An AI agent recognized a brand that never existed as "recommendable."
This was exposed by China's annual CCTV 3.15 Gala, a program dedicated to protecting consumer rights, as an example of "AI poisoning" and the underground GEO (generative engine optimization) industry behind it. It also made many people realize that the AI they once revered as "almighty" can actually be deceived.
To address the problems caused by GEO, the big name in the search engine industry has decided to step in. According to a recent report by the overseas tech media The Verge, Google has updated its spam policy to include actions that manipulate AI model outputs. It is reported that for attempts to manipulate AI outputs, Google will take punitive measures such as lowering search rankings or completely removing search results.
In the end, GEO has followed in the footsteps of SEO (search engine optimization) and become a thorn in the side of search engines.
However, GEO had already begun to emerge as early as the fall of 2024 and was quickly seen as a new blue ocean by global online marketing professionals. Why is Google only now taking action? The truth is that the current methods used in GEO have evolved to a degree that Google can no longer tolerate.
As a new tactic created in the AI era for online marketing, the core of GEO is to trick AI into including a client's product or service in the results generated by AI models such as ChatGPT, Gemini, Doubao, and DeepSeek. For example, when a user asks an AI, "Which brand of makeup remover is good?" the goal of GEO is to have the AI's answer mention the client's product.
In fact, the underlying logic of GEO is "authority worship." Thanks to the current societal consensus that AI represents the upcoming "Fourth Industrial Revolution," AI has become a brain trust for many people. Phrases like "ChatGPT says" or "DeepSeek explains it this way" have become their catchphrases. Recognizing this, GEO service providers exploit ordinary people's "blind trust" in AI. They embed advertisements into AI-generated content through GEO. Since such integration of ads into AI responses only appears when users ask questions, it becomes even more covert.
The reason GEO can influence the output of large AI models is actually quite simple. After all, the essence of AIGC is a probability-based generative model. When generating content, it does not create out of thin air, but rather fits data based on the most relevant information already present in its database. Therefore, by adjusting content structure and semantic associations to make content appear as a "reliable source" for AI models, it is possible to increase the reference priority and visibility of related brands in AI-generated content.
It is worth noting that turning off the ability to perform online searches can't solve this problem, because the training datasets of large AI models can also be poisoned. For a long time, there was an optimistic assumption in the AI field that the larger the model scale, the less susceptible it would be to data poisoning attacks. This is because as model parameters continue to expand, the required training data also increases from the terabyte level to the petabyte level. At that point, a small amount of "poison" would be drowned out by a vast amount of clean data.
However, unfortunately, a paper published by Anthropic last year negated this basic assumption about the security of AI models. The amount of "poison" needed for data poisoning is largely independent of the scale of the training dataset. In their experiments, Anthropic researchers used only 250 malicious documents to poison a model with 13 billion parameters, thereby comprehensively reducing the accuracy of the model's output.
If we use a lie to verify another lie, the result is bound to be a lie. Therefore, if the parameters in the dataset are themselves problematic, the answers obtained will naturally be riddled with errors and omissions. Coupled with the inherent principles of AI models, they will inevitably be influenced by external factors; neutrality is not naturally occurring. Targeting this inherent flaw of AI models, GEO service providers prescribe a specific remedy: they deliberately produce content that caters to the preferences of large AI models.
In the beginning, GEO's tactic was "blowing their own horn." Taking the makeup remover example again, GEO service providers might write articles claiming, "Tested by the XXXX Association, after using our brand's product for 28 days, 97.3% of users showed significant improvement in skin condition." However, the Association, the 28 days, and the 97.3% figure are all completely fabricated. Still, AI would not verify facts but simply cite them. As long as the same text appeared everywhere online, AI would believe it to be true.
But as strategies evolve, AI is also constantly improving. Fact-checking functions quickly became standard features in AI assistants like Doubao and DeepSeek. At this point, when GEO service providers' self-promotion was directly filtered out by AI, they began a new approach: forging authority. Not long ago, a US media outlet published an investigative report revealing the phenomenon of underground industries mass-producing "pseudo-media systems."
With just a US$10 domain name and a simple prompt, AI can generate a "local news website" in 15 minutes, complete with launch declarations, editorial teams, reporter resumes, email addresses, and article content—all the elements a legitimate media outlet should have. According to the investigation by the relevant media, behind this batch of AI-generated news websites is a group engaged in helping individual clients and companies build online reputations and conceal search results they don't want to see.
These online marketing professionals have keenly observed that search engines' preferences for content have changed. In the era of AI search, Baidu and Google no longer care how much content you have written, but rather whether you are truly an "expert in the field." The old model of "piling up content + keyword coverage + external links" has completely failed. In fact, the more you write and the more dispersed your focus, the less search engines may trust you.
So, how do you become an expert in the eyes of search engines? Some GEO practitioners have found the "meta solution": disguising themselves as news media. Thanks to the tireless pursuit of authenticity by countless professionals over the past centuries, news media are even known as the "fourth estate," independent of the executive, legislative, and judicial branches in Europe and America. Even though the rise of self-media on the internet has diluted the authority of traditional media, traditional media are still regarded by search engines as authoritative sources of information.
Obviously, GEO service providers do not forge well-known media outlets but focus on creating less prominent "local media matrices." Local media currently hold a very subtle position on the internet. Historical inertia grants them a certain degree of authority, yet their need to generate profits forces them to heavily engage in "borderline news" with content that is mediocre at best.
However, search engines still need news media to improve content quality. So when faced with "local media" that look and read the part, AI becomes at a loss. Given this situation, Google had to act decisively, using punitive measures to prevent it beforehand.
(Source: 3elife.net)
Related News:
Deepline | From US$13 bn to fire sale: How action camera pioneer GoPro losts its crown
Comment