If you’ve ever found yourself watching TikTok for a couple of hours, you’re not alone. According to the “State of Mobile 2021” report released by our portfolio company AppAnnie, TikTok was the most popular social networking app in Japan in 2020. It surpassed even LINE, Twitter, Instagram, and Facebook in monthly active users (MAU). In addition, according to app store data from Sensor Tower, a U.S. mobile app research firm, the total number of downloads for TikTok and its Chinese domestic version Douyin (抖音) has surpassed three billion.

But what exactly is the secret sauce that keeps bringing users back to the app again and again and again? It’s TikTok’s powerful AI recommendation feature.

A Scarily Accurate Experiment

In mini-documentary published in July 2021, The Wall Street Journal created more than 100 bot accounts to investigate exactly how TikTok’s algorithm learns its users’ preferences.The study reported that when users first sign up, they see popular videos in their feeds, and as the app gets to know their users better, users begin to see more niche and customized content. TikTok’s algorithm began to understand the preference of most bots in less than two hours, and in the fastest case, in less than 40 minutes.

TikTok’s algorithms worked faster and better than those of other platforms because not only does it learn preferences, it can even detect users’ vulnerabilities, according to Guillaume Chaslot, a former Google engineer and data scientist who helped design YouTube’s algorithm. He also mentioned that on YouTube, over 70 percent of views come from recommendation engines, but TikTok‘s views from recs are about 90 to 95 percent.

A bot named “@Kentucky_96”, programmed to be interested in sad and depression-inducing videos, actively watched videos with the hashtags #sad #mentalhealth #anxiety. And by the time @Kentucky_96 had watched 224 videos in 36 minutes, the TikTok algorithm had figured out that he was interested in videos about depression and mental helth, not videos about relationships and breakups. After that, 93 percent of the videos in its feed were about sadness and depression.

TiikTok claims to intentionally play videos unrelated to user preferences in order to maintain content diversity, but the remaining seven percent of videos in @Kentucky_96’s feed were ads.

How It Works

TikTok’s June 2020 blog on how recommendations work listed some factors that helped determine which videos were recommended, and that included user interactions such as liking and sharing videos, followed accounts, posted comments, and content created. It would also survey information in the video that the users would interact with, such as captions, sounds, and hashtags. Other determing factos were device and account settings, such as language, country settings, and device type.

All of these variable are then processed by TikTok’s recommendation engine to determine each factor’s importance. For example, whether a user watched the video all the way is weighted more important than whether the user and the video creator are in the same country.

TikTok collects a large number of indicators from each user and uses its own algorithm to recommend videos that appear in our feeds.

The Money-Making Algorithms

In July 2021, Financial Times reported that ByteDance has begun to quietly launched a service called BytePlus Recommend in June. One of the case studies on their site, claims a “40% metric improvement” of “increase in conversions per user.”

There is no detailed information about the implementation process and fees are not public, but the SOP begins with BytePlus begins by learning about their clients’ needs, decide what kinds of data need to be validated, training models, and eventually launching their recommendation service on their clients’ platforms.

In addition to the recommendation algorithms “BytePlus Recommend,” BytePlus also offers “BytePlus Effects,” which provides various SDKs for AR experiences, “BytePlus Translate”, a machine translation service that can be easily integrated into applications and websites, and “DataRangers”, which provides A/B testing and user behavior analysis tools.

Before ByteDance launched BytePlus, it also launched a similar Chinese version of this service called Volcano Engine, which is implemented by services operated by ByteDance, such as TouTiao, Xigua Vide, DongCheDi, JianYing, FanQie Novel, and PiPiXia.

Business Strategies of Successful Social Enterprises

It’s worth noting that ByteDance, the company behind TikTok, one of the most popular social services in the world, is now in the 2B business. (Even before the launch of BytePlus and Volcano Engine, ByteDance also offered workspace applications such as Feishu and Lark.) This isn’t altogether an uncommon route for tech giants.

AWS is too well known to mention as an example, but there are several others, including Epic’s Unreal Engine. Two common reasons behind this strategy particular are to offset the repercussions of of rapid growth, which hits a ceiling in the global market or causes lower growth, but also to mitigate the risk of becoming dependent on one business model.

One company that employed this strategy was Snap. In June 2018, Snap announced the launch of their Snap Kit. Ironically, it was around the same time Snapchat’s daily active user count began to decline quarter-over-quarter.

Facebook, which relies heavily on the advertising model, saw its market value drop by about $60 billion in two days last year due to a boycott by a number of advertisers, including Starbucks. Compared to the advertising model, the 2B business is expected to have a relatively stable income.

It may be a good heads up that social services, which have such a large number of users, are likely to develop new 2B businesses using their infrastructure, in addition to the original business models such as advertising and commissions.

Subscribe to our newsletter and stay up to date with all the latest Headline Asia news.

IAT006