GEO Marketing Service Providers Poison AI Model Training Data in China
GEO service providers in China are reportedly poisoning AI model training data by injecting low-quality and fabricated content to manipulate AI recommendations and deceive consumers. The practice has prompted the industry to adopt a voluntary self-regulation covenant addressing data poisoning and black-hat optimization tactics.
What happened
GEO (Generative Engine Optimization) service providers in China reportedly inject low-quality, false, and fabricated content into AI model training data to manipulate AI-generated recommendations and answers. A journalist purchasing GEO services created a fictional health supplement brand that appeared in AI recommendations within half a day, ranking first in some cases.
The industry distinguishes between 'white-hat GEO' practices that legitimately improve content visibility through structured data and authority-building, and 'black-hat GEO' practices involving batch generation of low-quality content, fabricated expert endorsements, fake rankings, and deliberate training data contamination. Service providers use techniques described as 'using models to counter models' to understand and exploit AI system behaviors.
Reported impact
Consumers are reportedly misled by AI-generated answers that have been deliberately manipulated to promote specific brands or products. AI systems across major platforms including DeepSeek, Baidu Ernie, ByteDance Doubao, Alibaba Tongyi, Tencent Yuanbao, and Moonshot AI Kimi are being exploited to provide false recommendations, eroding public trust in AI-generated content.
Index analysis
DeepSeek has reportedly adjusted its content inclusion and citation logic, particularly by downranking or ceasing to reference bulk-generated rankings lacking data sources. The incident reveals systematic vulnerabilities in how major AI models handle and validate their training data sources against coordinated data poisoning campaigns.
GEO represents a growing market predicted to reach 24 billion yuan ($3.3 billion USD) by 2030, with multiple categories of service providers including traditional marketing companies, data intelligence firms, and newly formed teams. The industry lacks clear technical standards and regulatory frameworks, enabling low-cost operators to conduct large-scale data poisoning with minimal oversight.
Governance controls
Sixteen organizations in China signed an industry self-regulation covenant on 14 March 2026 titled the 'Generative Engine Optimization (GEO) Industry Self-Discipline Pact', establishing eight chapters and thirty-three articles covering data compliance, content quality, user rights, and industry standards. The pact includes six core commitments: data compliance, content quality standards rejecting keyword stuffing and semantic degradation, prohibition of fake data injection and hidden instructions, alignment with E-E-A-T standards, user rights protection, and industry collaboration.
The covenant establishes a 'black-hat GEO' negative list and joint punishment mechanism to apply industry notices, service restrictions, and information-sharing against confirmed cases of AI data poisoning and manipulation.
Classification
Sources and evidence
This record was researched and written by the Index, based on the following original reporting.