Easier said than done, of course.
One disadvantage of being an online store was that it made browsing difficult. If you knew what you were looking for, you could just search for it. But if not, finding movies was surprisingly difficult. You could only view one page at a time, and there was a limited number of movies you could fit on a page. You had to make a snap judgment based on the cover art or a synopsis. This was a problem in brick-and-mortars, too, of course. According to Mitch, most people walked into video stores completely unsure of what they were looking for, and simply drifted from section to section. But in a brick-and-mortar, you could ask a clerk for help. Or, at the very least, you could wander the aisles, and hope that you’d serendipitously stumble across something that looked promising.
We wanted to make browsing easier, and we also wanted to connect users with recommendations and reviews. So Christina, the editorial content team, and I designed content-rich landing pages for a variety of genres. If you were looking for a thriller, we had an entire page dedicated to them, replete with top-ten lists, reviews of recent and classic thrillers, and highlighted selections from our inventory. If you liked Tom Cruise movies—same deal. The idea was to provide gentle suggestions and guidance, something akin to what a sympathetic (and knowledgeable) video store clerk could offer.
We wanted to offer a personalized touch. The problem was, it was enormously expensive—not to mention time-consuming—to do it all manually. When we had 900 titles, it was somewhat feasible to create content to match. But by late 1999, we had almost 5,000 movies to work with. It was hard to keep up, and even harder to browse.
Reed, in typical Reed fashion, pushed for automation.
“Forget the landing pages,” he said. “We’re redesigning the site anyway. Instead of hard-coding pages, how about we just do it like this: Create a frame on the home page that has slots to display four movies at a time. Each slot can show the cover of the movie, run time, date of release, a little capsule synopsis—the data we already have. Then just make a list of fifty movies you might want to have appear there, and have the site randomly pick which four to display. Or better yet, just define how to build the list—maybe call the list ‘thrillers’ and let the system randomly pick from any movie we have that is tagged as a thriller.”
If I recall correctly, I reacted with horror to this suggestion. I hated it. It seemed cold, computerized, random—all the things we weren’t trying to be.
But have you used Netflix lately? Reed’s slot structure survives—with alterations. The most crucial of which is that the films in the slots aren’t randomly chosen. They’re the product of a complex algorithmic matching service, one that’s calibrated to both your taste and Netflix’s needs.
That algorithmic matching service can be traced directly back to 2000 and Reed’s slots. Because he was right, of course—users needed a more efficient, easier way to find movies they would like, something even more intuitive than an editorially curated landing page. Putting DVDs into slots was a start. Now we just needed to figure out some way to arrange them that wasn’t random.
In talks all that fall, we discussed ways to build a service that would give users movies they’d love while also making our life as a distributor easier (and more profitable). When users sat down to decide which movies to order next, we wanted them to see a list of films that had been customized to their taste—and optimized for our inventory. If we could show customers what they wanted to watch, they’d be happier with the service. And if we could also show them what we wanted them to watch? Win-win.
Put simply: Even if we were ordering twenty times more new releases than any Blockbuster (an enormously expensive gambit), we wouldn’t be able to satisfy all demand, all the time. And new releases were expensive. To keep our customers happy and our costs reasonable, we needed to direct users to less in-demand movies that we knew they’d like—and probably like even better than new releases.
For example: Say I rented (and loved) Pleasantville, one of the best movies of 1998 and a clever dark comedy about what happens when two teenagers from the nineties (Tobey Maguire and Reese Witherspoon) are sucked into a black-and-white television show set in 1950s small-town America. The ideal recommendation engine would be able to steer me away from more current new releases and toward other movies, like Pleasantville—movies like Doc Hollywood.
That was a tall order. The thing about taste is that it’s subjective. And the number of factors in play, when trying to establish similarities between films, is almost endless. Do you group films by actor, by director, by genre? Release year, award nominations, screenwriter? How does one quantify a thing like mood?
I worked with Reed and the engineers for months on a solution. The problem was coming up with an algorithm that actually spat out movies that made sense together. Since it could only use the data available to it—things like genre, actors, location, release year, language, and so forth—the algorithm often made suggestions that made sense to a computer but didn’t really take into account any kind of real-world similarity. Or, it would give unhelpful suggestions: “You like Top Gun? Here’s another movie that came out in 1986!”
In the end, we realized that the best way to give users what they wanted was to crowdsource data from them. At first, we did what Amazon did. Using a process called “collaborative filtering,” Amazon would suggest products to you based on common buying patterns. They still do this. Essentially, if you buy a wrench from Amazon, it groups you with other users who have bought a wrench, and then suggests that you buy other things that they’ve bought.
Here’s how it worked with rentals: Let’s say Reed and I each rented three movies from Netflix. I rented Armageddon, The Bridges of Madison County, and Casablanca. And Reed rented Armageddon, The Bridges of Madison County, and The Mighty Ducks. Collaborative filtering would say that since we’d both rented two of the same movies, we would probably each enjoy the third movie that the other person rented. Therefore, the site would recommend that I rent The Mighty Ducks and that Reed rent Casablanca.
The problem with this method, of course, is that filtering for rental history doesn’t really tell you whether I liked Casablanca, or if Reed liked The Mighty Ducks. It just tells us that we both rented those movies. We could have hated them. We could have rented them for our kids (or our wives).
If we were going to use collaborative filtering to group customers and recommend films, we needed to know what customers enjoyed rather than just what they rented. We needed a reviews system: a movie rating system. Grouping customers by ratings—by “clustering” users according to overlapping positive or negative reviews—meant that we could efficiently recommend films to users based not on what they’d rented but what they liked. Ultimately, the algorithm would become much more complex than that. But for it to work at all, we needed users to review movies—lots of them.
Ultimately, we decided that we would ask our customers to rate movies by assigning each movie from one to five stars. Five stars for a movie they loved. One star for a complete time waster.
It sounds simple enough, but that stupid star rating system was the source of hundreds of hours of argument. More battles about fewer pixels have never been waged. Could you give something zero stars? Should we offer a half-star option? When you gave a rating it was whole stars, but when we predicted a rating, should it be in whole stars or in tenths? When should a user be prompted to review a film? Where should the widget go?
In the end, we asked Netflix users to review films early and often. We would ask them to rate films whenever they visited the site, whenever they returned a movie, and whenever they rearranged their queue. The great thing about movie rentals is that you don’t have to rent a movie to have already seen it—unlike buying a wrench, a review didn’t have to be tied to a sale. Theoretically, a user could review every movie he or she had ever seen—even if he or she had never rented a single movie from us. And it turns out that people love to be asked for their opinion. Everyone’s a critic.
It was remarkably easy to amass enough reviews to build a collaborative filtering function that could actually predict—with reasonable accuracy—what someone might like. After that, Reed’s team went to work integrating these taste predictions into a broader algorithm that made movie recommendations after weighing a number of factors—keyword, number of copies, number of copies in stock, cost per disc.
The result—which launched in February of 2000 as Cinematch—was a seemingly more intuitive recommendation engine, one that outsourced qualitative assessment to users while also optimizing things on the back end. In many ways, it was the best of both worlds: an automated system that nonetheless felt human, like a video store clerk asking you what you’d seen lately and then recommending something he knew you’d like—and that he had in stock.
Actually, it felt better than human. It felt invisible.
If it sounds like two of the most innovative and influential developments in the history of Netflix happened quickly, hot on the heels of Reed and I deciding to run the company together—well, if it sounds that way, that’s because it’s true.
Reed and I came to our CEO/president agreement in September of 1998. Within a year, the subscription plan was live. Within a year and a half, it was the only way to rent from Netflix—and a redesigned site was connecting with customers using an innovative algorithm that gave them exactly what we knew they’d want…and what we wanted them to have.
Those two key innovations would be enough to prove to almost anyone that we’d made the right choice when it came to running the company. We were really singing together. The team I’d built was bursting with creative ideas to connect with our users, and Reed’s had a singular focus in streamlining our vision. Reed’s laser focus helped us concentrate on the future. My goal was to make sure that however quickly we moved, however efficient we got, we were always fundamentally seeking to connect with our users.
Past and future, heart and brain, Lennon and McCartney—Reed and I were a perfect pair.
15.
Drowning in Our
Own Success
(September 2000: two and a half years after launch)
ALISAL RANCH MIGHT NOT be at the end of the earth—but you can see it from there.
If you want to see for yourself, head to Santa Barbara. Then drive thirty miles north, up the 101. Veer east when you reach the town of Solvang and its rows of faux-Danish storefronts. Leave the quaint markers of civilization behind, and continue on a single-lane secondary road, through brown bunchgrass meadows dotted with California oaks. Kick up clouds of dust for what seems like hours. And just when you think you’re really and truly lost, you’ll find yourself coming around a sharp bend and there it will be: the Alisal Guest Ranch. Ten thousand acres of rolling California foothills in the middle of nowhere.
I don’t know what we were thinking—or even who was thinking it—but the Alisal Ranch is where, in September of 2000, right as the last bit of air was streaming out of the dot-com balloon, we decided to have our first corporate retreat.
There was plenty for us to retreat and talk about that September. Earlier, in the spring, we’d raised another $50 million in financing—our Series E—bringing the total amount of money invested in Netflix to more than $100 million. The share price for the Series E had come in at nearly $10 per share. And since I still owned a shitload of shares, I was now worth an absolutely obscene amount of money…at least on paper. Since I couldn’t sell any of my stock, it was just imaginary—funny money. Still, it reduced the frequency with which Lorraine brought up the idea of selling the house and moving to Montana.
Netflix now had more than 350 employees, and we had long since passed the point where I knew everybody. We’d continued on our streak of making major talent hires—the most recent being Leslie Kilgore, whom Reed had convinced to leave Amazon to head our marketing efforts as CMO, and Ted Sarandos, who now managed our content acquisition.
Since walking away from à la carte rentals, our no-due-dates, no-late-fees program had steadily built up steam. Users loved Cinematch, our recommendation engine. We did, too. It kept our subscribers’ queues full—and nothing, we found, correlated more to retention than a queue with lots of movies in it. We were now approaching nearly 200,000 paying subscribers. Our other metrics were looking pretty impressive as well. We now carried 5,800 different DVD titles and shipped more than 800,000 discs a month, and our warehouse was packed with more than a million discs. Tom Dillon was making headway on a method to ensure that users could get access to those discs within a day of ordering them.
Earlier in the year, at the height of the dot-com boom, we’d seen the bankers circling us like vultures with briefcases, and had even flirted with going public. More than flirted, really. We’d chosen Deutsche Bank to manage the offering, had hired accountants to go over our books, and had drawn up an S-1 (also known as a registration statement), the document laying out to the Securities and Exchange Commission a summary of our business: what we did, how we did it, and what our risk factors were.
We’d even started changing Netflix’s identity to appeal to risk-averse banks and their customers. The big trend, in the late 1990s and early 2000s, was for internet companies to be portals—entry points on the internet for a specific niche. The popular wisdom at the time was that if you wanted to be a successful website, you had to be all things to all people—if you wanted to chase money, you had to chase traffic first. That meant that Netflix couldn’t just be a rental service aimed at helping people find the DVDs they’d love—it had to be a place for movie lovers of all stripes.
The VCs on our board had told us that if we wanted to go public, we needed to think big: Movie showtimes. Movie reviews. A monthly column by Leonard Maltin, king of the video guides. Etc. We’d done all of it, but I’d been unable to shake the suspicion that we were getting distracted, salivating over dollar signs and eyeing possible valuations.
Then the bubble had burst. From its high in March, the Nasdaq stock exchange—which is where most technology companies are listed—had entered a period of steady decline, punctuated by a terrifying 25 percent drop the week of April 14. That was the exact week that we filed our S-1 with the SEC, asking permission to go public. Over the following months, as the market continued to slump, Deutsche Bank had continued to project false enthusiasm, offering steadily weakening reassurances that we would be okay.
By the fall, though, it had become obvious to everyone that the numbers we’d been excitedly bandying back and forth as the solution to all our problems—$75 million? $80 million?—had evaporated into thin air. I’d gotten the call that Deutsche Bank was pulling the offering one rainy Saturday morning in September, while I was shopping with Lorraine in Carmel. Needless to say, we didn’t buy anything.
At the time, not going public had felt like a major blow. But in retrospect, it was possibly one of the best things that ever happened to us. If we’d gone public in the fall of 2000, we would have been tied to the portal idea and to the unrealistic financial expectations that we had built around it—and that would have been a disaster. We never would have been able to make any money as an all things to all people site. Becoming a “movie portal” was the complete opposite of the Canada Principle. It didn’t allow for the rigorous focus that set us apart—and which ultimately provided us with the business model to succeed on our own terms.
We quietly started abandoning most of our portal-oriented efforts. If the people who wanted movie showtimes and in-depth critics’ reviews and top-ten lists were just the banks and their customers, and the bank didn’t want to take us public, then why keep them?
So by September, we were back to where we started. We didn’t have $75 million in our pockets, and we were losing money—lots of money. Up until now, with Reed at the helm, we’d been able to find funding fairly easily, and we had long reassured ourselves that as long as we could continue to use Silicon Valley money to support our growth, we would be okay. But in a post-bubble era, getting it from our usual VC sources
would be hard. Really hard.
I worried about how the bursting of the dot-com bubble would affect Netflix’s finances. But I have to confess that I wasn’t sad to see some sort of adjustment coming down the pike. All of the dot-com hype had seemed crazy to me. Watching the Super Bowl with Lorraine back in January, I’d kept a running tally: No fewer than sixteen companies with dot-com in their names had advertised during the game, spending upward of $2 million for each spot. That was more money out the door per spot than Netflix had spent its entire first year.
At the height of the dot-com boom, the prevailing attitude at many companies had been spend now, worry later. It had become routine for companies to spend lavishly on parties, promotions, and facilities. No one put it better than Stephan Paternot, CEO of TheGlobe.com. After his 1998 IPO, he famously said, “Got the girl. Got the money. Now I’m ready to live a disgusting, frivolous life.”
That wasn’t us. Netflix had long since outgrown its card tables and beach chairs, but we were still pretty frugal—when we moved to Los Gatos, we’d bought used cubicles and secondhand furniture. The only nod to lavish decoration was a popcorn machine in the main atrium, and even that didn’t work most of the time. I’d never understood what companies were thinking when they spent tens of thousands of dollars on carpeting or bought thousand-dollar Aeron chairs for every employee. Frankly, I don’t understand it even now.
It had been an age of decadence, in other words. And like all decadent eras, it hadn’t lasted. By the time we were headed to Alisal Ranch, there wasn’t much decadence left. Boo.com, the online clothing e-tailer, had filed for bankruptcy after spending more than $175 million in only six months. The rumor mill had Pets.com on the verge of collapse after the company spent more than $150 million in the first half of the year. Webvan’s stock would fall from $30 per share to 6 cents per share after the online grocer spent nearly a billion dollars on expansion. Drkoop.com, the online portal founded by 82-year-old retired surgeon general C. Everett Koop—which had managed to go public with not a single penny in revenue—was losing tens of millions of dollars every quarter.
That Will Never Work Page 23
