Your recommendation engine might be doing a good job. It learns what a viewer likes, serves more of it and the numbers initially look healthy. Then slowly, the home screen starts to look the same every evening for your users.
That is the rabbit hole. A viewer who watched two crime dramas now only gets crime recommendations. After a few weeks, it feels repetitive and repetitive is one step from boring. This is supported by the fact that 32% of UK streaming viewers say platform recommendations aren’t useful for finding content they want to watch.
For OTT platforms, that matters. Viewers who feel they have seen everything watch less, browse longer without finding something to watch and spend less time on the platform. Narrow recommendations can quietly turn into lower engagement and a real risk of churn.
Some content left on the table
The rabbit hole hurts twice. Viewers get less variety and a large part of your catalog never gets seen. Which is wasting your licensing and budget. If the engine never surfaces it, it earns nothing. Viewers who only see a narrow slice of your library conclude the library itself is thin. For example, Altman Solon found that the top 3% of titles generate about 37% of hours viewed, while a long tail of more than 10,000 titles accounts for roughly 4% of viewership.
Viewers perceive that there is nothing to watch. Your catalog can hold thousands of titles, Nielsen found that 19% of people will abandon a viewing session if their content search is not successful. The problem usually isn’t the content. It’s what the engine chooses to show.
How the loop forms
The cause is a feedback loop built into how most engines learn. The system recommends what it thinks a viewer will like and the viewer watches some of it. That viewing becomes the evidence for the next round of recommendations. One true-crime documentary watched on a slow Sunday can gradually turn into a home screen full of true crime. The engine isn’t learning the viewer’s full taste. It’s confirming its own guesses and every round makes the next guess narrower.
How to not get stuck in a hole
According to Nielsen, consumers globally spend 14 minutes searching for what to watch. There’s no single switch that fixes this. However, a combination of certain tactics can keep recommendations personal without letting them collapse into one lane. Here are some of the principles we use with our XroadMedia clients.
- Real-time: Many engines retrain every few hours or even overnight, so one binge can shape a viewer’s home screen for days. An engine that incorporates feedback instantly can tell the difference between a passing mood and a lasting preference.
- Go beyond similarity: “Because you watched X” is useful, but it is also the engine of the rabbit hole. Look past surface similarity to the deeper profile: which languages a viewer is open to, which moods they return to, which creators or formats they enjoy.
- A/B test everything: Run variations with different levels of diversity, different business rules and different row layouts, then measure what actually happens to plays, completion and retention. Your audience will tell you where the balance sits and it will likely differ by segment and by brand.
- Social trends: Identify relevant data from social platforms like X and channels like Google Trends to present your content that fits topics and themes trending across socials and that the audience will enjoy.
- Use context, not just history: Watch history says what someone liked before. Context says why they want to watch. The same viewer wants something different on a weekday lunch break on a phone than on a Friday night on the living-room TV with family. Signals like time of day, device, location and session behavior help the engine infer mood and activity and recommend accordingly. Locally relevant content, such as regional news or sport, becomes easier to surface too.
How you present your content matters!
In one XroadMedia trial with a large operator in Asia, the personalized rows were placed quite low in the user interface, below a number of existing rows. As expected, although the individual personalized rows performed better, the low UI visibility meant they still didn’t deliver the results we wanted to achieve.
Interactions drastically increased when the personalized rows were unpinned and gained more exposure in the UI as a result of the users’ preferences. Some assets that previously had little engagement and were positioned more effectively by our solution saw an increase of up to 600%.
For example, for those users who have shown interest in anime, our solution automatically moved one row containing mainly anime content from 17th to 5th position. That led to a significant increase in engagement and consumption. This shows that even with niche content, you are missing out on potential engagement.
Together, these small changes doubled the Effective Catalog Size across the AVOD catalog of the operator.
