Fifteen years ago television networks collected viewer data through hosting focus groups. Based on randomly chosen viewers’ self-reported emotional responses to shows, networks would either push, delay, or cancel different series. This data was collected in what now seems to be a hilariously analog format: the network stakeholders watched viewers through one-way glass and had them report emotional responses via a dial. Termed Audience Dial Testing (ADT), this method for predicting a show’s success started in the 1960’s, and persisted for over 40 years despite its frequent failures — such as predicting that Seinfeld would flop. Today, however, streaming platforms like Netflix, Hulu, and Amazon have eradicated ADT and replaced it with a much more reliant system: predictive analytics.
The Netflix Effect
Last week, Netflix became a $100 billion company. It added 8.3 million new subscribers in the fourth quarter — six million of which came from overseas — and its subscriber base continues to outgrow those of its competitors, including Amazon and Hulu. These results can be attributed to numerous factors, including the provider’s original series — such as Stranger Things (over 15 million viewers in three days) and The Crown; the company’s high stream quality; and its uncannily accurate personalized suggestions — to name a few.
The key differentiator between Netflix and its competitors, though, is that Netflix measures everything, and uses predictive analytics to gauge the future success of the shows it invests in. That’s why it outbid competitors for House of Cards (including AMC) without even seeing a pilot. Using data points such as plot theme, viewership trends, director, and political climate, Netflix predicted that the show would be a success based purely on analytics — not personal opinions. In fact, the network credited its monster Q4 to precisely this, citing its “original content slate” as the secret to success.
The video platform hasn’t stopped at using predictive analytics for choosing shows, though; it also shows different users personalized content (such as trailers, suggested shows, and movie art) based on behavioral data to get them to watch these shows. For instance, when House of Cards premiered, cohorts of users were shown different trailers based on their tastes and viewing habits. The data used to inform what content viewers were shown went beyond genre tastes, into granular data such as volume, how often a user pauses, whether the viewer uses subtitles, and even data on users’ tastes in pace and camera shots. These data points are not random; Netflix regularly tests the predictive viability of various metrics, and then uses those that are predictive in conjunction with others to constantly finesse its algorithm. In fact, the company runs several million A/B tests per day and measures the impact of these tests daily to optimize user satisfaction.
Step by Step: How Netflix Actually Gains User Insights
Netflix collects data from two different verticals: viewers and the content itself. The company uses AI to A/B test interactions between users and genres of content in order to draw predictive conclusions about what shows and movies different users will enjoy (or at least, engage with).
Step One: Understand Viewers
- Collect metadata on viewing behavior including how much users watch, when they watch, and what they watch.
- Use this metadata to create behavioral customer segments.
- Measure degree of engagement based on remote signals including raising/lowering volume, skipping forward, re-watching, and stopping before the end.
- Gauge viewer enthusiasm movie by movie.
Step Two: Understand Content
- Collect metadata generated from IMDB and movie credits including director, actors, length, and genre.
- Collect unstructured on-screen data including luminosity, movement intensity, sound intensity, and character vs. scenery intensity.
Step Three: Actionable Insights For Future Predictions
- Create micro predictions — i.e. how a given customer will react to a specific piece of content. These predictions include the likelihood that the customer will watch a movie/TV show, the degree of engagement the customer is likely to exhibit, and the viewers behavior based on type of content (i.e. a customer is more likely to watch a TV series on weekdays and movies on weekends).
- Create macro predictions — i.e. how an entire audience will react to a given piece of content. These predictions include how many views a show/movie will gain, how popular a show/movie will be, and which cohorts of viewers are most likely to engage with/ create buzz around a piece of content.
These two levels of prediction accomplish unique yet related goals: improved user experience and high likelihood of success for Netflix original content. In the end, though, the outcome is the same: higher user retention and acquisition, and thus increased revenue for Netflix.
How Netflix Manages to Test and Analyze so Much Data
The Netflix content suggestion algorithm is generally considered the holy grail of personalization. But aside from the company’s seamless workflow outlined above, the salient reason why it’s so good at predictive analytics is the degree of automation in this workflow. Using what the team terms the “Experimentation Platform,” Netflix even tests the images associated with titles, sometimes resulting in 20% to 30% more viewing for that title. All of this testing is highly reliant on systems such as this platform, which automate the testing workflow and ensure that tests don’t interfere with each other and pollute the data.
What can you take from this? That automating real-time data collection for constant A/B testing allows you to analyze and gain insights that can inform highly impactful business decisions (like investing in a new show). When you can make both macro and micro conclusions about your users you can ensure that what you give them will satisfy their needs.
Replacing Human Emotion with Data
The implications of the Netflix algorithm expand far beyond binge-watching Grace and Frankie. Any company can implement a similar system of predictive analytics to make smart investment decisions and give users content that is statistically likely to be engaging. Global data doubles in size every two years and is projected to reach over 44 trillion gigabytes by 2020— a positive trend for the future of predictive analytics. The more data one has, the easier it is to automate the process of identifying patterns. And companies that leverage behavioral data to make predictions will gain a similar market advantage to Netflix.
The Netflix model also demonstrates the implications of constantly iterating on cohorts, user tastes, and audience preferences. Netflix never stops testing and never stops incorporating new metadata into its insights. Over time, people change on both a personal level and an audience level. Netflix is poised to keep its hold on user tastes even as they evolve over time. It is this attention to behavior that makes the company stand out.
That said, many leaders and companies default to instinct or opinions over data. Less than 0.5% of all data is ever analyzed and used, and as IBM CEO Ginni Rometty recently pointed out, “a third of your decisions are really great decisions, a third are not optimal, and a third are just wrong.” The focus groups of the past may induce familiarity-based nostalgia, but the reality is that an algorithm is better at predicting behavior than an executive is. This doesn’t mean we should do away with humans and let robots run the world, of course. It means only that a healthy interaction between code and analysts can result in more user friendly and audience pleasing products.
In 2018 Netflix plans to invest another $8 billion on content, all without the use of the almost obsolete ADT — and chances are that this investment will, once again, prove worth it.
The data and insights in this article are sourced from NewtonX experts. For the purposes of this blog, we keep our experts anonymous and ensure that no confidential data or information has been disclosed. Experts are a mix of industry consultants and previous employees of the company(s) referenced.