Zooming in on Video Personalization: KPIs, Use Cases, and Best Practices

Gabriella Vas Gabriella Vas
11 min read | May 10, 2022

We’ve come a long way from the online videos of yore that required a robust connection and nerves of steel: they would freeze at the most exciting moments, forcing us to wait helplessly for the next few seconds of action to load. With the advent of streaming technology, online videos became a commodity for viewers and, at the same time, a lucrative source of revenue for publishers. The sudden abundance of video content and the ensuing drop in attention spans made recommendations indispensable. 

Today, video personalization is the lifeblood of streaming platforms. But it’s a tough challenge to serve real-time, relevant recommendations to millions of users, many of whom share profiles with family members, switch devices, or behave confusingly in other ways. Read on to see the big picture about video personalization: its potential, its KPIs, its functions, and its solutions to the issues mentioned above.

How Video Personalization Creates Value for Streaming Sites

There are a number of ways to consume video content online: on social video platforms like YouTube or DailyMotion; in the video sections of news portals; using OTT (over-the-top) streaming services like Netflix; or as part of the IPTV (internet protocol television) offering of telco providers. From a revenue perspective, though, it all boils down to two main business models: free content with ads, or video on demand, accessible either on a pay-per-view or on a subscription basis.

In the case of the ad revenue model, the goal is to get viewers to watch as many videos as possible, exposing them to pre-roll and midroll ads. It’s down to video personalization to keep them watching one clip after the other by serving relevant, intriguing recommendations.

For subscription-based streaming services, the objective is to increase customer engagement and loyalty. If viewers cancel their subscription, it’s because they feel that interesting content on the site has run out; that they have seen all there is to see. Given the vast libraries of most streaming platforms, this is hardly the case. So video personalization is geared towards enabling subscribers to discover more content they’re bound to like. 

And more doesn’t necessarily mean new. Streaming service providers are under tremendous pressure to keep adding new films in order to retain viewers. But entertaining everyone with exclusively fresh content all the time would not be feasible. Video personalization helps to make the most of existing material by interspersing recommendations with earlier movies or shows that are in some way similar to the most sought-after, latest releases.

Choosing what to watch is always a bit of a headache for viewers, whether they’re newbies or seasoned series addicts. This is because the specificities of the video format make it difficult to assess movies at a glance. Whatever the preview includes – artwork, description, cast, trailer, even ratings – may not provide enough information to convince someone that watching the given film will be time well spent. Friends’ recommendations matter more to viewers, or, lacking those, the choices of like-minded people. The collaborative filtering algorithms of video personalization engines cater precisely to this preference. 

The KPIs of Video Personalization and the Streaming Sector

It follows from the two business models described earlier that the most important metrics for video personalization measure click-through rate or watch time as a result of recommendations. 

  • More specifically, click-through rate (CTR) is the percentage of recommended videos users clicked on, of all recommendations displayed. It indicates the relevance of recommendations, and, indirectly, the quality of user experience. The higher CTR a video personalization engine can achieve, the more ad revenue it can help generate for providers of free content. 
  • Watch time through recommendation is a natural number: the total of seconds users spend watching recommended videos. (That is, watching a video within 24 hours after clicking on it in a recommendation widget.) Assessing or A/B testing the performance of video personalization this way is especially useful for subscription-based streaming services. 

However, the go-to metric for streaming platforms is average watch time per user. Calculated as the total time spent watching videos divided by the number of all users on a site, it is the indicator of overall user engagement. Bear in mind, video personalization is but one of several factors contributing to this metric, and its precise impact is difficult to measure.

Video Personalization Use Cases

Helping Viewers Get Started

As the first and most important touchpoint, the home page plays the lead role in hooking first time visitors and in engaging returning users. Therefore, streaming platforms invest significant resources in perfecting the layout of the home page with video personalization in mind. 

Apparently, the most efficient home page layout to date, used by many streaming sites but made famous by Netflix, is a series of rows (also referred to in some regions as “trays” or “swimlanes”). Each row contains a number of videos grouped according to a certain logic (“More Films Like…”, “Because You Watched…”, or “Continue Watching”), or selected from a specific category (“Award-Winning TV Shows”) or genre (“Romantic Comedies”). Users can browse each row horizontally for more videos within that group, or scroll through rows vertically to discover new topics. This way, the home page enables both exploitation and exploration: a viewer can dig deeper in the movie types she has shown interest in before, or scroll on to find new material that may fascinate her.

Needless to say, implementing video personalization on the home page requires constant experimentation. It starts with identifying a host of possible video groupings. The next step is figuring out which of these are relevant to a given user, and within each group, what specific titles. Data shows that it helps to add justification (“evidence”, in Netflix terminology) for including a certain row in a user’s personalized home page. 

The arrangement of the rows and of the items in each row is also important. Because users scroll from top to bottom or from left to right, it makes sense to place the most relevant recommendations in the first few positions of the top rows, where people tend to look first. 

The simplest way to populate the home page with videos is the rule-based approach, meaning, every user is shown the same rows in the same order, but with a personalized assortment of videos in each row. The most sophisticated solution is to rank the rows themselves, as well as the videos they contain, in a personalized way. This presents a complex algorithmic challenge; it takes many iterations to find the approach that provides the optimal balance of relevance versus diversity on the personalized home page. 

Extending Watch Time

Once the viewer has finally settled on something to watch, the next challenge for video personalization is to keep her glued to the screen, even after the chosen movie or clip is over. One way to achieve this is a function called continuous play. In this case, the next most relevant video starts playing by default after the first one ends. 

It’s reasonable to wonder: did this form of video personalization lead to the phenomenon of binge-watching series, or was it the other way around? According to our experts, the drive to consume entire seasons in one sitting, fuelled by the need for instant gratification, was a characteristic shared by many people even before personalization tapped into it successfully. 

In any case, recommending series – basically, up to dozens of videos bundled into one item – is another smart move to keep viewers watching. The logic is pretty straightforward: a person who has watched episodes of a series will be prompted to continue watching the series on subsequent visits to the home page. If she grows bored with the series in question, the video personalization engine detects the pattern: the more time passes between watching episodes, the less likely it is that the series will appear again among her recommendations. 

Recommending playlists is yet another way to extend watch time. As opposed to continuous play, it gives more control to the viewer in deciding what kind of content she wants to keep watching.

The Challenges and Best Practices of Video Personalization

Making the Most of User-Generated Content 

On social platforms like YouTube or DailyMotion, users can publish their own videos, and this presents a specific set of challenges for video personalization. 

Smart as it may be, the recommendation engine cannot “see” the actual content of videos. Metadata (video descriptions, ratings, tags, categories) is meant to help with ranking new items. In the case of user-generated content, however, this metadata is often inconsistent, patchy, or missing altogether – unstructured, in data science lingo. Some users add lengthy descriptions to their videos, others write one-liners, or none at all. One influencer may classify her morning skin routine tutorial under Beauty and Lifestyle, while another one may categorize similar content as a How-to. 

Video personalization systems work their way around this complication by ignoring metadata and focusing instead on viewer behavior to generate recommendations. By evaluating previous interactions with a video, collaborative filtering algorithms can predict who else might appreciate it – much more accurately than by parsing descriptions. 

That’s one problem sorted, but there’s more. Because social platforms reward their most popular contributors with a share of the ad revenue, as a result, many user-generated videos are low-quality, clickbait affairs with no apparent value or purpose other than ramping up the number of views. If the platform’s business rules favor conversion over CTR (that is, actual views, not just quick clicks), the video personalization engine can filter out the clickbait pieces based on their (typically short) watch time.  

Adapting to Shared Accounts and Devices

In an ideal world, people would behave predictably, and recommending movies to them would be a breeze. Alas, this is not the case. 

A video personalization system is at its best when it can identify individual viewers and build elaborate profiles based on their movie choices and ratings. However, many people are not ready to create their own accounts, or can’t afford to do so, or just can’t be bothered. Sharing video streaming accounts with family or household members is common practice, as is sharing or switching between devices. 

But it takes more than this to confuse a good recommendation engine. It can read between the lines of contextual data: even if several people log in to the same account to watch films on the same tablet or laptop, the time of day or the choice of genres can help to separate the various viewer profiles. Cartoons in the afternoon, thrillers after dark? A well-trained video personalization system gets the idea, and comes up with appropriate, time-sensitive recommendations. 

In other cases, although the account is shared, it’s accessed from multiple devices, some used exclusively by one person. Again, it’s about connecting the dots: if watching certain genres or categories of movies can be linked to one specific device, the video personalization engine will display similar recommendations on that screen, regardless of the preferences of other users under the same profile.   

Understanding Negative Implicit Feedback

As the quip often misattributed to Einstein goes, “The definition of insanity is doing the same thing over and over again and expecting different results”. In the same vein, it doesn’t make much sense for a video personalization system to keep recommending items that viewers have ignored before. 

Therefore, on top of tracking everything viewers do on a video streaming platform, personalization providers, Yusp included, also observe and measure whatever they don’t do. The lack of clicks on a recommendation is interpreted as negative implicit feedback. 

This is the kind of response that is only implied through user behavior and assumed by the provider. Whether someone skipped a recommendation because they didn’t like it, or because they were distracted by something else, we can’t tell. And just because they watched a film, we can’t be sure that they actually liked it. 

On the other hand, explicit feedback, like when a viewer rates the movie he has just seen, is way more straightforward. Then again, it can also be biased. For instance, someone may be inclined to positively rate a movie they deem “worthy”, like an Oscar winner, but reluctant to show appreciation for a trashy rom-com, even though they enjoyed the latter more. 

There you have it: implicit feedback is uncertain, while explicit feedback is unreliable. In terms of quantity, though, the former is definitely more abundant. All in all, implicit feedback paints a fairly good picture of viewer behavior and as such, it is also taken into account by video personalization systems. From the end users’ perspective, this translates into recommendations that adapt dynamically to their preferences – their active choices as well as their passes. 

Scaling Up to Serve a Large Audience

Popular movie sites, especially in populous countries, entertain millions of viewers at any given moment. Case in point: Hotstar, India’s largest premium video streaming platform, boasting up to 100 million daily active users. 

As Hotstar’s video personalization system, Yusp faced (and aced) the challenge of serving this humongous crowd of viewers relevant recommendations in real time. In fact, the task had a few particular aspects that made it, if possible, even more daunting. 

Besides a library comprising 100 000 hours of movies, Hotstar also provides live coverage of major global sports events, and these, especially cricket tournaments, are the primary attraction for Indian viewers. The video personalization challenge is to give them reason to linger after the cricket match (often lasting several hours) is over. Given the sport’s universal appeal on the Indian subcontinent, watching a cricket game is hardly a differentiating factor collaborative filtering could make use of. How to tell what else would interest and ideally convert a visitor? 

To segment the audience and recommend videos accordingly, Yusp used contextual data such as type of device or location. In addition, detecting language settings on viewers’ devices also helped to generate relevant recommendations. (To serve India’s multilingual population, Hotstar offers content in 17 languages.) 

Live cricket coverage serving as the gateway to Hotstar meant not just a giant viewer base but also sudden spikes in demand for recommendations. During match breaks, or when a game ended, the video personalization engine had to entertain millions of viewers at the same time. Thanks to its ability to scale, Yusp was able to serve more than one billion recommendations daily, and its data centers in the region ensured real-time responsiveness. 

That’s All, Folks!

As we have seen, video personalization plays a key role in converting new viewers and in minimizing subscriber churn on streaming platforms – essentially by taking user experience to the next level. When faced with challenges like messy metadata, confusing user behavior, or a tsunami of simultaneous requests, video recommendation engines like Yusp find innovative ways to stay relevant. 

Learn more about our solutions, and don’t hesitate to get in touch if you have specific questions. 

What to read next

Join our newsletter

Get to know the ins and outs of personalization