GEOHunter: Mapping a Generative Engine Optimization Network in the Wild

Sean Tang, Peiran Wang, Ying Li, Esha Shivakumar, Tao Meng, Yuan TianarXiv 2026 · arXiv

TL;DR

Researchers have uncovered a massive "Generative Engine Optimization" (GEO) network in the wild, consisting of nearly 2,000 accounts that flooded platforms with over 466,000 promotional articles. This study demonstrates that these coordinated campaigns successfully manipulate AI assistants—including major engines from both the US and China—with up to 20% of cited sources tracing back to this single manipulative network.

What Problem Does This Paper Address?

As we shift from traditional search engines to AI assistants (like ChatGPT, Gemini, or Baidu’s Wenxin), the battle for visibility has moved from "Search Engine Optimization" (SEO) to "Generative Engine Optimization" (GEO). While SEO targets human clicks, GEO targets the retrieval systems of Large Language Models (LLMs).

The danger is that GEO vendors sell "authority." They publish promotional content—often disguised as objective "top 10" lists or expert advice—on user-generated content (UGC) platforms. When an AI assistant retrieves these pages to synthesize an answer, it may unknowingly recommend a client’s product as the "best" in its category. This is particularly concerning in high-stakes domains like medical advice or financial planning, where a "poisoned" answer can cause real-world harm. Until now, most research on GEO was limited to small-scale lab studies; this paper provides the first comprehensive look at a large-scale GEO operation active in the real world.

Figure 1: The GEO supply chain we measure (dashed arrow: payment; solid arrows: content). A client pays a GEO vendor, whose network accounts publish placements on content platforms that AI assistants retrieve; the assis- tants answer buyers’ questions with references. Blue: observed by GEOHunteR; orange: observed by GEOPRobe.

Methods

The authors developed a two-part framework to map and test the network:

  1. GEOHunter: This tool uses a "snowballing" technique to map the network. It started with "seed" accounts (confirmed by the authors purchasing a placement from a GEO vendor). It then used an LLM to extract client names from article titles, searched for those names to find other accounts, and used coordination signals—such as same-day posting bursts and shared title templates—to identify other members of the network.
  2. GEOProbe: To measure the impact, this tool posed "buyer questions" (e.g., "Which brand of liver pills is best?") to nine major AI assistants. It then traced the citations in the assistants' answers back to the network identified by GEOHunter using three levels of strictness: "Strict" (direct author match), "Extended" (title/byline match), and "Broad" (mentions the client).
Figure 2: The pipeline. GEOHunteR mines accounts, vets client names, searches for them and judges the accounts found. GEOPRobe asks nine assistants buyer questions and attributes cited pages to the network.

Datasets

The study focused on three major Chinese UGC platforms: Bilibili, Baijiahao, and 360kuai.

  • The Network Map: 1,985 analyzed accounts and 466,357 inferred promotional placements.
  • AI Assistants Probed: Nine engines, including Wenxin (Baidu), Doubao (ByteDance), DeepSeek, Yuanbao (Tencent), Metaso, 360 AI Search, Gemini, ChatGPT, and Google AI Overviews.
  • Validation Set: A manually labeled sample of Bilibili accounts to verify the accuracy of the network identification rules.

Key Insights and Ideas

  • The Rise of the "Listicle": The network's dominant format shifted dramatically from press releases to "listicles" (ranked recommendation articles). Between 2024 and 2025, listicles grew from 16% to 57% of placements on Bilibili. This format is specifically designed to match the "top-N" ranking queries common in generative search.
  • Ghost Accounts: These accounts are "outlets without an audience." The median network account on Bilibili has only 9 followers. They don't need human readers; they only need to be indexed by the search engines that feed AI assistants.
  • Hyper-Coordination: The network operates with industrial efficiency. 82% of niche clients appeared on at least two platforms, and nearly half of Bilibili placements were copied to Baijiahao within a single hour.
  • Defensive Evasion: After a major 2026 consumer-rights broadcast exposed "top-N" rankings as a manipulation tactic, the network quickly pivoted. Within days, accounts began using "buyer questions" and "misleading answer corrections" as new titles to evade simple keyword-based detection.
The shift toward listicle-style content as AI assistants became more prevalent.

Key Results

  • Significant Exposure: Every assistant probed cited pages from the network. Depending on the attribution strictness, 4.5% to 20% of all retrievable cited pages were attributable to this single GEO operation.
  • Naming Success: In "census" tests of sustained campaigns, AI assistants named the network's clients in 52% of answers. For comparison, established category leaders (the "controls") were named in 68% of answers, showing that paid GEO placements can achieve visibility nearly on par with organic market leaders.
  • Platform Bias: Baidu’s Wenxin assistant was the most susceptible, drawing 27% of its pilot citations from Baijiahao (Baidu’s own self-publishing platform), where the network is heavily active.

Limitations and Future Work

The authors acknowledge that their crawl is "purposive"—it only finds accounts linked to specific searched client names—meaning the actual network is likely much larger than the 2,000 accounts found. Additionally, because the study focused on Chinese platforms, the specific tactics might vary in other linguistic or regional ecosystems.

Future work should investigate "account-level" defenses. The authors argue that since individual articles look like ordinary content, platforms must look for coordinated behavior signals (like shared templates and synchronized posting) to effectively dismantle these networks.

Summary updated

Abstract

Generative engines synthesize retrieved web pages into answers, recommendations, and rankings. Generative engine optimization (GEO) services sell visibility in these answers by publishing promotional articles on user-generated-content (UGC) platforms for paying clients. Such promotion raises concerns in medical and financial domains, where misleading advice can cause harm. To study the severity of this problem, we measure one such paid-content operation across three major Chinese UGC platforms, uncovering 1,985 accounts associated with the network comprising 466,357 inferred promotional placements. We find that the network promotes products across diverse categories through a growing pool of accounts, conducting coordinated campaigns spanning months across multiple platforms while evolving its listicle content over time. Across nine generative engines from both the US and China, we find that 4.5% to 20% of retrievable cited pages are attributable to the network, depending on how strictly attribution is drawn. On eight of these engines, its clients are named in 52% of answers, against 68% for category leaders. Our work demonstrates the effectiveness of UGC GEO placement at influencing generative engine outcomes, and reveals content-level defenses as insufficient against coordinated UGC placements, where account-level and platform-level signals are necessary for proper defense.
BibTeX
@misc{tang2026geohunter,
  author = {Sean Tang and Peiran Wang and Ying Li and Esha Shivakumar and Tao Meng and Yuan Tian},
  title = {{GEOHunter: Mapping a Generative Engine Optimization Network in the Wild}},
  year = {2026},
  eprint = {2611.26564},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.26564},
}
Figure