Behind The Rise of TikTok
Product design, algorithm, content and network effect
TikTok’s global blitzscale has been a massive success. Starting in China under the name Douyin just 4 years ago, TikTok has now reached more than 850 million monthly active users globally. It was consistently ranked as the most downloaded app on App Store and Google Play throughout 2019 and 2020.
I’ve been reading and thinking about the reasons behind TikTok’s success both in terms of product design, content, network effect, and distribution. Below is my take on this.
1. TikTok’s product design - The path of least resistance
Water flows from high altitude to low altitude, simply because it’s the path of least resistance. Similarly, watching Netflix is easier than doing homeworks because it’s the path of least resistance. The path of least resistance means the easiest way, in which you spend the least amount of effort to do something just like water flows naturally.
This concept is usually translated and applied to product design as always design your product flow with as least friction as possible or as the title of a classic book on UX by Steve Krug stated - “Don’t make me think!”
The path of least resistance rule is tightly integrated into TikTok’s product design. When you open the app for the first time, you can immediately watch videos. You don’t even have to create an account to start like most other social networks as it creates a shadow profile based on your device ID.
Furthermore, you don’t even have to pick what to see or who to follow to get started. Right off the bat, Tiktok decides what should be recommended for you => You just literally don’t have to think. The only decision you have to make is whether to swipe down or continue watching.
By actively sitting in the driver seat itself, TikTok removes the decision paralysis for the user. Decision paralysis aka The Paradox of Choice is a state in which the user is inundated with too many options to choose, making them unable to decide anything at all. This happens all the time to me whenever I browse Netflix, I sometimes spend 10-15 mins but still can’t decide on which one to watch since there are just too many.
As Barry Schwartz argues in his book - Paradox of Choice, eliminating instead of adding more choices can greatly reduce anxiety for shoppers and increase conversion rate.
TikTok understands and applies this really well into its product. Of course, this is in part because TikTok is really confident in its recommendation algorithm, which is our next stop.
2. Algorithm & Data touchpoint advantage
TikTok’s feed recommendation is so good, there is no doubt about that. But to get why we have to, first, understand its parent company in China - ByteDance.
ByteDance’s first product Jinri Toutiao (“Today’s Headlines”) launched in August of 2012. Basically, it’s like the Facebook newsfeed but only consists of articles gathered from different websites in China personalized for each user.
“ New users logged in with their Sina or Weibo account, which Toutiao scraped for initial interests and friends. It then used each user's individual usage data (how they tapped, swiped, or paused, time spent per article, their comments, location, time of day, and much more) to serve each user the most relevant content. It changed the titles, cover images, and even shortened most articles. This got users to an 80% read rate on each article in less than one day, which contributed to the 45% lifetime user retention it boasted early on. This was all instant and free compared to human editors and gave Toutiao a 10x better product at a 10x lower cost.
It peaked in mid-2018 at around 200 million DAU’s using it for an average of 74 minutes per day; nearly twice as long as Facebook, Instagram, and Snapchat. ”
The Rise of TikTok and Understanding ByteDance by Novak Turner.
Leveraging the strength of Toutiao’s personalization models, TikTok had a really good starting point regarding the algorithm. However, as you might know, a good algorithm is only half of the equation when it comes to training ML models. The other half lies in having massive clusters of user data for the model to train on, this is when the nature of short videos comes in handy for TikTok.
Since TikTok videos are usually very short from 15 seconds to less than 2 mins => in a session of using the app, an average user may consume many videos at once. => This naturally creates more data touchpoints for TikTok to feed into its ML models to understand its users in a very short period => which in turn makes the recommendation on its feed more accurate and addictive => hence making users spend more time on the app => generating more data points.
This flywheel of data generation and recommendation improvement is a huge advantage of TikTok’s short-form video nature compared to other video-first platforms like Youtube or Netflix whose the number of content units consumed per session is much smaller.
Another reason that makes TikTok able to build a really good feed is its low cost of bad recommendation. Imagine wasting 30 seconds to watch a boring video on TikTok versus spending 2 hours on Netflix just to realize that you’ve just watched a really lame movie, what would piss you off more? the latter one, right?
Since a core tradeoff in online personalization is when to explore vs. exploit a user's preferences, having a higher risk tolerance of users when it comes to exploring user’s interest makes TikTok well-positioned to understand its users.
Eric Stromberg articulated this in his essay on Deep Understanding:
“TikTok, Instagram, and the Quadrant of Deep Understanding
A core tradeoff in online personalization is when to explore vs. exploit a user's preferences. Example: Let's say we are building a system for music personalization. A new user listens to a country song and "likes" it.
We can exploit that learning and recommend another country song, or we can explore the user's tastes and recommend a pop song. Exploiting what we know will deliver immediate engagement but the user may eventually get bored. Exploring can reveal a new preference but is risky; if we make a bad recommendation the user may close the app.
This explore/exploit process — executed at scale and algorithmically — is key to how apps learn about their users. User understanding is deepest when the number of signals per session is high, and the cost of a bad recommendation is low. Cost of Bad High Recommendation Low” - Screenshot Essay by Eric Stromberg.
3. Non-perishable content
Content can be divided into two categories - perishable and nonperishable.
Perishable content just like food gets rotten and loses its value over time. It has a limited lifespan of circulation, such as daily news.
Non-perishable content, on the other hand, doesn’t lose value over time. For example, classic novels, music, hilarious videos, scenery images, etc. You can watch these kinds of content now or 5 months later or even 5 years later and can still appreciate its beauty and value. Non-perishable content doesn’t have an expiring date, therefore can be reused.
The ratio of perishable/non-perishable content on different social networks varies significantly depending on its social graph and users’ curation. From my personal experiences, Facebook has the largest amount of perishable content since most information on its newsfeed comes from what is happening at the moment.
Youtube and Twitter have a more balanced mix of timeless wisdom/ junky news on their feed. However, this might vary significantly from user to user since what’s recommended on each user’s newsfeed depends on their choice of who to follow.
When it comes to TikTok, its feed looks like a huge collection of hits. Most of TikTok’s hits are either hilarious, cute, thrilling, or about attractive boys/girls => they’re mostly non-perishable content. Because of that nature, TikTok can reuse its content over and over without worrying about them become obsolete.
You might not notice this but out of these platforms, only TikTok and Youtube reuse their old content by actually recommending these videos on users’ feed. All content shown by algorithms on Facebook, Twitter, or Instagram’s feeds was newly created at the time of browsing.
However, TikTok goes so far on reusing its timeless hits that it’s the ONLY online social network that doesn’t display the timestamp of content shown on its feed.
A friend of mine who just started using TikTok recently excitedly shared with me hilarious videos on his feed. It turned out that I had watched most of them months ago, and it’s now being recommended to my friend - a TikTok newbie.
4. Global network effect
Local v Global Network Effects
When speaking of network effects, the words local and global do not just refer to geography. A global network is an entire network that captures every single user. A local network is a subset of users who are clustered together around a common modality (such as interest, socio-economic class, or physical location).
Illustration of Global network effect (Airbnb) vs Local network effect (Uber). Source: Cornell University
The more a network is fragmented into local clusters—and the more isolated those clusters are from one another—the more vulnerable a business is to challenges. Consider Uber. Drivers in Boston care mostly about the number of riders in Boston, and riders in Boston care mostly about drivers in Boston. Except for frequent travelers, no one in Boston cares much about the number of drivers and riders in, say, San Francisco. This makes it easy for another ride-sharing service to reach critical mass in a local market and take off through a differentiated offer such as a lower price.
Now let’s compare Uber’s market with Airbnb’s. Travelers don’t care much about the number of Airbnb hosts in their home cities; instead, they care about how many there are in the cities they plan to visit. Hence, the network more or less is one large cluster. Any real challenger to Airbnb would have to enter the market on a global scale—building brand awareness around the world to attract critical masses of travelers and hosts. So breaking into Airbnb’s market becomes much more costly.
The bottom line about local vs global network effect is that the more global network effect a product possesses, the stronger its defensive moat is.
Back to our case with TikTok, with very generic types of content that most people can relate to, TikTok's global network effect both in terms of geography and content category is much stronger than other social networks.
For instance, It’s much harder for a tweet from a VC on Twitter to resonate well with a random teenager in Vietnam, but it’s more likely that a video of a cute cat or a dancing girl anywhere in the world will be liked by someone else across the globe. Most videos on TikTok only have action and background music, everyone can feel and enjoy these. It’s not a tweet or a long audio-video that requires users to understand foreign languages to relate to. That’s why you can see many videos on TikTok reached millions of likes even though the creators are ordinary people coming from all corners of the world. In contrast, reaching millions of likes on Facebook, Instagram or Twitter is the status that is usually reserved for celebrities only.
My favorite example of TikTok’s global network effect is about a farmer living in the countryside of Southern Vietnam. Ytiet is a farmer making a living by raising cows, in his meantime, he shoots short videos as a side hobby. Little did he know that one of his TikTok counting videos in which he literally counted from 1,2,3 to 26 would become highly viral and catapulted him to fame.
Stories like that of Ytiet are not uncommon on TikTok, blessed with inherently viral content in its nature, TikTok’s global network effect is one of the strongest among all social platforms.
To be continue
So far we’ve delved into the rationale behind TikTok’s rise from the perspective of product design, algorithm, and content. In the next post, we’ll see how TikTok combined money and viral growth hack tactics to pull off one of the most impressive growth stories ever in the consumer tech world.
See you then.
Minh Phan
Special shoutout to Novak Turner for his long essay on TikTok and ByteDance, definitely one of the best writeups on this topic. Since our take is quite different, I hope this somehow serves as a complement for your pieces.
—
Preview image by Zion, Interface Market and Kimmi Lo via Behance.