Structured, but time-consuming.
Audio is not intuitive. Users always have a lingering concern before deciding to listen: What if I don't like it after clicking? Although one can comfort oneself by thinking that even if disliked, it's still a way to kill time, people always want to have their cake and eat it too. They prefer to enjoy their time-wasting activities; otherwise, they end up losing both time and mood, leaving them empty-handed.
In such scenarios, AI becomes an excellent assistant. Although audio itself is not intuitive, it can be converted into text, transforming it into an NLP problem. This allows large language models to showcase their strengths by understanding content, summarizing key points, and presenting the content intuitively, helping users filter and judge.
In fact, as long as it meets the conditions of ### "structured" + "time-consuming", AI can be brought in to play a role. For example, WeChat Reading has also launched an AI book questioning feature integrated into the platform. When users are interested in a certain concept or question, AI book questioning can link to specific content pages of relevant books within WeChat Reading, facilitating users to read and learn in-depth. This is also a kind of mini-search based on WeChat Reading's own ecosystem.
However, in Amazon's attempt, it further utilizes the dialogue capabilities of large models for precise recommendations. Similarly, the domestic podcast platform Xiaoyuzhou is also trying this approach, launching the "Ask Xiaoyuzhou" beta version.
This feature is currently not integrated into the Xiaoyuzhou client side but is a separate webpage with a unique design style, mimicking early browser web pages. It creates an effect of "Although you've received AI help, you should still open the podcast and listen instead of continuing to browse the web."
Compared to structured content, AI will be more meaningful in mining and integrating fragmented content, though it also faces more challenges. This is also the significance of various content platforms successively integrating AI search into their platforms.
The most representative is Xiaohongshu, which launched two features in succession: a dedicated AI assistant "Da Vinci" and "Sousou Shu" for search.
We have done evaluations before, and both features have their strengths and weaknesses. Currently, they are still quite preliminary, and the recommended content cannot be fully adopted. Users need to jump to the cited notes to confirm and verify the content. A common feature is that they both revitalize the rich note content on Xiaohongshu.
For content ecosystems rooted in specific environments, it's a thorny area for traditional search. On one hand, due to ecosystem protection, search engines can't "reach" them. On the other hand, built-in search functions are generally not user-friendly. For example, Weibo's advanced search function is still based on basic information such as time and location, with very limited accuracy.
This is because social media brings a large amount of content, but it's very fragmented. This both provides more room for AI search to play a role and poses greater challenges.
Unlike more regular podcast and audiobook products, taking Xiaohongshu as an example, the content forms on such social media platforms are diverse, including images, text, videos, and live streams. Moreover, this content itself comes from personal experiences and feelings, mixed with a large number of internet memes, emoticons, etc. - users with slow internet speeds might not even keep up.
When users expect recommendations that are more in line with current tastes, existing recommendation algorithms, usually based on long-term user interest modeling, need to slowly collect user preference and behavior data to build profiles, tending to recommend preferences that users have already shown.
In comparison, AI search is a good starting point, getting feedback through users' search behavior. Especially for frequent internet surfers who often only vaguely capture certain hot topics.
At present, the user interest sparked by popular discourse is a little, but not much, and needs further understanding to supplement. At this time, when users take the initiative to step out for a search, search based on large language models can play a better role.
Search is a query-response process, while recommendation is a continuous dynamic process. The intersection of the two lies in the goal of being more personalized based on user needs. The burden of "creating information cocoons" that traditional recommendation algorithms have always carried may be improved through integrated AI search.