The visual cortex takes up 30% of the human brain, so people naturally process images 60,000 times faster than text. As a result, we’re all guilty of judging books by their covers, which is a crucial element of Netflix’s engagement strategy.
Netflix has a seemingly unlimited supply of content, which can be a double-edged sword. That volume of content produces choice paralysis , where people have so many options that they can’t decide what to watch. They ultimately don’t watch anything, which underlines the importance of Netflix’s sophisticated recommendation algorithm. Netflix customizes everything on the homepage, right down to the title images.
“We don’t have a product. We have hundreds of millions of products  because we deliver personalized experiences,” says Tony Jebara, Director of Machine Learning. “We’re looking to not just make recommendations, but have members believe in them.”
AI driving engagement
Machine learning drives Netflix’s algorithms, which play a huge role in the company’s success. By recommending the most relevant content, they increase engagement significantly, saving the company an annual $1 billion .
The algorithms dictate the personalized homepages and “top picks,” taking each subscriber’s viewing habits and preferences into consideration. The company utilizes a process called interleaving , which identifies the most promising ranking algorithms from a large set of initial ideas and then A/B tests using the pared-down algorithms. Compared with traditional A/B testing, interleaving works faster with smaller sample sizes.
“You end up kicking yourself because during the whole period of innovation, you’re not giving the best possible machine-learning algorithm to the users,” says Jebara.
What does this mean for title images?
Netflix used to use the generic title images, provided by studio partners. They were often scaled-down versions of DVD cover art that didn’t necessarily fit . The company has since performed tests to determine the images most likely to catch people’s limited attention. Engagement metrics include click-through rate, aggregate play duration, and what percentage of views had short durations.
Images with expressive facial emotions perform well, as do those featuring particular people. For example, Unbreakable Kimmy Schmidt got the best engagement with an image including both the titular character and fan favorite Tituss Burgess.
What performs even better are the images that are specifically selected for you.
“Let’s say Good Will Hunting is recommended. If we know someone really loves comedies because they’ve watched Zoolander and Arrested Development, they title image might have a picture of Robin Williams,” explains Jebara.
Do you watch a lot of movies with John Travolta or Uma Thurman? That factors into which Pulp Fiction title image you see.
A quick experiment
At The Next Web’d TNW event, Jebara’s presentation included visuals. But they were the “official examples” from Netflix. I was curious, how did this factor into my own personal Netflix homepage?
As a Shameless fan, I searched for William H. Macy’s biggest movies, like Fargo and Magnolia. When none were available to stream, I decided to focus on the filmographies of less-established cast members.
Mall is 2014 movie I’ve never heard of about disaffected suburbanites who come together in a mall following a shooting spree. Look at the difference between Mall‘s movie poster and my title image on Netflix. Guess what TV show the redhead is on.
The Netflix experience revolves storytelling, in a way.
“Storytelling is the heart of the human race,” says Jebara. “It’s how we share information, pass on our culture and teach the language. It started with cavemen drawing and now we’ve got PowerPoint.”
Netflix views title images as “the gateway to the stories” and the machine-learning algorithm serves the right ones for each user .
- ^ volume of content produces choice paralysis (hbr.org)
- ^ We have hundreds of millions of products (www.clickz.com)
- ^ saving the company an annual $1 billion (www.clickz.com)
- ^ a process called interleaving (medium.com)
- ^ that didn’t necessarily fit (www.clickz.com)
- ^ the machine-learning algorithm serves the right ones for each user (www.clickz.com)