Thursday, April 26, 2012

Why is Discovery so hard to implement for video services?

Last week, Google said it was trying to tackle one of the hardest problems on the internet -- video Discovery.

Looking at consumer video services (Netflix, Hulu, Amazon and even GoogleTV) and their second screen counterparts (Matcha, Fanhattan, BuddyTV, etc), the admission of the challenge is painfully evident in the user interface the consumer faces and the result of the Discovery process.

But let's back up a bit first. What is Discovery? How does it relate to Search and Recommendation? I think we will find wide agreement that the concept of Search is one where you know what you are looking for and are trying to find it. Now this can be more complex than "Where can I find a legal version of Mission Impossible: Ghost Protocol that I can watch in my living room right now?" (which itself can be challenging in today's service offerings).  It is not usually as complex as the problem Shazam solves in the music industry ("what is the name of that song that sounds like..."), but can be difficult (I know the actor who was in the movie or what it was about). Search is decidedly a "lean forward" experience, and as most of us have found out over the last 5 years, it incredibly difficult to implement on a 10-foot remote experience, with various virtual keyboards or fancy remotes trying to help us solve this problem.

Recommendation is also relatively straight forward as a concept. Usually, it starts with features I have already seen that I like or genres I know I like and then asking a friend or a service to recommend something similar in hopes that the movie or TV show may also appeal. The simplest approach here is what we all know as the "Amazon" approach (people who watched that movie, also watched this movie). It can also be incredibly complex, taking into account your social network, what is currently popular, what movies or TV shows you have already seen, and what genres you like.

So what is Discovery? In this blog, I have often defined it by the service's ability to suggest relevant and interesting content to the user in a very simple, "lean back" user experience. In the real world, this might be akin to the difference between "do you have filet mignon?" (Search) vs. "I heard the fresh lobster here is amazing" (Recommendation) vs. the chef preparing a tasting menu of courses based on some consideration of what you don't like/are allergic to and his particular culinary skills (Discovery). Discovery suggests that you are going to end up watching something you didn't know you were looking for, but will very likely enjoy.

So why is this so hard to accomplish that even the mighty Google is struggling with it? The challenge is very likely in the detailed nuances of our likes and dislikes of videos. For example, "Space Balls" and "Aliens" are both science fiction, but hardly even remotely similar movies. Additionally, you may have liked the latest Mission Impossible because it is a spy movie, or an action movie, or a Tom Cruise movie--or it could have been you actually prefer movies where there is suspense, drama, gunfire, intrigue, and scantily clad heroines--but that may be difficult for you to describe yourself, but you know what you like when you see it. Lately, some of the video services have started to attack this nuance a little more directly by featuring complex genre options such as "goofy family comedies" (Netflix) or "epic heists" (Fanhattan).

Another contributing problem is the user interface, or more broadly, the user experience. For example, imagine the PC market before the mouse and graphical interface--no amount of amazing desktop publishing algorithm could solve a problem without a change in the way the user interacts with the program.

So let's consider what we are trying to accomplish again: a "lean back" experience akin to the way we surf channels these days in the living room, but that quickly and easily delivers something for us to sample with ever increasing probability of success. This very likely means the majority of Discovery experiences will need to take place on a 2-foot remote (tablet/smartphone), especially since its more mature siblings Search and Recommendation are likely to be more effective there.

Let's look at what might be important in that user interface. "Lean back" implies quick and easy, without much thought. Let's agree here that we want to find something worth sampling at least in 3 clicks/gestures. Additionally, let's consider the famous Columbia & Stanford University "When Choice is Demotivating" which concluded that consumers faced with 6 choices had a reasonable buy rate, but when faced with 36 choices, those same consumers shut down and walked away. So, as we look for a good UI, let's agree that any presentation that requires a significant amount of processing (by our brain), a significant number of clicks/typing/gestures, or presents too many choices at any given point is not going to work well.

So what is out there then?

Netflix said recently in their tech blog that 75% of their streamed content is the result of their recommendations. I think this is more likely a factor of the consumer use case for Netflix (I am bored and want to watch something, almost anything) than an indication of their algorithm's success. While the recently implemented Just For Kids vs. Adult UX is a good start, the recommendation service still struggles because there is no distinction of which members of the household are using the service at any given point, giving me recommendations on Dinosaur content and Gossip Girl and suggesting spy thrillers to my 8-year old son and wife. Further, their UI is based on a concept of rows of choices related to an algorithm choice ("Top 10 for this account", "Popular with Members Like Me", "Like: a recent title I watched") that presents at least 18 choices on the screen at once with a scrolling access to 12 rows (72 choices).  Too much.

Hulu has an anemic "Featured" (stuff someone is paying them to put in front of you) and "Most Popular" (think Top 40 radio) set of categories. At first glance you seem to be presented with only 9 choices, but the scrolling begs you to look down through row after row, presenting hundreds of choices.

Amazon has been absent on the iPad, with only their web browser as a 2-foot interface (no app), and the experience is absolutely painful. The company that built the recommendation culture has fallen flat here. To be fair, their PS3 experience has "Best of Prime", "Popular Movies" and "Popular TV", but that doesn't match with my expectations from them.

Vudu comes up significantly short with only "Top Picks" as a real category on the iPad experience. Their PS3 experience is more like Amazon's (popular rental titles, popular purchases, etc).

BuddyTV, a popular second screen app known for enabling you to Simply control your DirecTV and AT&T set top box and to recommend shows on right now from Netflix, Amazon or your channel service provider, does have a rather cool seeding process for their algorithm (asking you your preferences on a short list of movies and integrating your Facebook Likes to try to guess what you like) and presents them directly to you with only 5 initial choices (good). They do fall prey again to the scrolling process (seemingly endless choice) and separated visual choices for OTT video services from the channel line-up (oddly enough).

Fanhattan, a well-regarded second screen app known for its ability to provide tons of Stimulating content about a movie or TV show, does provide a decent filterable genre and category approach, including things like what your friends most or recently liked, but its "similar to this movie" recommendation feature is buried deep in the UI and while a powerful way to discover content, is too complex ro be "lean back" and get you there in 3 clicks/gestures.

Matcha is an app that is designed to be a "2nd screen as your 1st screen" recommendation experience, linking directly to your Netflix, Hulu and Facebook, and in theory launching you directly to those services when you pick a feature (except for Amazon, since they do not have an app). They actually do a slightly better job than Netflix, but they fall prey to the "row UI" approach, cluttering your decision field with 3 rows of 6 choices at any given time and oddly burying the recommendations row down below the initial screen, but at least limit the rows to 5 in numbers (but seemingly infinite choice to the right). They do a very good job of indicating the logic behind some of their recommendations (showing small thumbnails of Facebook friends if they have seen it or the Rotten Tomatoes and IMDB logos and ratings for the most popular content.

So apparently implementing Discovery is very hard, as the great list of companies above haven't really licked the UI/UX yet, and despite claims on performance of the recommendations themselves, the average consumer would not rave about the results either.

I think you will see two major efforts in this space in the near future:
  1) a continuous effort to improve the UI so that consumers can be presented with multiple paths to a recommendation (genre, friends, popular, new), but that allows a limited number of clicks/gestures to get to an increasingly better set of results quickly, and

  2) further effort on the algorithms themselves to better harness the nuances of the videos we like to watch and integrating that information more seamlessly with input from our social networks, our stated preferences, and external events (new releases, movie awards, etc).

I am sure all of us wish Netflix and the 2nd screen apps the best of luck in solving this since all of us would have a happier 37 hours in our week...



  1. Amazing analysis Chuck. I am saving this. Agree wholeheartedly with your take on why suggestion engines don't really work.

    As for Discovery, put yourself back in time to around 1997. You've just waked into your local Blockbuster. You may have a specific movie in mind that you are looking to rent. Or you may be in the mood for a specific genre. Or you may spend an hour wandering through the aisles because nothing in particular is striking you.

    Same thing with VOD: sometimes you know what you want and you want to be the one driving the selection process. Other times, you're happy to let someone else drive. The industry seems to focus exclusively on the latter scenario. Jinni, the Israeli site bought by Microsoft is the only site I know of that does a good job of helping me find a specific type of movie when I have a topic or genre in mind. Jinni can get very specific: movies set during the Civil War made in the 1970s.

    I wish more effort was given over to that side of the equation, which seems algorithmiclally easier to achieve.

    1. Alan, I think your POV is dead on. I hadn't though of the two use cases (I have an idea of what I want to watch such as a relatively new action/thriller vs. I have no idea).

      I also agree that the biggest problem Blockbuster presented was TOO MUCH CHOICE. They solved the new release dilemma by having 2 or 3 huge walls of only 5 or 6 new features (and tons of copies). But on the catalog, you were stuck roaming and ultimately feeling frustrated.

      I think the use case where I want to watch something, almost anything, Netflix recommendation is working. But in the use case of I don't know what to watch and only want to watch something that is time well spent, it is still very hard to find a good recommendation.

    2. Chuck/Alan... This is an excellent summary of what every pay TV operator, OTT service and consumer electronics group is attempting to tackle. Part of the issue is matching the experience with the medium. For example, the notion that television is meant to be interactive defies what an entertainment experience through your TV was intended to be.

      If you look at the history of TV, consumers could count on less than one hand the channels that were available… it was easy to find something you wanted as your choices were limited. In a sense, your recommendation was pre-decided based on the limited amount of content. Over the last 50 years, cable, satellite and telco operators continued to bring more content to the consumer... the issue, however, was that the navigation experience barely advanced... flip through the channels, watch a scrolling grid or select from an interactive grid. The analogy is a bit like the negative military term of 'Spray and Pray'… The idea of firing automatic weapons haphazardly and hoping you hit something. Thus, the notion of "I have 200 channels and can't find anything to watch" was born. When someone sits down to watch TV, they want to be entertained... passively. They are not looking to interact with their TV. If, however, the person sits at a computer/tablet, they expect to interact. This is a fundamental difference in the medium. So, how do you solve the issue of getting back to what TV was meant to be- a medium for entertainment and relaxation? More broadly, how do you match discovery with the medium and accomplish the goals of those providing the content?

      Discovery is the current word the industry is using. Discovery, however, means many things as Chuck points out. It includes recommendations, search, social, etc. What is going to win? No one knows the technology for sure, but the point is that it must match the medium. Consumers want to be guided and entertained in a TV experience- they are not looking to interact and work for it. Interaction occurs on the second screen… specifically when a tablet or computer is involved.

      There are many more methods that need to be considered other than topic or genre. Digitalsmiths, for example, brings the most powerful recommendations combined with personalized search (I.e., you don't have to work for something good out of your results… and only provides search on a medium that is conducive to interaction) as well as genre filtering and newer areas of research such as social buzz, personalized linear content and more. One reason discovery is so difficult is that it means aligning the medium, the message and driving a successful business case. There will be multiple winners and exponentially more that fail.

      At Digitalsmiths, we're excited where the industry is going… and make no mistake, discovery is tough (and Chuck succinctly points out). As the saying goes… if it was easy, everyone would be doing it.

  2. This comment has been removed by the author.

  3. Intersting and beautiful blog lovely presentation thanks for sharing your views...Samsung Leptop Support | samsung support

  4. Great analysis Chuck and great point on Netflix--they essentially picked a low-hanging fruit because of their business model. Subscribers have already paid, so they don't mind sampling things that are being recommended. Different story if one has to pay several dollars a piece to do that.

    Much remains to be understood about how people make their viewing choices in different contexts, and then how to model that in the discovery engine. Also everyone is different so the discovery engine has to consider the diversity.

    To borrow the Blockbuster example, you really need a helpful and knowledgeable shop-keeper. If you are a frequent customer, he knows your taste very well (not just based on the usual genre classifications), and can quickly come up with 10 things you will like, out of which you can pick a couple. If you are a new customer, he will ask a few questions, and quickly provide a shortlist of recommendations.