Image source: generated by Unbounded AI

When you open this "Little Red Book", you will not "click" on anything.

This is not a learning machine, but an AI search product "Dandian" that Xiaohongshu is currently testing internally. Open the search box of Xiaohongshu, and users can already see an entrance with a "dot" tail in the drop-down related search topic, and click to enter the AI ​​search dialogue interface. In this interface, you can ask questions at will, and click on the answers given by the AI ​​at will. If you don't understand, click here to get further information.

A community enterprise focusing on content eventually developed into "Little Red Book" and even became a challenger to traditional search. When AI search becomes a trend, the most fashionable Xiaohongshu will naturally not let it go.

This is not the first time Xiaohongshu has tried AI search. Throughout 2024, starting from Da Vinci Agent, to Sosousu, and now to Diandian, the entrances and paths occupied are different.

To be precise, Diandian has been in internal testing since August last year, and has launched APP and WeChat mini-program models. At first, it operated in a low-key unofficial capacity, quietly publishing content and accumulating users. A Xiaohongshu blogger broke the news and Diandian approached him several times to promote it, but he did not know it was official.

After a series of promotions in September and October, it was not until December that it was "officially" given the official title of Xiaohongshu. At the end of December, Diandian just announced the marketing and promotion plan of "Ask Diandian when you are out and about", which shows the importance Xiaohongshu attaches to it.

Different from previous Xiaohongshu AI products, the completeness and positioning of Diandian’s products are clearer and it has received greater internal marketing resource support.

Can this belated AI search product Diandian achieve a bigger search dream based on the high-value content of Xiaohongshu?

1. Integrate UGC sources in real time, even comments can be captured?

Silicon Star people started testing it and found that Diandian Search does have something to offer.

First of all, in terms of positioning, Diandian abnormally avoids the basic product catalog of Xiaohongshu. Instead, it focuses on "asking Diandian when going out", using the role of life search assistant to solve user travel strategies, Scenarios such as common sense in life.

Planting grass means screening and comparison, which requires higher understanding and analysis capabilities of AI. In life scenarios, what is more important is reality. information and comprehensive considerations. For example, we first asked a photo check-in question, "Where can I take a complete bird's eye view of the Forbidden City with a mobile phone?" As a result, Diandian was given a very complete guide, including the specific location. "Wanchun Pavilion in Jingshan Park", and mentioned the shooting time, equipment parameters, crowd conditions, and transportation methods and routes to the park. It is very reliable..

But when we asked Diandian "How to buy whitening products for oily skin", Diandian could only give a few very one-sided product recommendations and reasons. On the contrary, there was no comprehensive user post in the answer source, at least. The post contains clear whitening ideas, product combinations, and detailed information on how to use it.

On the day of the test, it happened to snow in some parts of Beijing. Because it was too "local", those who saw the snow and those who did not seemed to be experiencing two worlds. So, we asked "Where did it snow in Beijing today?" Diandian not only quickly identified the North Fifth Ring Road, Yanqing, Xiangshan and other places, but also paired it with fresh videos shot by users, and the date understanding was very accurate.

This seemingly simple answer reflects the three core characteristics of Diandian: using UGC content as the source of information, real-time information, and the ability to understand comments.

When Diandian answered the question about snowfall in Beijing, it did not simply call weather data, but integrated the field visit notes of local users - someone posted the first snow in Xiangshan, and some photographers shared it The composition of the snow scene in Yanqing. This kind of content integration based on real experience is the barrier established by the Xiaohongshu community.

And answering real-time questions is also an ability that many AI searches do not have.

Take Secret Tower Search and Kimi as examples. When asking about the content of a certain day, you can only answer a rough time period, and the information is often lagging behind.

In addition, since Xiaohongshu’s UGC content is composed of text, pictures, videos and even comment interactions, this also requires AI to have a more comprehensive understanding of these unstructured data. For example, in the snowing question, the snowing location was mentioned by the user in the comment area, but not in the text or title. This capability also makes Diandian search results more credible.

If a user searches for "Which hot pot restaurant is the best?", Diandian can not only read the text, but also analyze the feedback in the comment area. Even if the main note of an internet celebrity store has rave reviews, if negative feedback such as "the queue is too long" or "the service has deteriorated" has frequently appeared in the comment area recently, it would be amazing if AI can promptly remind users to pay attention to these changes.

Of course, multimodal understanding capabilities not only appear in the understanding of search results, but also include the understanding of user input (query). Currently, Diandian supports voice and image input to ask questions. For example, we uploaded a photo like this, and Diandian could clearly understand that it was a border collie, and there were some irregular pieces of paper on the ground.

In this regard, Diandian understood that the dog may have caused damage, and proactively guided users to discuss how to avoid pet chewing. This ability to "look at pictures and speak" makes the search process closer to natural conversation.

2. Where not, "click"

At the same time, Diandian's "exploration function" design also opens up a new interactive method for AI search.

Dandian does not copy the question and answer model of traditional search engines. Instead, it adopts a guided interaction of "click where you don't know where to click". It does not require a new round of voice or text input. By clicking on the key in the answer words, lower the threshold for secondary questions, and extend the search with fool-like operations.

However, these functions seem powerful, but there is also an obvious illusion of a large model. For example, when we asked him about check-in places near Tongzhou Collection Park, the graphic and text information he gave was not within the park. , but the format looks neat and correct. What's more, random associations will occur and users will complain about them.

It is undeniable that in the past year, Xiaohongshu has made a lot of AI search products, from DaVinci, to Sosousu, to Diandian, you can see Xiaohongshu’s Cautious and confused. Each iteration is trying to answer the same question: How to combine new interaction methods with old user habits?

Is it to sort the content of posts, or to further optimize past search efficiency? Should we optimize quality through multiple rounds of dialogue and increase AI thinking, or should we present user content as neutrally as possible? Xiaohongshu has gone through many iterations on these issues.

Take the AI ​​assistant Da Vinci embedded in the main website as an example. Through the form of Agent, it conducts multiple rounds of dialogue with users to help users search for information and summarize information; while Sosousu does not have a dedicated traffic entrance. , equivalent to a post-search summary, is relatively marginal in terms of optimizing user experience.

The Diandian stage is more like a combination of Sosousu and Da Vinci.

On the path, there is a high probability that Diandian, like Sosousu, will summarize the content after searching. That is, using traditional search to find the corresponding content, AI will understand several links and summarize the complete answer. We asked Diandian and conducted traditional searches with the same questions, and found that the sources of AI answers were basically the top positions in the original search.

In terms of interaction, Diandian is a chatbot like Da Vinci, but its ability to have multiple rounds of conversations is limited; and Diandian has both an independent App and a main site search entrance. It can be said that Diandian wants to take advantage of the main site traffic. , and worry that the AI ​​function will affect the main website experience.

These product-level swings will inevitably affect organizational structure and resource allocation.

Of course, the repetition and swing of constantly changing vests makes users feel confused, but in the earliest stages of product development, continuous exploration is more likely than standing still.

3. AI search and community attributes once again staged "love and kill each other"

We found that with the UGC real-time content, the usability is obviously stronger than other products with homogeneous content on the Internet.

This leads to the followingThe question is, does AI search need to use a browser as a carrier like Perplexity, or does it need to use a content platform as a carrier like Xiaohongshu?

On the one hand, the value of search lies in the quality of the source itself, and on the other hand, it lies in how to maximize the value of these contents. Compared with general-purpose AI searches such as Perplexity, Xiaohongshu has accumulated a large amount of life-related UGC content in vertical fields, which is its biggest advantage in AI search, especially since there are massive real needs and uses within the Xiaohongshu platform itself. Scenario, according to the latest data reported by the media, the average daily search volume of Xiaohongshu has been close to 600 million times in the fourth quarter of 2024, which has doubled from 300 million times in mid-2023.

From beauty and makeup to travel guides, from food shopping to home renovations, these contents are not simple product introductions or service descriptions, but real user experiences and scenario-based solutions. For example, a note about "Walking the kids in Shanghai on weekends" may include weather advice, transportation routes, suitable restaurants, activity arrangements, and even emergency plans. This kind of scenario-based and experiential content is difficult for traditional search engines to accumulate.

These are also the capabilities that Perplexity has always coveted. Even though Perplexity is stronger, it still has to face its dependence on search engine APIs and the fundamental problem of content homogeneity.

So when Xiaohongshu does AI search, most people will think that this is a natural action. Because the capabilities it has accumulated are basically exclusive capabilities that can be used for AI search.

More importantly, Xiaohongshu, which does not rely on search advertising, does not have the commercial concerns of traditional search engines - there is no phenomenon that once it develops AI search, it will be equivalent to slashing itself with a knife.

But Xiaohongshu’s AI search also has its own problems.

From a technical perspective, unlike other AI searches, Xiaohongshu, which guards users and data treasures, needs to think about how to maximize the value of this content, especially to understand a large number of unstructured long-tail content. Detailed content.

Currently, when we open the content source of Diandian search, we can see that the source content is basically the same as the result of traditional search directly on Xiaohongshu. It is equivalent to AI helping users conduct traditional search and read Several links are provided and a summary conclusion is given. This obviously does not fully release the value of Xiaohongshu’s UGC long-tail content, but is more of an extension of Xiaohongshu’s original search method.

From the perspective of community attributes, it is difficult for the little sweet potatoes who are accustomed to real user feedback to accept the formatting and uncertainty of AI search. There have been many conflicting voices: some users think this This kind of intelligent summary makes the user's personal experience invisible in it, damaging the most important sense of reality of the social platform. It is large and general, but loses its persuasiveness.

In addition, when users are accustomed to AI’s quick answers, will they still be willing to browse notes immersively? Will this lead to a decline in user interaction with real posts?Will it reduce the data value of the user's original posts and impact the content community itself?

In terms of AI, Xiaohongshu is likely to once again experience difficulties similar to "difficult to balance community and commercialization".

In order to avoid a repeat of the past, Xiaohongshu needs to determine the priority of the two. Which is more important, the community? Is AI search just a small function that optimizes user experience to serve the community? Or does AI search have a bigger future? , the user base of the Xiaohongshu community can be used as a ticket to open a new world of search and reconstruct the traffic field.

At present, several AI search products launched by Xiaohongshu are still in an unfinished state, and it is estimated that more "dots" will appear in the future.