Unpacking the Psychology of Successful Content Recommendations
Do you have a friend or family member whose taste in movies is just like yours, and if they recommend a show, you know you’re going to love it? Conversely, have you got another friend who is always urging you to watch their new favorite TV show, but you’ve learned from painful experience that if they love a show then it’s 99% certain to bore you to tears? It’s human nature to apply a whole load of our own experience and assumptions to any third-party recommendation. Those made by streaming services are no exception.
Personalized recommendations can be a powerful driver of content engagement—some of our customers have seen a 300% improvement in conversion from personalized rails. The use of personalization is also on the increase.
When I interviewed Caroline Carodozo, Director of Personalization & Discovery for Sky Europe, she told me that over 50% of the rails in Sky’s various UIs are automated or personalized and that the figure was increasing all the time. She also revealed that the Sky team refers to personalization internally as a “halo feature,” one that they hope will be “a comforter and delighter,” for consumers, helping them to find content they already know and love, but also new material they hadn’t heard of before which could soon be their new favorite.
Do your content recommendations measure up?
But just as you can’t target all consumers with the same recommendations, you can’t take a “one size fits all” to all the different areas of your app. It’s important to match your approach, and your algorithms, to your target audience and your specific objectives.
For most streaming services, their first experiments in personalization come with an algorithmically derived content rail called “Recommended For You.” They may also add a “Most Popular” or “Trending” section that is data-driven, but often not personalized per user. Even here, in these first steps into maximizing engagement using the data already available in your streaming service, there is subtlety that can affect your success rate.
FOMO in action
Row titles like “Most Popular,” “Trending,” and “What’s Hot Right Now” work a little like a five-star rating on a product in your online shopping search. They are a “convincer,” telling you that a series or movie is worth taking a chance on. They also play on a user’s fear of missing out (FOMO) on the current hot show that all their friends will be talking about. But the data models behind them should be subtly different.
For the majority of streaming services, a “Most Popular” row is driven by pure statistics—the content items that have had the most views in the past seven or 30 days (depending on the frequency of visits by your average viewer). For some streaming services, that means the first five slots could often be filled by the five most recent episodes of their top show. That’s great for fans of that show who want to reach it quickly, but you need a different model for everyone else.
“Trending” should involve more subtlety. We usually recommend to our customers that the model behind their trending rail is what’s popular right now—we find it works best when measured over the last 3 hours. This means that if you’re visiting the app in the middle of the day, you’ll see a very different trending section (perhaps involving pre-school kid’s shows and home-improvement shows) than if you visit later that same night when you might get dramas and more adult oriented content.
Creatures of habit
“More Like This” (MLT) and “Because You Watched” (BYW) are driven by a content-to-content data model. In data science terms, you’re placing a greater importance on metadata correlation than any of the other data science techniques used in generating recommendations, like user collaborative filtering for example.
Both MLT and BYW are designed to make inspiring connections between content items to help fans of one item find more things to watch that are thematically similar. These recommendations work on the theory that humans are creatures of habit and find comfort in watching more content in the same sphere. We always want to see more Scandi-noir dramas in my household! Frame of mind is also important. Even if a user can’t describe the mood they’re in, they know that cool series they loved last year is the kind of thing they’re after, so they’ll be looking for something like that to give them a comforting hug.
The choice of name you use for this model depends on the placement within your app. MLT is best for a deeper content page. If a user clicks on the Top Gun page to wallow in some nostalgia, a list of recommended content like this may convince them to try something new, but reassuringly similar in the catalog. This expands the range of viewing and promotes the perceived value of your service. BYW does the same thing but is better suited to the home page or genre pages within your service. A row of content entitled “Because you watched Top Gun” allows someone who is randomly browsing to short-cut their decision-making process. We tend to base BTW on the last three shows that that household watched and rotate the show between sessions.
Content-to-content recommendations (MLT and BYW) also feel less “creepy” for certain groups of users. When we say that one group of shows is similar to something you’ve already watched, it’s easy to understand our thinking. However, for some people, the idea that a streaming service can “recommend a show just for you” is a bit challenging. They can have a much more cynical reaction to “Recommended for you” rails and may respond better to content-to-content recommendations which they consider less invasive.
Expanding your horizons
With “Recommended For You” (RFY), the model should be less about metadata matches and more on the behavior of other users. This is less about FOMO and more about surprising the user with a recommendation for something they’d never go searching for, but actually find very interesting--without going so far as to recommend horror movies to someone who exclusively watches gardening programs!
You also need to think about where RFY is on that customer’s journey down the page. They have probably passed over the hero banner, the “Most Popular and “Trending” rails to get to where they are now. So, they have likely seen all the big, new, shiny content already. RFY is about exploring the range of content you have that works for that individual, so it shouldn’t be too heavily weighted towards popularity metrics.
The data science term we use is “collaborative filtering.” RFY is a great way to get your customers to try something outside their normal sphere of interest. The model can determine that a significant percentage of people who watched a particular gardening show also watched several of the home improvement series in your catalog, for example. Choosing RFY over MLT or BYW will also help to keep your app looking fresh. If you only have a small number of gardening shows on offer, eventually the metadata matching will run out of fresh content to suggest, and your recommendations start to look stale. With RFY you can make the leap to other parts of your library.
So, which is best for increased engagement?
There is no single answer to this question. But you can use all of them in different scenarios and to tackle different specific challenges. Used correctly, these models can be combined to improve the user experience and increase engagement. As I warned previously, some audiences find “Recommended for you” a bit intrusive. Other services have equal success with both RFY and BYW.
We can make recommendations based on our experience with other streaming services, but the only way to really know the best route for your particular audience is extensive A/B testing with different models and different labels. We’ve even had customers who’ve seen a decent bump in engagement purely by renaming a content rail in their app from something like “Trending” to “What’s Hot Right Now” to freshen the UX very slightly. As I said at the beginning of this article series, recommendations can’t be a “set it and forget it” process.
In my next article, I’ll look at an often-neglected area of app real-estate where personalization can have a big impact on consumer “stickiness” - the post-play moment. If you’d like to learn more, you can also download 24i’s e-guide: Five engagement-boosting strategies every streaming service should adopt right now.
[Editor's note: This is a contributed article from 24i. Streaming Media accepts vendor bylines based solely on their value to our readers.]
Personalized content recommendations are a hot topic in streaming. Extending them to your emails and other marketing messages is a logical way to increase repeat visits to your apps and counteract churn.
Metadata quality is often the biggest challenge for any streaming service looking to increase content engagement with targeted recommendations. Can machine learning come to the rescue?
Many streaming services have got binge-watching nailed for episodic content, keeping viewers hooked on their hottest series with auto-play of the next installment. But are you wasting a golden opportunity to recommend new content at the end of a series or when a viewer finishes watching non-episodic content?
Going viral in the streaming space is a taller order than when audiences first coined the term "binge-watching" as streaming platforms started dropping full seasons of content all at once. But what if, even in an increasingly crowded space, content owners and buyers could use data to determine whether content will drive viewership—or that it might even be binge-worthy?
Higher engagement with content is streaming's holy grail. More eyeballs on a greater range of content is great for tackling churn and for growing advertising revenue. But where should you invest your effort--in editorial curation or data-driven recommendations?
Amid fierce competition, streaming services are looking to new personalization strategies to increase content engagement, tackle churn, increase advertising revenue, and encourage users to return to their apps. What new and emerging data-driven approaches should streaming services use to boost their content discovery user experience?
Companies and Suppliers Mentioned