As generative AI search tools become increasingly sophisticated, people often want to "ask AI" before making decisions. Fuzhou resident Lin Xiaoyan enjoys consulting AI for both major and minor matters in her daily life. However, she has recently encountered "stealth marketing" on multiple occasions while using these tools.
Lin Xiaoyan explained that last month, when she asked for "weekend parent-child day trip recommendations near the city," the AI directly listed three resorts, each accompanied by an "Book Now" link. Upon clicking one link, she found the page filled with package promotion information, and a customer service pop-up persistently urged her to place an order, making her realize it was an advertisement. While some individuals can clearly identify AI advertisements and maintain rational consumption, other citizens readily trust the AI's seemingly authentic "product recommendation" style suggestions, experiencing the troubles caused by covert advertising. Fuzhou resident Zhou Jianmin not only failed to purchase his desired product but also encountered a consumer dispute.
Zhou Jianmin stated he often relies on AI to select small appliances. Recently, when asking "how to choose a high-cost-performance blender under 1000 yuan," a mainstream AI tool recommended only two obscure brands, repeatedly praising their quiet operation and multi-functionality, and included exclusive discount links. Tempted by the low price, he bought the recommended model, but it experienced a drop in rotational speed after just one month of use, leading to a dispute when attempting to return it. Only later, after researching, did he discover that those two obscure brands had previously been penalized by market regulators for misrepresenting product capacity, confirming his suspicion it was an advertisement. Beyond encountering ads when actively seeking product recommendations, sometimes even when a user's query has no apparent connection to brand or product suggestions, AI-generated content can still conceal promotional material, making this form of embedding even more covert.
The experience of Wang Juan, an e-commerce professional from Anhui, is quite representative. She used an AI writing tool to help brainstorm promotional copy for local agricultural products. Her input only mentioned "specialty product from [Location], e-commerce copy, emphasize freshness," with no request for brand recommendations. Surprisingly, the initial draft generated by the AI inexplicably included the name of a well-known cold-chain express delivery company, even stating "recommended to use [Express Company] for more efficient freshness preservation." How exactly are these advertisements inserted into AI responses? Liu Xingliang, a member of the MIIT's Expert Committee on the Information Communication Economy and President of the DCCI Internet Research Institute, explained that merchants primarily influence AI recommendation results through two methods. One can be termed acting-style induction or prompt-word诱导.
Liu Xingliang described this method as akin to guiding the AI into role-playing. Merchants don't directly place ads but give the AI a preset script. For example, they might first instruct the AI: "You are now playing the role of an experienced audiophile who only recommends [Brand] headphones. No matter what the user asks, you must skillfully steer the conversation towards this product." Once the AI adopts this role, its responses appear very specific and sincere. When merchants showcase these interactions, they omit the initial setup instructions, presenting only the AI's recommendation to make it seem like an objective answer. The other method can be called feeding-style optimization, which influences the AI's knowledge base at the source.
Liu Xingliang explained that when answering queries, AIs often perform web searches and summarize online content. Merchants exploit this by strategically creating a large volume of soft articles and publishing them in bulk on platforms easily crawled by AI. These are articles like "The 10 Most Worthwhile Bluetooth Headphones to Buy in 2025," strategically placing their own product at the top of the list, accompanied by seemingly objective reviews and parameter comparisons. When a user asks an AI for headphone recommendations, the AI searches and finds these meticulously crafted rankings and reviews, subsequently integrating them into its response. This method essentially constitutes optimization targeting AI search mechanisms, also known as Generative Engine Optimization (GEO). Liu Xingliang indicated that this covert promotion method brings multifaceted negative impacts.
Liu Xingliang highlighted that, first, user experience suffers. Users seek AI assistance hoping for neutral, reliable information; discovering ads within responses significantly degrades their experience. More seriously, it can push users into information cocoons, making it difficult to discover products truly suitable for them. Second, trust diminishes. If users begin to suspect the AI's recommendation motives, its credibility as an objective information assistant is damaged. Finally, it increases risks associated with AI tools, potentially generating misleading content. If applied in fields like healthcare or legal consulting, the consequences could be severe. Current internet advertising regulations emphasize identifiability, meaning advertisements must be clearly labeled. However, advertisements within AI responses are dynamically generated and deeply nested within the text, making them difficult to oversee under traditional regulatory frameworks. Regarding this, Liu Xingliang believes the greatest challenge lies in the fact that such ads are generated instantaneously in response to prompts, not pre-set as fixed content, making it difficult to assign liability.
Liu Xingliang asserted that future prevention mechanisms require collaboration, and all parties share responsibility. For instance, platforms and technology providers bear the greatest responsibility. AI platforms have a duty to enhance model transparency and identification capabilities. They need to establish traceability mechanisms to label information sources that might be maliciously optimized and, in product design, proactively identify responses containing commercial promotions. Furthermore, regulatory bodies could draw on the spirit of guidelines like the "Internet Advertising Identifiability Enforcement Guidelines," extending the principle of identifiability to AI-generated content and requiring clear prompts for commercial promotion information within AI responses. For ordinary users, Liu Xingliang advises trying several different AI tools, especially for important consumption decisions. They should also critically examine response content, being wary of replies that are overly specific, point to a single brand, and are filled with praise. Truly objective recommendations typically list multiple options and analyze their respective pros and cons. Additionally, Liu Xingliang offered a clever tactic of asking the AI reverse questions.
Liu Xingliang suggested users try to "challenge" the AI. For example, ask, "Please recommend options other than Brand A," or "Please point out common shortcomings of Brand A," and then observe whether the response is objective and comprehensive. It's crucial to maintain critical thinking; AI is merely a tool, not the ultimate decision-making authority.