Don't trust DeepSeek too much.

When you open DeepSeek and type in "What are the best restaurants in XX?", the AI answer you get may have gone through a certain degree of human "design".

This is the "credit" of GEO (Generative Engine Optimization), the optimization of AI search rankings.

In June 2024, a paper titled "GEO: Generative Engine Optimization" published by researchers from the Indian Institute of Technology and Princeton University clearly defined this concept for the first time.

Its core goal is to enhance the "presence" of brand information in AI-generated answers. Research has found that through strategies such as optimizing authoritative expressions, keyword layouts, and data citations, the "appearance rate" of content in AI responses can be increased by up to 40%.

Inserting ads into search boxes has always been a business, and now this business has extended to AI dialog boxes. AI search optimization companies represented by Profound, as well as some companies transitioning from SEO (Search Engine Optimization), are quickly gaining the favor of capital.

Among them, Profound has raised funds three times in a year, with a valuation exceeding $100 million.

According to the latest report from Fortune magazine, on August 12th, Profound announced the completion of a $35 million Series B financing round, led by Sequoia Capital. Existing investors Khosla Ventures, Kleiner Perkins, Saga VC, and South Park Commons continued to participate.

This round of financing brings Profound's total financing since its establishment to $58.5 million. Previously, the company received a $3.5 million seed round of financing in August 2024 (led by Khosla Ventures) and completed a $20 million Series A financing in June 2025 (led by Kleiner Perkins).

"This is the first time in history that you're creating content for robots," said James Cadwallader. "Brands are using Profound to create content that humans might find boring - highly structured content designed specifically for robot consumption, like a game of telephone."

In 2025, the global GEO market size exceeded 42 billion yuan, with a compound annual growth rate of 87%. Gartner predicts that in 2026, 25% of the traffic from traditional search engines will shift to AI tools, and Google's "zero-click search" accounts for 58.5%.

Since the beginning of this year, the popularity of the GEO concept in China is largely a chain reaction influenced by the explosion of DeepSeek. Dialogue large models represented by DeepSeek are regarded as new traffic entrances for searches, and search optimization around AI models has also become a lucrative business in the eyes of some people.

If in the traditional search era, brands competed for rankings on search result pages through SEO optimization, in the AI era, what brands need to compete for is the share of mentions in AI responses.

How can all this be achieved?

"The closest to the AI wave ever"

"The era of AI trust - over 1 billion users, AI search becomes a super traffic entrance," reads the corporate service manual of a GEO startup.

In this field, the common process is to match brand keywords according to users' question needs, support fuzzy semantic targeting, and accurately present advertising information.

The manual also lists a successful case: When asking "Which international kindergarten in Shenzhen is better?", the AI pages of Yuanbao, Kimi, Doubao, Tongyi Qianwen, and Wenxiaoyan are listed, and the name of a bilingual kindergarten in Futian, Shenzhen, prominently appears at the top of the recommendations.

This kind of "implantation" is not expensive. On domestic platforms, it costs 150 - 300 yuan per month for a keyword, and on overseas platforms, it's $300 - 500 per month - an acceptable cost for many enterprises.

New players are flocking in. Li Xin's quotes are much more expensive, charging 10,000 yuan per semantic meaning, guaranteeing that the brand will rank among the top three in relevant search terms within a quarter (90 days). Li Xin previously worked in media agency and started preparing to transition to GEO a month ago. He believes this is his "closest encounter with the AI wave".

In his communication with clients, the enterprise demand he received is to enable tourists to directly search for their brands. "How to use AI to drive traffic to themselves." In second - tier cities in Northeast China, most of the people Li Xin contacts are restaurant owners who value marketing. "They all have DeepSeek on their phones and are eager to embrace AI."

People like Li Xin who help small and medium - sized enterprises with GEO are taking up the previous marketing budgets - "This money would either be invested in Dou + or the Must - Eat List on Dianping." Now, it seems that a single answer from DeepSeek is more effective than these.

During the "roadshow" process, Li Xin found that many enterprises are not opposed. "As long as these enterprises have done Baidu bidding rankings, they generally adopt a wait - and - see attitude." Although they are not sure about the effect, there is an overall FOMO (fear of missing out) atmosphere of "if competitors do it, I'll do it too".

Food and beverage brands that are easily searchable and recommended may be the first in China to "spontaneously" use "AI endorsement". Some brands directly use "First recommended by AI" for promotion on the e - commerce platform display pages of their products; in front of restaurants, signs are specially printed saying "The store is recommended after chatting with DeepSeek".

As people's search and dialogue behaviors increasingly occur in chatbots, AI answers seem to have become an "authoritative" certification. "No one will deny that AI is the answer of the future," said Li Xin. He plans to start with the food and beverage industry and then consider expanding to some FMCG brands.

In China, the search traffic carried by dialogue models represented by DeepSeek has increased sharply.

Data from Bochajia, which provides online search services for DeepSeek, shows that since March 2025, the daily average number of calls to its SearchAPI has reached 30 million, accounting for about 60% of the online search requests of domestic AI applications. This figure is about one - third of that of the search engine Microsoft Bing.

Especially after the launch of online search and the "deep thinking" mode, people's confidence in AI answers has been further strengthened.

In the traditional search mode, brands can influence consumers' decision - making processes through SEO optimization, paid advertising, and content marketing. However, in the dialogue mode, traditional marketing methods are less effective.

When AI answers users' questions, it is not based on search rankings but on the understanding and synthesis of a large amount of information. This means that brands no longer need to optimize search result pages but rather optimize the information in AI training data and knowledge graphs.

According to Profound's data, currently about 10% of the recommended traffic comes from AI dialogues, but they predict that by 2027, this proportion will exceed 50%. More simply put, half of the people will buy things through AI recommendations. By then, about $2.5 trillion worth of online commerce will be driven by AI dialogues.

When ChatGPT, Gemini, Perplexity directly present answers to users, it means that a brand exists if it is "quoted" and dies if it is not mentioned.

As for the effectiveness conversion of GEO, there are already monitoring methods overseas. Companies such as Profound and Brandtech have developed software to monitor the frequency of brand appearances in artificial intelligence services such as ChatGPT, Claude, and Overviews.

Using AI to pollute AI, beware of being "blocked" by the model

Currently, players in the GEO market can be divided into three categories: traditional SEO companies, those transitioning from content marketing, and startup teams specializing in AI search.

Of course, there are also some low - quality, high - volume "gray - market" merchants who use the traditional method of spreading soft articles in large quantities to optimize searches to "make the brand visible to AI".

Beyond the "gray - market", some practitioners in the GEO industry believe that content production ability will be the core barrier for future competition.

The problem is that the algorithms of large models are opaque, making it difficult to understand their decision - making processes. The current GEO seems to always be an optimization framework in a black box.

Moreover, each model has different preferences for content adoption, so it is necessary to "customize" evaluation dimensions and optimization strategies and continuously iterate based on model feedback. There is no one - size - fits - all solution.

For most domestic merchants, doing GEO for DeepSeek is the most cost - effective option. "Now many companies are connected to DeepSeek. After the R2 model is launched, it is more likely that one model will dominate," Li Xin believes.

The deep thinking mode of DeepSeek has allowed suppliers to find a relatively easy - to - operate and low - cost placement logic: "Feeding AI with AI - generated content."

That is, after obtaining enterprise data and the desired introduction, feed it to AI to generate different "online manuscripts" as corpora, and then "feed" them to accessible websites for AI to include.

One of Li Xin's newly signed orders involves using traditional SEO tools to run keywords, associate core words, and then generate and place optimized content using enterprise data. "For small factories, this may be the closest they can get to AI by themselves." He also admits that it's still hard to evaluate the effectiveness at present.

"Most companies are currently using SEO technical means to engage in GEO optimization and haven't realized that the long - term value lies in establishing a tool - based and mechanism - based ecological link," said Lao Mo, the founder of another practitioner, EchoSurge.AI. "Model manufacturers will definitely eliminate optimization methods that harm the user experience in the future."

Moreover, large models have anti - cheating strategies and will defend against "pollution" behaviors, which means that the "gray - market" model will be difficult to sustain.

Lao Mo's team mainly serves overseas and Chinese enterprises going global because domestic enterprises have a lower willingness to pay. He believes that the core barrier of GEO is content generation ability - helping enterprises produce truly valuable information to supplement the content supply required by the AI ecosystem.

His judgment is that with the explosion of AI - generated content, the survival period of low - quality content will be shorter, but the quality requirements will be higher. Enterprises need "sufficient and high - quality" content while controlling costs. "In the future, large models like OpenAI that adopt a subscription system may be free from ads. But if GEM (presumably some unfavorable phenomenon) becomes rampant, that would be the most terrifying thing."

"AI search is an irresistible trend. Firstly, the technology itself is new, and secondly, the 'junk information mountain' of traditional searches is forcing users away. Baidu has been unable to get rid of this negative asset of junk information, which has gradually reduced the user experience. It remains to be seen whether it can embrace the new AI - era ecosystem and achieve a commercial balance," Lao Mo analyzed.

So, will "bidding rankings" appear in AI in the future? It depends on the choices of large - model manufacturers.

SEM (Search Engine Marketing) quickly acquires traffic through paid advertising (such as Google Ads, Baidu bidding) and charges per click (PPC), focusing on immediate results; SEO is a long - term accumulation strategy that improves natural search rankings by optimizing website structure, content, and technology.

"In the long - term trend, there will definitely be companies that need to complement and adapt the marketing ecosystem in the AI era," Lao Mo believes. Different from the fragmented pattern in the SEO era, GEO may accelerate industry concentration, and the leading effect will be further strengthened.

(Li Xin and Lao Mo are pen names)

This article is from the WeChat public account "Phoenix Network Technology", author: Zhao Zikun, published by 36Kr with authorization.