Disclosure: I’m a personal investor in Toutiao.
Using Machine and Deep Learning to Create and Serve Content, China’s Toutiao Created a Product with Engagement Similar to that of Social Networks - All without a Social Graph
Toutiao, one of the flagship products of Bytedance*, may be the largest app you’ve never heard of–it’s like every news feed you read, YouTube, and TechMeme in one. Over 120M people in China use it each day. Yet what’s most interesting about Toutiao isn’t that people consume such varied content all in one place… it’s how Toutiao serves it up. Without any explicit user inputs, social graph, or product purchase history to rely on, Toutiao offers a personalized, high quality-content feed for each user that is powered by machine and deep learning algorithms.
Going a step further than merely serving up content, Toutiao’s algorithms also create content: During the 2016 Olympics, a Toutiao bot wrote original news coverage, publishing stories on major events more quickly than traditional media outlets. The bot-written articles enjoyed read rates (# of reads and # of impressions) in line with those produced at a slower speed and higher cost by human writers on average.
The average user spends more than 74 minutes each day in Toutiao -- that’s more than the average user spends on Facebook1, and more than twice what they spend on Snapchat2. More than half that time is spent watching short-form videos; this coupled with over 10 billion video views per day makes Toutiao the YouTube of China (along with, of course, everything else it offers).
How did Toutiao do this? Especially without massive consumer platforms at scale like those orchestrated by Chinese conglomerates Alibaba, Baidu, and Tencent? In this post I’ll explore how Toutiao got to 120M daily active users. Toutiao doesn’t attribute its growth to any one factor, but rather to the interplay between many tactical and strategic decisions it made starting at launch; specifically, five key advantages, all of which I have outlined below. And while “super apps” aren’t as common in the U.S., I believe there are specific lessons in this case that can inspire others in building their own products and platforms.
But first, a bit of background
Toutiao launched in 2012. The app uses machine and deep learning algorithms to source and surface content that users will find most interesting. Toutiao’s underlying technology learns about readers through their usage – taps, swipes, time spent on each article, time of the day the user reads, pauses, comments, interactions with the content and location – but doesn’t require any explicit input from the user and is not built on their social graph. Today, each user is measured across millions of dimensions and the result is a personalized, extensive, and high-quality content feed for every user, each time they open the app.Sources: Snapchat - S-1 filing. Instagram - Recode. Facebook - Q1 2016 earnings report.
While timing is everything for a startup, it takes deliberate effort to build an addictive app. Toutiao’s timing was fortuitous, but its exploitation of this unique moment was deliberate. Toutiao launched as smartphone use was taking off in China: mobile penetration increased from nearly nothing in 2010 to 65% by 20143. Moreover, many of the largest content providers had not yet developed mobile apps or mobile-friendly sites, meaning that true mobile-optimized information and entertainment was rare. By mid-2012, there were only six significant news apps on the Chinese Android platform. Four of them were direct extensions of existing news portals with limited mobile optimization, and the other two were aggregators that relied exclusively on slow and impersonal editor input to determine what stories to show. Further, the Chinese audience's demand for content (both articles and videos) was underserved by Chinese social networks such as WeChat and Weibo. WeChat launched as a messenger and to this day has a closed social network (i.e. sharing/moments are private to friends only).
Toutiao stepped into this gap with an easy-to-use, personalized, informative, and addictive mobile-first app. From the outset, Toutiao was extremely easy to start using – all it took was a download. There was no need to create an account and password, to link it to social media (unless the user so desired), or to provide information on interests or preferences. The app’s simple design also made it intuitive to use with no prior knowledge or tutorials. For any app, driving initial engagement – moving from downloads to DAUs – is notoriously difficult. It’s typical to lose users at every step of the process due to discouragement, confusion, or annoyance.
The name of the app Jinri Toutiao (meaning “today’s headlines” in Chinese) and the icon of the app were catchy for users, resulting in excellent user growth. It was also the first time various news articles were aggregated in one place. From the very early days, Toutiao tracked information about each user - their taps, swipes, time spent per article and location to power the recommendation engine which we will discuss later in the post. One month after launch, Toutiao became a personalized news aggregator for several of its users. The product, the only one of its kind and delicately designed at that time, led to a rapid growth. They hit 1M DAUs only four months after launch. Toutiao gave new internet users something to “do” when their mobile time was still up for grabs. Toutiao updated the app almost weekly throughout its first year, as it consistently innovated, iterated, and improved its features and algorithms, and this resulted in improved retention over time.
In the years that followed, competition for user share of attention on mobile would drastically increase – the number of mobile apps available in China more than tripled in the three years from 2012 to 20154. But Toutiao’s early lead meant that, by the time competitors arrived, it already had an important and valuable foothold.
The image below shows the personalized feed of two different users.
You can have all the algorithms in the world, but without an addictive product there is no data, and without data, no algorithm can make the system better. Matt Turck has written about the power of the data network here. Simply put, the more users use your product, the more data they contribute. The more data they contribute, the smarter your product becomes. The smarter your product is (e.g., better personalization, recommendations), the better it serves your users and they are more likely to come back often and contribute more data — thus creating a virtuous cycle.
By building an addictive product, Toutiao generates engagement data from their users. That data is fed into Toutiao’s algorithms, which in turn further refines the products’ quality. Ultimately, the company plans to use this virtuous cycle to optimize every stage of what they call the “content lifecycle”: Creation, Curation, Recommendation and Interaction.
Ever since the invention of written language, content creation has been the exclusive domain of humans. Toutiao looks to change that. It’s begun with Xiaomingbot, an artificial intelligence that has already published more than 8,000 stories on the platform to-date. It debuted during the Olympics in 2016, where it published stories on major events more quickly (approximately 2 seconds after the event ended) than traditional media outlets. Indeed, the bot-authored articles enjoyed read rates (# of reads and # of impressions) in-line with those produced at a slower speed and higher cost by human writers on average.
Below is a screenshot of an article written by the Xiaomingbot describing the results of the tennis match between Andy Murray and Juan Martin Del Potro during the 2016 Olympics.
To achieve this, Toutiao had to overcome a couple of significant technical challenges:
First, writing stories on Olympic game results required data, and Toutiao pulled it from three sources: [a] real time score updates from the Olympics organization, [b] images from an image-gathering-company it had recently acquired to find relevant visual media, and [c] monitoring live text commentary about the game. It also started with four sports — Table Tennis, Tennis, Badminton and Women’s Soccer — that were easier to recap from a technical standpoint (Table Tennis, Tennis and Badminton are “turn-based” games and the rules of the games are simpler vs. other sports. Unique access to a high-quality data source for Women’s Soccer made that the fourth game covered.)
Second, Toutiao had to figure out how to combine data from these three sources to ensure an internally consistent and relevant story. This was a much larger challenge than even accessing and interpreting the data in the first place. Any selected image needed to be relevant to the results of the event, and also appropriate for the takeaways from the commentary. This, in turn, required Toutiao’s AI team to integrate natural language processing capabilities with contextual image recognition. They ended up with a combination of a grammar-based representation for generating story templates, a ranking algorithm to select relevant sentences from live text commentary, and an image-text matching algorithm to tie it all together. The system also employs convolutional neural networks to analyze content in candidate images. By training on historical data, the model is able to pick the most relevant and visually appealing image for the story. They also use sequence-to-sequence deep learning algorithms to summarize existing stories into daily highlights and suggest better titles for articles.The system employs recurrent neural networks to compute vector representation for sentences and these sentence vectors are further fed into a ranking model to pick concise summaries for each article.
The products of these efforts – 450 published stories with 500-1,000 words during the Rio Olympics – that were hugely successful. They enjoyed read rates (# of reads divided by # of impressions) on par with those produced at a slower speed and higher cost by human writers. Toutiao has extended this capability beyond sports to over 8,000 stories to-date, and is working hard to close some of the remaining technical loopholes that make human writers recognizable.
A major engagement driver for Toutiao in its early days was “soft news”– areas like celebrity gossip, pop culture and lifestyle articles. This was no accident. Contrary to official news, which was distributed by well-known state-owned news sources, soft content was distributed across the internet on a plethora of individual sites. In short, there was no central place to access the content: users who were looking for it would have to invest meaningful time in visiting different sites, and had no assurance they were getting the most interesting information. Toutiao changed that. In owning, centralizing, and optimizing the distribution, it reduced the time a user needed to find content to nearly zero, and it increased their confidence that they were reading the most interesting stories. This created real value for users.
At its core, content curation is a two-sided problem: the curator must find content, in addition to serving it to its users. The first requires visiting websites, identifying stories, and collecting relevant metadata. The second requires continuously updating a central repository of stories, and creating as many personalized versions as possible. Both are process-intensive tasks where algorithms have a distinct advantage over humans. Toutiao’s only meaningful competition in this space when it launched were web portals where human editors handled this work, and Toutiao’s use of algorithms gave it a major advantage over the manual competition.
The speed with which the system could do what took human editors much longer translated directly into value for Toutiao’s users. Toutiao could gather more content more quickly and at a lower cost, creating a major advantage in a business were customer value is directly tied to content quality, relevance and refresh rate. The use of algorithms also meant that each user could have their own, interest based and continuously updated profile – something that no human editor would ever have the time to do.
Toutiao also uses algorithms to identify and filter out low-quality content. A content distribution platform is only as good as the content it distributes. The days of mass-distributed cookie-cutter content (e.g., newspapers, magazines) are over. In Toutiao’s world, the distribution platform only serves what is interesting to its users. False reporting and spam are major issues in the media industry. Toutiao’s underlying technology uses a text classification algorithm to determine if an article is fake news, uses clickbait titles, or doesn’t meet Toutiao’s quality standards. Here, Toutiao also leans on user moderators to flag fake articles and employs human moderators to arbitrate on disputed reporting.
Content recommendation is the feature for which Toutiao is best-known, and to which it owes much of its success and reputation. The use of machine and deep learning algorithms at this stage of the content lifecycle is what has sets Toutiao apart from its peers, and is key to driving continued user growth and retention.
The question that the recommendation engine is trying to solve is simple: what are the one hundred articles the platform can recommend to each user that are most likely to result in continued engagement? This is a question with major consequences – the AI team has recognized that 100 headlines is a retention “threshold” (users that do not retain long-term tend to drop off dramatically after seeing ~100 headlines, similar to Facebook’s “10 friends” rule). It is also a question that humans are unsuited to answer: no human editor could ever regularly and quickly identify the optimal set of headlines for every one of the app’s new users.
As simple as the question may be, the solution is complex. For every new user, Toutiao blends signals from three key areas to create a feed that it hopes is engaging and will push users over the 100-headline threshold:
The underlying algorithms must then identify the strongest statistical match between the user’s profile, its own content profile, and context, and it must do this on a continual basis. This matching is meant to optimize the percent of articles a user reads (clicks on) and the percent of articles that a user finishes (measured by the time spent on the page). When a user first opens the app, the system uses the basic data in the profile for the matching: a user in Silicon Valley, for example, may be more likely to click on articles about tech. The system also makes sure to show a variety of articles to assess interest/disinterest– in doing so, can help users discover previously unknown content and test their potential interests. Over time, as the app collects user information, these recommendations get further and further refined. The engine learns quickly – for most users, it takes less than one day to successfully learn their interests (indicated by 80% read rates). The result is the case of strong user retention (>45%) that is similar to social networks and one of the largest time spent per user apps in the world.
As Toutiao has grown, interaction on the platform has become more and more central to its user value proposition. Rather than leaving it to the users to find each other, Toutiao uses underlying algorithms to help enable meaningful connections. Nowhere is this more relevant than in its recently developed question-and-answer feature, where the AI team was tasked with developing a matching engine that links a question-asker with someone who can answer them. Toutiao recently published a paper for the ACL conference touting these results. Their proposed “Conditional Focused Neural Question Answering with Large-Scale Knowledge Bases approach” achieves an accuracy of 75.7% on a data set of 108K questions, and outperforms the current state of the art (better than the Memory Network and LTG-CNN methods on the benchmark dataset) by an 11.8% margin.
Toutiao’s underlying technology not only creates a better user experience, but also serves to strengthen the company’s competitive moat. More compelling content and interactions meant users would spend more time on the platform, and the more time they spent on the platform the better the use of algorithms became. The smarter the system is, the better it can distribute content – and the more content creators it attracts. This, in turn, drives more users to the platform. And thus is born a strong data network effect - the power of the system grows exponentially with the scale of the system. There are competitors who have launched since then (especially after seeing Toutiao’s success), however it has been difficult to match the accuracy and efficacy of the Toutiao recommendation engine leading to continued rapid growth for Toutiao.
It is not uncommon to see apps strive to move from content aggregation to content destination. However it is extremely challenging from a brand and creative strategy to make that happen. Here is how Toutiao did it. Toutiao offered two significant benefits to content contributors over the platforms.
Strong incentives via revenue sharing that enabled writers to make money from very early on. In 2014, Toutiao rolled out incentive programs to attract more content creators to the platform. These ranged from offering office space, tools, minimum guarantees per month if they hit certain key milestones (e.g., # of articles, read rates) to sharing revenue via monetization. Toutiao began monetizing via ads since 2014 and this enabled revenue sharing opportunities with their content contributors.
This was the function that launched Toutiao, but as it has grown, Toutiao has transitioned into a deeper platform for content generation, consumption, and connections. Today, it hosts more than 800,000 Toutiaohao accounts – professional media outlets, bloggers, and influencers who use the platform to share articles, images, and videos with Toutiao users . It hosts many more users sharing short posts through Wei Toutiao. The result is the wide variety of content that Toutiao hosts today ranging from news to stocks to science to relationships. Top 20 categories account for only 60% of the content supply and no single category contributes over 10% of the content.
Below is an example of a variety of content that a user can choose from (the screenshot only displays the 40 of the 50+ channels users can choose from):
Larger and more relevant audience than other platforms that directly translated to increasing brand presence for content contributors. Almost all contributors create and distribute content on all platforms. But for many contributors, they have the ability to attract more traffic from Toutiao due to the strong recommendation engine. One example is " 欢子tv“ ( Huanzi TV). This creator creates short videos about folks' lives and customs in the countryside of China. Each of his videos has an average of 700,000 views, while the views in his Wechat official account is less than 1/40 of that on Toutiao. Toutiao has enabled the long tail of contributors to reach their most relevant audience more seamlessly than any other platform in China.
Instead of being stubborn about their core format (e.g., listicles, long form content and news), Toutiao was quick to expand to other formats when the data suggested they should. In 2015, at the time where most video platforms in China are focusing on long-form videos, Toutiao added video capability and started to support PGC short video content (typically 1-5 mins) on its platform. Toutiao had observed an increase in supply of video content in 2014 as connectivity and infrastructure had improved significantly by 2014. Additionally, Toutiao rolled out several incentive programs to promote video content creation on its platform. The transition from text to image to video was similar to what most US platforms have seen to date.
Later in March 2016, Toutiao launched Toutiao Video (which is now renamed to Watermelon Video), a separate PGC short video app powered by the same algorithm engine as Toutiao. Similar to written content, the underlying algorithms recommend the most relevant videos to users based on their interest graph. Toutiao is now the "go-to platform" for PGC short video content. More than half of its 74 mins daily usage for each user is spent watching short-form videos and Toutiao is close to exceeding 10 billion video views every day.
Toutiao has reached unprecedented scale in revenue in a short time frame (5 years since launch and 3 years since they began monetizing) and it is remarkable that they are doing it without leveraging any social graph or product purchase history. Toutiao is on target to hit more than 15B RMB (>$2.2B USD) in revenue this year - one of the fastest growing apps in terms of revenue in the history of the internet.
Of the many things that Toutiao does, one element that is core to its model more than any other: it is good at identifying what its users want to see. It is fitting, then, that its business model maps perfectly to that strength. Toutiao generates revenue by matching relevant ads to users, using the same proprietary technology behind their content targeting. This has three important benefits:
First, it reduced the impact of monetization on the user experience – and may have actually improved the experience! Users typically consider ads as intrusive and degrading to their experience, but ads aligned with user preferences are less so. In serving ads that are highly relevant to a user’s interests, Toutiao in many ways acts as a product discovery mechanism.
The second is that it increased the rates that Toutiao could charge advertisers. One of the key problems in advertising is identifying how to selectively place your ads in front of the highest potential customers, and advertisers spend countless hours and enormous sums of money trying to target effectively. Toutiao’s technology, which solves this targeting problem natively, represented a solution and saves advertisers from paying a big premium for it.
Third, since the primary use case is to read and view content, users are more receptive to seeing relevant targeted ads and therefore there is more inventory available to advertisers.
The combination of all these three factors results in much better CTRs (Click Through Rates) on Toutiao vs. competitors. Third party survey data estimates Toutiao’s CTRs to be 200% better than its peers.
Impact on the Future of Content Discovery
Toutiao is chipping away at their end goal, which is essentially to wipe away the concept of search and just serve up aggregated, hyper-relevant content. We’ve seen “content aggregators” come and go in the U.S. but it is possible that they are an idea whose time is yet to come – and that better algorithms will be the catalyst for success. Facebook and Twitter are both critical sources for consumption of news in the U.S. today. The other giant in the room here is Google, which in July announced that the feed in its mobile app would be increasing its use of machine learning to better show their users the information they will find most relevant and interesting – a feed that incorporates all types of news.
Special thanks to the Toutiao team, Sharon Pope, Craig Cannon, Sonal Chokshi, Kat Manalac, Daniel Gross, and Ram Parameswaran for reading multiple drafts of this essay.
Founded in March 2012 in Beijing, Bytedance is at the global forefront of innovating artificial intelligence technologies. Bytedance is dedicated to optimizing the connection of people with information, as well as promoting content creation and communications. Its flagship product, Toutiao, is the largest AI-powered content discovery platform in China, it delivers personalized content recommendations to every user based on their interests. Bytedance owns a series of products celebrated by the users around the world, including Topbuzz, Flipagram and a series of UGC short video apps.
Bytedance established an AI Lab in 2016, leveraging extensive and complex datasets to conduct state-of-the-art research in artificial intelligence.
1. Source: Facebook Q1 2016 Earnings Call.↩
2. Source: http://time.com/4272935/snapchat-users-usage-time-app-advertising/.↩
3. Source: http://www.businessinsider.com/china-has-more-smartphone-users-than-us-brazil-and-indonesia-combined-2015-7.↩
4. Source: https://www.statista.com/statistics/315485/china-number-of-mobile-apps-available/.↩