Source | Kong Shou The internet has brought an information explosion, and AI and algorithms have further changed the way humans access information. Looking back at every major change in advertising and marketing history, almost all have been brought about by information revolutions. The goal of brand marketing, in the final analysis, is to achieve efficient connections between products and consumers. To achieve this connection, one is to use media to convey product information, allowing consumers to understand, recognize, and identify with the brand; the other is to use channels to distribute and sell products, making it convenient for consumers to purchase. With the development of the times, information dissemination is becoming more important than channel reach. Don E. Schultz, a professor at Northwestern University and the father of integrated marketing communications, strongly emphasized the importance of integrating the "information flow." He strongly advocated that consumer communication is the entirety of marketing, believing that "the information that exists in the consumer's mind is the true marketing value because all other marketing variables, such as product design, pricing, channels, and promotion, can be imitated or copied by competitors; information dissemination is the only sustainable competitive advantage for an organization." But now, the amount of information we obtain in a day is equivalent to what a person in Shakespeare's time obtained in a lifetime. Faced with massive amounts of information and users, a fundamental issue has emerged: How can we connect people and information more efficiently? How can we help specific users quickly and accurately find the information they are interested in? In the PC era, the mainstream modes of information distribution were portals and search engines. Portals filtered information through website editors, categorizing, organizing, and sorting it before pushing it to users; search engines relied on users to actively search for information based on keywords. In the mobile era, social networks and recommendation algorithms became the mainstream modes. Social networks refer to users obtaining information by following friends and subscribing to blogger accounts, and then obtaining more information through friend recommendations, reposts, and other behaviors; Recommendation algorithms refer to systems that proactively discover user interests through intelligent methods and push information based on that. The emergence of algorithms has brought about a significant improvement in the efficiency of connecting people and information, becoming the mainstream of major social platforms, and this will also bring about a new round of changes in brand marketing. Why is labeling significant for brands? First, labels are the foundation of algorithms and the infrastructure of the big data era. Initially, the basic principle of recommendation algorithms was the "label mechanism." On one end of the algorithm are people, and on the other end is information (content, products, other people). Common algorithms find similarities among people and similarities among content, label each user and each piece of information, and then through data computation, push information with corresponding labels to people with corresponding attributes. Label-matched people and information will meet under the system's recommendation; this is the algorithm. In layman's terms, this is "birds of a feather flock together." With the development of machine learning technology, algorithms are also rapidly iterating and changing. Now, algorithms can directly predict user behavior without understanding content types or real-world semantics, that is, predicting users' interest and preference for information. On March 30 and April 2, 2025, Douyin and Video Account respectively announced their short video algorithm mechanisms. Douyin mainly relies on learning from users' behavior of watching content, using neural network calculations to predict users' next actions; Video Account emphasizes the role of social relationship chains, using friend recommendations to filter high-quality content and enrich users' information sources. However, to improve algorithm efficiency and break the "information cocoon," labels remain important. Generally speaking, for a video content to be selected from massive information and pushed to users, it must go through three stages: content pool, recall, and ranking. First, a newly published video must undergo preliminary evaluation to enter the content pool; next, it enters the recall stage, which is to quickly rough-screen candidate video sets (from millions to hundreds) that users might be interested in from the massive content in the content pool. Finally, in the ranking stage, the algorithm scores candidate videos for each specific user and pushes the highest-scoring video to that user (i.e., the video with the highest probability of target behaviors such as likes, comments, shares, follows, etc.). The ranking stage updates in real-time at the minute level, thereby more accurately predicting user behavior. Since the recall stage requires processing large amounts of data, to improve computational speed and recall rate, the industry typically uses simpler algorithm models and adopts a "multi-channel recall strategy" that stacks multiple simple algorithms, commonly including collaborative filtering, interest tags, hot news, recent trends, friend preferences, etc. , thereby pushing more different types of videos to users and guiding the algorithm to break the information cocoon. In the ranking stage, the content to be processed is smaller, so more complex ranking models can be used to obtain precise ranking results. Take Douyin's most commonly used "Two-Tower Retrieval Model" as an example. The two towers consist of a user tower and a content tower. The user tower learns users' interest preferences to form user features; the content tower extracts features from video content information to form content features. These two towers are essentially labeling users and content with unique "digital labels." Then, in the deep learning model, they are encoded as points (vectors) in digital space based on the distance between different content and users. This process is vectorized representation learning. The algorithm thus no longer needs to understand the real-world semantics of video content; it only needs to process pure sets of numbers to predict a user's behavior toward a certain content. For an algorithm to work perfectly, the recommendation model is important, but it is not the entirety of the algorithm. The strategy of the recommendation system's recall layer is also important, and another very important aspect is how the recommendation model selects and processes user and information features; this is the feature engineering of the recommendation system. Features are labels; their essence is the abstract expression of relevant information in the process of users obtaining information. For recommendation algorithms, user browsing history, likes, comments, and other behavioral data, user demographic characteristics (gender, age, income, city, etc.), user interests and hobbies, user relationship data, as well as video content type, emotion, specific elements, titles, and other information, all need to be abstractly extracted and converted into digital labels for the recommendation system to use. In addition to content and user labels, feature engineering also needs to pay attention to the background information of user behavior, that is, the specific context in which the behavior occurs. For example, a user tends to watch learning content on the way to work, entertainment videos on the way home, comedy variety shows in the office during lunch break, and thriller suspense dramas at home late at night. This background information is also important for algorithm recommendations. Nowadays, recommendation models and feature engineering are increasingly integrated; feature engineering itself is part of the model. These labels retain the key features of content and users, discarding redundant information. The existence of these digital labels is more conducive to the work of algorithms. With the rapid development of AI, more and more consumers are starting to use AI to ask questions and search for products. Grass-planting and SEO services targeting AI large models are becoming new business formats. In this ecosystem, labeling is becoming increasingly important for brand marketing. This is the second point I want to emphasize. As we all know, consumers always collect information before shopping to assist their purchase decisions, such as understanding product features and basic principles, collecting information on major brands in the market, and comparing performance parameters, user reviews, cost-effectiveness, etc. of different brand products. Almost everyone does this, especially for expensive and technically complex products, which can consume a lot of time and energy. From collecting information, screening brands, to weighing comparisons and finally making a decision, the entire decision-making process is time-consuming and laborious, and the decision cost for consumers is extremely high. Therefore, Herbert Simon, the father of artificial intelligence and a master of economic portfolio decision management, proposed the "bounded rationality theory" and "decision theory," for which he won the 1978 Nobel Prize in Economics. Simon believed that to achieve the optimal purchase outcome, consumers must thoroughly study all brands and products on the market, extensively collect and deeply analyze relevant information. Considering the costs involved, "satisfying is the most reasonable shopping choice." (Herbert Simon) But the emergence of AI can solve this dilemma for consumers. Before shopping, we can ask AI to help us collect information, compare products, or even directly tell us the answer. AI can process large amounts of data in a very short time, perform complex calculations, and comprehensively, systematically, and deeply analyze a matter or a product without any effort on our part. Especially for products with a high degree of standardization, such as cars, home appliances, mobile phones, and computers, AI is particularly suitable for comparing performance parameters and helping us make the best purchase decisions. Another Nobel laureate, Israeli-American cognitive psychologist Daniel Kahneman, mentioned that the basis for consumer decision-making is two cognitive systems in our brains: System 1 is fast thinking, dominated by intuition, automatically and quickly helping us make judgments to handle simple tasks in daily life; System 2 is slow thinking, centered on rationality, requiring time and energy for deep analysis to make rational judgments. (Daniel Kahneman) System 1 is a low-energy thinking activity, but it is not reliable and is prone to cognitive biases. When we encounter complex problems and System 1's operation is hindered, System 2 is activated. System 2 is rational, but it runs slowly and has high cognitive consumption. Systems 1 and 2 each have advantages and disadvantages, but AI can combine the advantages of both, helping us make decisions that are both good and fast, and completely effortless. Therefore, Pedro Domingos, author of the bestseller "The Master Algorithm," professor of computer science and engineering at the University of Washington, and co-founder of the International Machine Learning Society, said a famous quote: "AI is your System 3" (AI is your System 3). AI can greatly expand human cognitive abilities and reduce our decision costs. (Pedro Domingos) More and more consumers are using AI to recommend products to themselves and gradually relying on AI to help them make purchase decisions. A study by the Boston Consulting Group shows that about 28% of consumers frequently use AI large models to recommend cosmetics and other products. In January 2025, Xiaohongshu launched an AI search tool called "Diandian," which references the real-life experiences and shopping insights of hundreds of millions of people on Xiaohongshu to generate various life answers covering food, clothing, housing, transportation, entertainment, and more. For example, if you arrive in a new city, you can directly ask "Diandian" what delicious food, shopping, and fun things are nearby, and let it recommend restaurants, attractions, and City Walk routes. If you want to buy something, you can directly ask "Diandian" to recommend. I recently bought a pair of sports earphones. Before buying, I asked "Diandian." I told it my needs, usage scenarios, and budget range, as well as the several products I was considering, and then asked "Diandian" to compare their product parameters and pros and cons, and directly recommend one. In the end, I bought the one recommended by "Diandian." "AI search" will sooner or later become the mainstream way for people to obtain information and an important tool for purchase decisions; it is only a matter of time. And business owners will eventually realize that we are in a new situation of "AI grass-planting," and every brand must find ways to win AI's "favor"—How does AI evaluate our brand? Will AI recommend us? As a marketer, if you don't pay attention to this issue, then you can wait for your boss to come with DeepSeek results to question you—Why didn't AI mention our brand? Or the description points are wrong. When DeepSeek was booming, in February 2025, someone asked it "Which is the best Japanese ramen in Shanghai?" Among the Japanese ramen shops recommended by DeepSeek, "Cunwu Ramen" was ranked first. Sensing a business opportunity, Cunwu Ramen directly printed this answer as a poster and placed it at the store entrance, promoting itself as a restaurant "officially certified" by DeepSeek. This form of advertising may be a temporary wave, but it is enough to show that AI-generated content has commercial value and can influence user purchase behavior. As consumers asking AI for product recommendations and companies using AI for promotion become unstoppable, the marketing landscape and gameplay will eventually change. At least, a new marketing behavior has been born: when consumers ask AI, how to make AI recommend my brand and say "good things" about my products. This is called GEO. In the PC internet era, consumers used search engines like Google and Baidu to learn about product information. To ensure that corporate brands were found in the front rows by consumers, companies optimized keywords, which in marketing is called SEO (Search Engine Optimization). In the AI era, companies naturally also want their brands to appear in prominent positions in AI answers, which requires optimization for AI's information retrieval and content generation. In June 2024, scholars and independent researchers from the Indian Institute of Technology, Princeton University, and others published a paper titled "GEO: Generative Engine Optimization," formally proposing the concept, framework, and related experimental design of GEO. Moreover, as of now, marketing technology companies that optimize generated content for different AI large models have emerged both domestically and internationally. GEO can certainly be achieved because AI still needs to call network data when answering user questions. But the principle of GEO is different from traditional SEO. Classic SEO methods involve adding more keywords from the query into the content, influencing page weight through keyword density, webpage structure, title tag optimization, and PageRank and hyperlink weights to achieve higher rankings. But keyword stuffing is ineffective in GEO because users use natural language when using AI search. Therefore, AI generally adopts new search engine architectures, using hybrid retrieval and re-ranking to provide large models with relevant and sufficiently complete knowledge. Moreover, AI pays more attention to the semantic relevance between content and user questions, thereby generating more reliable Q&A results. Also, don't forget that AI search not only has retrieval but also generation steps. Therefore, to do GEO well, companies need to achieve two points: First, content library construction. Because the quality of generative AI answers highly depends on its training data, and different large models cite different network information, users asking the same question to different AIs will get different results. Therefore, brands must view their text content across the entire network from a more macro and holistic perspective, improve the quantity and quality of their information sources, gain AI's "trust," and increase the weight and probability of being called by AI. The content library should be comprehensive, systematic, diverse, and continuously updated, covering BGC, PGC, UGC, etc. Companies should not only do their own content construction, such as corporate websites, product parameter libraries, case publications, industry white papers, etc.; They should also do third-party content such as authoritative media reports and academic papers, which have higher weight for structured data; They should also do user content matrices, such as social media posts, user reviews, ratings, etc. In addition, they should promptly follow up on dynamic content such as news reports, latest products, prices, and technology releases, so that AI can crawl in real-time. The content library should also comply with the EEAT principle, which is Google's algorithm for evaluating webpage content, including Experience, Expertise, Authoritativeness, and Trustworthiness. Content that complies with the EEAT principle is more likely to achieve high rankings in search engines. In the AI era, this rule still applies because AI also cleans data through denoising, deduplication, and other techniques, filtering out low-quality content to ensure the accuracy and timeliness of content entry. Second, the content library should have more structured label construction to ensure the brand is correctly recognized. The principle of generative AI is based on large-scale language texts on the internet, predicting the most likely next word under a given input for output. Therefore, brand building should not only create content that consumers love to see, but also create content that AI "loves" to see. This means that brands must find ways to embed clearer, more distinct, and consistent labels in all their communication content, build their brand discourse power, and only then can they influence AI's content generation involving themselves. If their communication content across the entire network covers all directions, it is easy to cause confusion and misjudgment in AI large models. For example, there is a Scottish whisky brand Ballantine, mainly sold to the mass market, but Meta's AI large model Llama classified it as a top high-end brand. To this end, Ballantine's brand agency specially produced a large amount of content highlighting its popular characteristics, hoping to reshape AI's understanding through social platforms. So in the era of GEO, corporate content marketing cannot rely solely on keywords, but should shift to content library construction, create unified brand labels in different content, and find ways to compete for AI large models' "favor" and "cognitive priority." In the final analysis, whether for the human brain or computers, the process of processing information is a process of abstracting key features from specific content and extracting labels. Label extraction is the essence of human cognition of information and the working principle of the human brain. Our cognition of the external world is not like a camera copying it as is. When we see something, the brain processes the received objective information, first decomposing it, then abstracting and purifying it, and finally simplifying it into key features. We choose simple information to remember this complex world; this is the survival strategy humans have evolved over the past millions of years. The external world is complex and ever-changing, with massive amounts of information encompassing everything. We must quickly extract key points to make judgments, which is crucial for human survival and reproduction. Moreover, the brain is one of the organs that consumes the most energy in the human body, so the brain also likes to be lazy, and the pursuit of simplification is engraved in human bones. And our tool for simplification is labeling. Labeling is the way humans cognize things and a habitual action. For example, when we meet a new friend, we may be told a lot of information, such as name, occupation and position, resume, interests and hobbies, etc. We also actively obtain a lot of information, such as gender, approximate age, height, appearance, dress style, etc. In fact, it is difficult for us to remember so much information at once. What we can remember is often an important label on this person, such as top student, rich woman, stay-at-home dad, big shot, cat lover, E person, J person, post-00s, non-mainstream, show-off, second generation, middle-aged greasy, Sagittarius player, big factory 996, 985, returnee, diamond bachelor, etc. These labels are extracted based on this person's characteristics, and they constitute our "first impression" of a person. Because many labels often carry certain emotional colors and value judgments, when we label other people and things, it also reflects our own likes, dislikes, and positions. Conversely, labels will also affect our cognition subconsciously, making us take sides implicitly. For example, in some social events, the parties involved are labeled as "BMW man" or "Range Rover woman." Just such a word can deeply change our cognition and affect our views on this person. Therefore, in daily life, we must be aware of avoiding being swayed by labels in our judgments, but in brand marketing, we should also recognize the important role of labels in influencing consumer perceptions. 🔺Scan code for ticket consultation🔺