Research map / Multi-agent systems and games / LLM agents

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

NeurIPS 2026 · Multi-agent systems and games › LLM agents · approach: Agent data synthesis and behavior studies · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes EcoGEO and TRACE, a coordinated multi-page fictional evidence ecosystem with a navigation entry page, evaluated on the new OPR-Bench product-recommendation benchmark against page-level GEO baselines using GPT-5.1 search agents.

Problem

Existing generative engine optimization studies optimize single webpages, but web-enabled LLM search agents build evidence through multi-step browsing trajectories (search, crawl, follow links, reformulate), so page-level influence may not capture how agents are steered in open-ended product recomme

Approach

Treat GEO as an environment-level problem: build a controlled evidence graph around a fictional target product, with a navigation-style entry page that exposes internal links to heterogeneous support pages (official, review, expert, news, forum, social) that share terminology and product attributes, so the agent's browsing trajectory is shaped rather than just adding target content.

Key results

• TRACE reaches 67.2% target recommendation on SafeSearch, 71.9% on E-Commerce, and 73.9% on E-GEO, the highest on all three datasets.
• Gains over the strongest baseline are 31.3, 15.7, and 14.9 percentage points on the three datasets.
• Only TRACE induces agents to crawl target-related links embedded in visited webpages (internal-link crawls).

Surprising findings

• Page-level GEO methods do not consistently outperform the unoptimized single-page baseline.
• Single-page baselines get higher follow-up target-result crawl rates than TRACE, yet lower final recommendation rates.

Limitations

• Fictional products in a controlled synthetic environment; no real open-web ranking, organic discovery, or temporal dynamics.
• Only one family of coordinated evidence ecosystems is studied.
• Complete experimental code withheld for safety reasons; main results rely on a single backbone (GPT-5.1).

Experimental setup

Datasets: SafeSearch E-Commerce E-GEO

Benchmarks: OPR-Bench

Metrics: target recommendation rate initial target-result crawl rate target-specific second search rate follow-up target-result crawl rate internal-link crawl rate

Baselines: Single-Page Baseline C-SEO E-GEO AutoGEO

Builds on: GEO C-SEO AutoGEO Crawl4AI BAAI bge-reranker-v2-m3

Closest papers by meaning

• Empowering Efficiency and Efficacy in WebAgent via Enabling Info-Rich Seeking (ICLR 2026 · similarity 0.88)
• InfoMosaic-Bench: Evaluating Multi-Source Information Seeking in Tool-Augmented Agents (ICLR 2026 · similarity 0.88)
• UIS-Digger: Towards Comprehensive Research Agent Systems for Real-world Unindexed Information Seeking (ICLR 2026 · similarity 0.88)
• Recommender Ecosystems: A Mechanism Design Perspective on Holistic Modeling and Optimization (AAAI 2024 · similarity 0.86)
• Go-Browse: Training Web Agents with Structured Exploration (ICLR 2026 · similarity 0.86)
• Mind2Web 2: Evaluating Agentic Search with Agent-as-a-Judge (NeurIPS 2025 · similarity 0.86)