GEO) to infers user intent directly from the document. Our approach approximates document's intent space and identifies content that may be missing yet relevant to answer potential user queries. Evaluation on GEO-Bench and Extended GEO-Bench demonstrated improvements across objective and subjective metrics. QI-GEO improved objective scores by up to 15.9% and subjective scores by up to 17.6%, while yielding nearly twice as many citation gains as citation losses. These results suggest that document-derived approximations of user intents can improve visibility without relying on explicit query inputs.
Shilpa Ramakrishna, William B. Andreopoulos
Department of Computer Science San Jose State University San Jose, CA, USA