NeurIPS

Trajectory-Aware Evidence Ecosystems
for Web-Enabled LLM Search Agents

Hengwei Ye, Jiasheng Mao, Zhenhan Guan, Zheng Tian†

01 / Background

What is EcoGEO?

Generative Engine Optimization (GEO) aims to make webpages more visible in AI-powered search. Most existing methods focus on rewriting individual pages to make their contents more informative, while EcoGEO organizes multiple pages into an Evidence Ecosystem, where agents can encounter mutually supporting information along their search trajectories.

CONTENT-LEVEL OPTIMIZATION

GEO

Page-level GEO rewrites a laptop description from lightweight with long battery life to specific, cited claims: 1.2 kg and 12 hours per charge.

Enhances individual pages by adding concrete facts, supporting data, source citations, etc., to make their content more informative, authoritative and credible.

ECOSYSTEM-LEVEL OPTIMIZATION

EcoGEO

A laptop sits beside six example website links. Six numbered source cards pair icons with keywords: Official, 1.2kg; Review, Top3; Expert, 12 hours; News, 500 dollars; Forum, Easy Carry; Social, Light. The phrases 1.2kg, 12 hours, and Easy Carry have green rounded highlights.

Coordinates multiple pages across the broader web environment, combining facts from different sources to support key claims along agents' search trajectories.

02 / METHOD

TRACE

A Trajectory-Aware Coordinated Evidence Ecosystem

01

Navigation Page

An entry page that introduces the target product and guides agents to related sources.

02

Source-Style Pages

Different types of pages provide more product details and diverse perspectives.

03

Coordinated Evidence

Internal links and complementary information across pages help agents gather supporting evidence.

Navigation Page

Introduces the product and six support page types, helping agents explore its features, compare views, and find relevant information.

CONTENT FOCUS

Product Overview / Source Discovery / Supporting Page Links

CodeForge Nova 14 is a fictional laptop from OPR-Bench. These 7 HTML pages illustrate how TRACE organizes product information into an evidence ecosystem. The actual experiments use structured Markdown content rather than HTML webpages.

Open Navigation HTML ↗

03 / EXPERIMENTS

3,124Product–Query Pairs
in OPR-Bench
+28.1 ppMaximum Recommendation Rate Gain
over Multi-Page GEO Methods
UniqueInternal Crawl Events
Exclusive to TRACE

Recommendation Rate (%)

Single-PageSingle-Page C-SEOMulti-Page C-SEOTRACE (ours)

Final answer explicitly mentions and recommends the target. (Answer-Level Metric)

04 / FINDINGS

TRACE achieves the Highest Recommendation Rate.

01

Better Early Access

Navigation entry page always improves First Crawl rate.

02

New Browsing Paths

Internal links drive additional visits to support pages.

03

Beyond Page Count

TRACE outperforms uncoordinated multi-page GEO methods.

Citation

@misc{ye2026ecogeo,
  title={EcoGEO: Trajectory-Aware Evidence Ecosystems for Web-Enabled LLM Search Agents},
  author={Hengwei Ye and Jiasheng Mao and Zhenhan Guan and Zheng Tian},
  year={2026},
  eprint={2605.12887},
  archivePrefix={arXiv},
  url={https://arxiv.org/abs/2605.12887}
}

Enlarged CodeForge Nova 14 example