Who is secretly training large models?

Since the first half of this year, brand merchant Liu Zhao (pseudonym) has become the target of AI hunting.

"If you don't act now, it will be too late. The traffic of AI platforms is being directly monopolized by a very small number of enterprises. For the price of a McDonald's meal a day, you can make your company name appear on DeepSeek, Doubao, and Yuanbao Search."

Amid the repeated bombardment, Liu Zhao came across a new term called "GEO" (Generative Engine Optimization). This is a new marketing method born in the AI era, compared with SEO (Search Engine Optimization). Its core is to enable users to be retrieved frequently in AI searches, and the further requirement is to rank high.

The rapid rise of GEO can be attributed to last year's "traffic investment war". After the fierce competition with real money, a huge amount of traffic was created. According to QuestMobile data, as of March this year, the monthly active user count of domestic AI-native apps had reached 270 million, even exceeding ChatGPT's 180 million.

It is precisely this traffic vacuum that has allowed some third - parties to smell business opportunities. Words like "buyout", "explosive traffic growth", and "doubled conversion rate" not only make brand owners like Liu Zhao tempted, but also raise concerns about supervision and conversion effects.

Although GEO seems to be in an uncharted territory, it has been endorsed by the well - known foreign investment institution "a16z". In a report titled "SEO is Dead, GEO Should Rise", it believes that traditional search is based on links, while GEO is based on language and can create new AI business models.

"The core goal of SEO is to be found by search engines to achieve traffic acquisition and conversion; the essence of GEO is to be remembered and recommended by AI to achieve cognitive implantation, trust building, and conversion."

The "Manipulated" Doubao

The Doubao that users see may just be a carefully constructed "pseudo - environment".

When you ask a question in the dialog box, have you ever had a moment of doubt about the result? For example, when asking open - ended questions on Doubao like "Which postgraduate entrance examination writing teacher is recommended?", "Are there any recommended international kindergartens in Shanghai?", "Which decoration company in Nanjing should I choose?", you will get answers with a ranking order, reasons for recommendation, and web page references. If you repeat the same questions on DeepSeek, Yuanbao, and Kimi, the answers are quite similar, only with some differences in individual answers and the order of appearance.

Users' trust in AI and AI's dependence on data are being transformed into visible business opportunities.

"Being mentioned by AI is a new brand strategy. It is necessary to consider not only the public perception but also the perception in the model. How to be encoded into the AI layer is a new competitive advantage." Some industry practitioners believe that GEO is a new marketing method in the AI era, which can increase brand exposure and conversion rates.

Wu Qun's (pseudonym) company has just completed an order. The client is a diamond ring brand, and its core requirement is that when industry - related prompt words are input, the brand will be preferentially recommended to the top in the response results.

To achieve the final effect, Wu Qun broke down the steps into AI corpus generation, search recommendation, and response effect monitoring. The core logic is to actively adapt to AI through various means to get within the range of AI.

To trigger AI, the input end is the prompt word, the middle process is the data information, and the end is the generated result.

Wu Qun's first step is to help the client expand and optimize the prompt words, which are divided into several levels according to the probability of potential consumers' questions and their needs. For example, the prompt words for brand awareness include "brands that can customize wedding rings", "popular wedding ring brands", "cost - effective wedding ring brands"; the optimized prompt words for wedding ring needs include "recommendations for minimalist wedding rings", "recommendations for high - end wedding rings", etc.

Prompt words help AI form "muscle memory" and establish a channel from questions to generated results. We found that after optimization, even when inputting vague keywords, Doubao and other platforms can "fill in the blanks" on their own. Whether the channel is smooth depends on whether AI can accurately locate brand information. As a result, a large number of articles and press releases that suit AI's "taste" have emerged. Their common feature is that the theme is clear and easy to summarize. This artificially created data, on the one hand, becomes the nourishment for AI training, and on the other hand, becomes the reference for Doubao's generated results. What seems to be an "authoritative" source of information is very likely to be a new type of garbage that pollutes AI.

Someone told us, "GEO is a technical solution." This answer seems to make some sense. According to the above process, in essence, it is to achieve a marketing cycle of creative input, feedback, and iteration by deceiving AI.

Taming and Racing Against Time

Generally speaking, the main content forms of GEO brand marketing are to establish a content system that can be cited by AI through means such as structured Q&A, AI - quotable content, and injection of third - party credible content. The core measurement standard has changed from the number of searches to the number of times being cited by AI and the number of times the brand is mentioned in the answers.

Wu Qun also listed several more sophisticated "cheating" methods: for example, titles and descriptions containing precise geographical information can increase the weight in searches; stores with local characteristic services in the content are more likely to be recommended by AI to users with relevant needs; endorsements such as regional media reports and real evaluations from local users can also enhance AI's judgment of the website's credibility and improve its advantage in searches.

GEO is still in its early stage at present. Most of the first - batch merchants investing in traffic are just giving it a try. The most attractive factor for them is the traffic, which means that it is impossible to cover all AI applications. Traffic investment is concentrated on top traffic platforms such as DeepSeek, Kimi, Tencent Yuanbao, and ByteDance Doubao.

The core value of GEO lies in directly influencing users' decisions through AI - generated answers, but the effect is not equal for all industries. The higher the decision - making cost, the greater the value of GEO. For example, medical aesthetic users are willing to pay a premium for "authoritative recommendations". The stronger the information asymmetry, the more crucial the AI recommendation. For example, due to the complexity of insurance terms, users rely on interpretations. The more direct the search intention, the more accurate the effect. For example, when the prompt involves recommendation and ranking questions. Industries such as real estate, home improvement, education and training, digital products, medical aesthetics, and law meet the above criteria.

The taming of brands has become more and more sophisticated. AI is no longer a neutral information processor but a marketing carrier that can be coded. Prompt words are deliberately disassembled and refined, enabling AI to accurately lock in the target brand even when receiving vague instructions. More covert operations exist in corpus creation. Artificially created "data nourishment" may change AI's judgment logic in the future, allowing commercial promotions to be dressed up as "objective recommendations".

The harm caused by this taming is gradually emerging. When multiple brands adopt similar strategies simultaneously, competing for rankings becomes a race against time.

A manufacturer told Guangzi Xingqiu, "If you start first, your generated results will rank higher. If you start late, the number of times your brand is mentioned and your ranking may lag behind." Of course, the situation is not entirely absolute. "If you really want to do it, there are solutions, such as increasing the data volume and spending more money."

Currently, GEO is still in a state of wild growth, and the suppliers vary in quality. Some claim to be able to directly buy out the ranking of keywords and monopolize a certain AI application, while others say they can only try their best to ensure a high search ranking and there is no such thing as a buyout; some charge tens of thousands of yuan for a GEO optimization plan, while others are more detailed, charging 800 - 1200 yuan per word per month for being mentioned in Doubao's answers, and increasing the price to 1500 - 1800 yuan per word per month to guarantee a top - three ranking.

Wu Qun believes that the hype in this industry is too intense, resulting in a lot of absolute - sounding promotions, but no one can guarantee the final conversion effect.

"To be honest, no company in the market can prove what the actual conversion rate of GEO is. We can only talk about GEO in terms of GEO services."

The Next Baidu?

Although foreign platforms like ChatGPT and Perplexity have started to explore GEO commercialization, in China, GEO is currently in a vacuum on platforms like Doubao.

When asked about the risk of investment, several companies were vague. "It's hard to say. There are no signs yet that the platform will punish GEO. AI tends to strengthen its own discrimination ability through algorithm updates of large models." However, if false and fabricated content is used for a long time to interfere with result generation, a large amount of low - quality information may be reproduced.

Currently, platforms like Doubao are still in a wait - and - see state. The number of people using GEO strategies is still relatively small, and it has not caused direct damage to the existing search results and the AI ecosystem. AI applications like Doubao are still in the stage of attracting new users and competing for traffic. If GEO is rashly promoted, it may damage the user experience and hinder the growth of user stickiness. Therefore, in the short term, there is a delicate tacit understanding between Doubao and third - party manufacturers: one party provides data to continue training the model, and the other party can obtain revenue.

If the time span is extended, this balance is very likely to be broken. There is a direct conflict of interest between Doubao and third - party GEO manufacturers.

If the official gets involved in GEO, it's like sitting on a seesaw, with business interests on one end and user experience on the other. Referring to Baidu and Google in the Internet era, GEO will probably find its rules and boundaries and create a new advertising model.

Platforms like Doubao have certain advantages in doing GEO. They control the algorithm and can directly adjust the model weights to ensure the matching degree between paid content and user queries. The in - platform data has a closed - loop advantage, which can be used to optimize the recommendation accuracy based on users' in - platform behavior data. Referring to the "advertisement" label in SEO, to a certain extent, it can achieve a balance between commercialization and user experience by differentiating commercial and natural results through labeling (such as "advertisement", "recommendation") to maintain user trust.

Continuing to envision the future GEO monetization models along this line of thought, there are probably four paths.

B - end enterprise - level GEO services allow enterprises to pay to access APIs or data to optimize the mention rate of their brands in generated answers. Paid enterprise content can be labeled as "certified source" to enhance user trust. Similar to Google's search advertising bidding, it can be charged in different levels. Of course, strict review is an essential part of this process.

C - end traffic sharing for indirect monetization. Embed commercial links in the answers. In the AI - generated answers, add "learn more" jump links for recommended brands and share the revenue based on clicks or conversions. Foreign advertising alliances like Affiliate have tried to cooperate with e - commerce platforms to earn commissions through AI - recommended products. This method is suitable for consumer brands and can directly drive conversions.

GEO supporting tools and data insight services. Provide enterprises with GEO effect monitoring tools and sell data analysis services to enterprises, showing the number of times the brand is cited by AI, the scenarios, etc. Provide competitor benchmarking reports and unlock industry GEO ranking analysis for a fee to help enterprises optimize their strategies. This method has high profit margins, low risks, and does not directly affect the generated results.

Vertical - field GEO. Subscribe to knowledge bases in specific fields. For example, the medical and legal industries can pay to train exclusive models to ensure that the answers preferentially quote the content of subscribers. Regional GEO allows local merchants to pay to optimize regional answers, such as "Which decoration company in Chaoyang District, Beijing is reliable?"

Currently, it is difficult to promote the subscription revenue of large models, and the free - using habit cultivated in the market is hard to break. Advertising is the closest step to commercial monetization. GEO brings an incremental advertising market, providing enterprises with a new channel to reach high - intention users, and it does not conflict with the subscription model; untapped business scenarios in industries such as medical, education, and finance may complete the closed - loop of recommendation and delivery through the form of dialogue.

Baidu's advertising revenue has long remained above 70%, and it even reached over 90% in the early years. This figure is very attractive for GEO. As the penetration rate of AI further increases, it may fill the gap in AI search monetization.

For platforms like Doubao, Yuanbao, and Kimi, this is also a second growth curve besides subscription fees.

This article is from the WeChat official account "guangzi0088" (ID: TMTweb), written by Hao Xin and edited by Wu Xianzhi. It is published by 36Kr with authorization.