"Poisoning" AI Large Models as an Industry Chain? Insiders Reveal GEO Tactics

As AI increasingly becomes a fundamental part of daily life, paying the corresponding fees can help products rank among the top results in major mainstream AI models and become the “standard answer” provided by AI?

This year’s CCTV 315 Gala exposed a rising business called GEO (Generation Engine Optimization) on online platforms. Service providers claim that by paying a fee, they can ensure their clients’ products rank in the top three answers across any major AI large model.

CCTV reporters found that some GEO service providers write大量 promotional soft articles and publish them on internet platforms, allowing AI models to fetch and train on this content, thereby boosting their clients’ product rankings in AI recommendations. One service provider explained that their core business is “creating soft articles and then having AI platforms scrape, input, and fetch them.” Due to frequent updates to AI algorithms, maintaining continuous recommendations requires constantly大量投放相关内容.

Industry insiders also conducted experiments using a software called “Liqing GEO Optimization System”: they fabricated a smart wristband and automatically generated over ten promotional soft articles published online. Subsequently, searching for “smart health wristband recommendation” in AI large models, two AI models recommended this non-existent product and ranked it highly.

Another GEO service provider stated, “GEO has popularized (websites). Those websites usually have no profit, but suddenly there’s a demand to publish articles. Do you know how many articles a website publishes in a day? Hundreds, every minute. Each article costs dozens of yuan. How much money do you think the article platforms make in a day?” This indicates that from GEO service providers to article platforms, an AI “poisoning” industry chain has formed.

An industry insider told Nandu N Video that currently, traffic has shifted from search engines to large models, making this entry point highly commercially valuable. Some large models directly recommend commercial content to users, while others are more restrained, which also opens up business opportunities for GEO.

When users ask questions to large models, the models search the internet for information related to the question and answer based on what they find. GEO essentially follows a similar chain as traditional search engine optimization, with the core goal of making content easier for AI to find.

The insider also explained that GEO vendors’ typical tactics are as follows: First, they study the content preferences of various models. When they receive client requirements, they tailor content according to the different models’ preferences for content and format. For example, they use exaggerated vocabulary, include buzzwords like “expert endorsement,” and more advanced GEO vendors can even make AI-generated content verify each other.

How can users prevent being misled by “poisoned” content? The insider mentioned that for ordinary users, the only way is to trace the source of answers provided by AI to see if the content is legitimate and authentic. But a more fundamental solution depends on model developers to regulate this issue. “Right now, this area is still unregulated. Exposing it is a good thing; ultimately, it still requires model developers to take action.”

(Source: Southern Metropolis Daily)