The Cold Start Problem: How to Start and Scale Network Effects
by Andrew Chen
30 popular highlights from this book
Key Insights & Memorable Quotes
Below are the most popular and impactful highlights and quotes from The Cold Start Problem: How to Start and Scale Network Effects:
Flinstoning” is a metaphor for this car, except in software, where missing product functionality is replaced with manual human effort.
Success comes with an inevitable problem: market saturation. New products initially grow just by adding more customers—to grow a network, add more nodes. Eventually this stops working because nearly everyone in the target market has joined the network, and there are not enough potential customers left. From here, the focus has to shift from adding new customers to layering on more services and revenue opportunities with existing ones. eBay had this problem in its early years, and had to figure its way out. My colleague at a16z, Jeff Jordan, experienced this himself, and would often write and speak about his first month as the general manager of eBay’s US business. It was in 2000, and for the first time ever, eBay’s US business failed to grow on a month-over-month basis. This was critical for eBay because nearly all the revenue and profit for the company came from the US unit—without growth in the United States, the entire business would stagnate. Something had to be done quickly. It’s tempting to just optimize the core business. After all, increasing a big revenue base even a little bit often looks more appealing than starting at zero. Bolder bets are risky. Yet because of the dynamics of market saturation, a product’s growth tends to slow down and not speed up. There’s no way around maintaining a high growth rate besides continuing to innovate. Jeff shared what the team did to find the next phase of growth for the company: eBay.com at the time enabled the community to buy and sell solely through online auctions. But auctions intimidated many prospective users who expressed preference for the ease and simplicity of fixed price formats. Interestingly, our research suggested that our online auction users were biased towards men, who relished the competitive aspect of the auction. So the first major innovation we pursued was to implement the (revolutionary!) concept of offering items for a fixed price on ebay.com, which we termed “buy-it-now.” Buy-it-now was surprisingly controversial to many in both the eBay community and in eBay headquarters. But we swallowed hard, took the risk and launched the feature . . . and it paid off big. These days, the buy-it-now format represents over $40 billion of annual Gross Merchandise Volume for eBay, 62% of their total.65
Launching “Buy It Now” was a large change that touched every transaction, but the eBay team also innovated across the experience for both sellers and buyers as well. With an initial success, we doubled down on innovation to drive growth. We introduced stores on eBay, which dramatically increased the amount of product offered for sale on the platform. We expanded the menu of optional features that sellers could purchase to better highlight their listings on the site. We improved the post-transaction experience on ebay.com by significantly improving the “checkout” flow, including the eventual seamless integration of PayPal on the eBay site. Each of these innovations supported the growth of the business and helped to keep that gravity at bay. Years later, Jeff became a general partner at Andreessen Horowitz, where he would kick off the firm’s success in startups with network effects, investing in Airbnb, Instacart, Pinterest, and others. I’m lucky to work with him! He recounted in an essay on the a16z blog that his strategy was to grow eBay by adding layers and layers of new revenue—like “adding layers to the cake.” You can see it visually here: Figure 12: eBay’s growth layer cake As the core US business began to look more like a line than a hockey stick, international and payments were layered on top. Together, the aggregate business started to look like a hockey stick, but underneath it was actually many new lines of business.
acquiring the hard side of the network and keeping them happy is paramount to standing up an atomic network.
Once Hopkins showed that this worked in creating one atomic network, the effort could be repeated in building the second, third, and so on: We proved out this plan in several cities of moderate size. Then we undertook New York City. There the market was dominated by a rival brand. Van Camp had slight distribution. In three weeks we secured, largely by letter, 97 per cent distribution. Every grocer saw the necessity of being prepared for that coupon demand. Then one Sunday in a page ad, we inserted the coupon. This just in Greater New York. As a result of that ad, 1,460,000 coupons were presented. We paid $146,000 to the grocers to redeem them. But 1,460,000 homes were trying Van Camp’s Milk after reading our story, and all in a single day. The total cost of that enterprise, including the advertising, was $175,000, mostly spent in redeeming those coupons. In less than nine months that cost came back with a profit. We captured the New York market.
The other question to ask is, if a user wants to reactivate, how hard is it? At Uber, we had a staggering statistic where several million users were failing their password recovery per week—how do you make this much easier, and treat reactivation with the same seriousness as the sign-up process? While reactivation is typically not a concern for new products—they should focus on new users, since their count of lapsed users won’t be large—for products that have hit Escape Velocity, there will be a pool of many millions of users to draw upon. Reengaging them can become as big a growth lever as acquiring new users.
This is an important tool that is unique to networked products. Traditional products that lack networks often struggle with this, because they rely on spammy emails, discounts, and push notifications to entice users back. This usually doesn’t work, and company-sent communications rank among the lowest clickthrough rate messages. Networked products, on the other hand, have the unique capability to reactivate these users by enlisting active users to bring them back. Even if you don’t open the app on a given day, other users in the network may interact with you—commenting or liking your past content, or sending you a message. Getting an email notification that says your boss just shared a folder with you is a lot more compelling than a marketing message. A notification that a close friend just joined an app you tried a month ago is a lot more engaging than an announcement about new features. And the more dense the network is around a churned user, the more likely they are to receive this type of interaction.
The network effect version of this in the technology industry happens when there is “overcrowding” from too many users. For communication apps, you might start to get too many messages. For social products, there might be too much content in feeds, or for marketplaces, too many listings so that finding the right thing becomes a chore. If you don’t apply spam detection, algorithmic feeds, and other ideas, quickly the network becomes unusable. But add the right features to aid discovery, combat spam, and increase relevance within the UI, and you can increase the carrying capacity for users.
successful network effect requires both a product and its network
A telephone without a connection at the other end of the line is not even a toy or a scientific instrument. It is one of the most useless things in the world. Its value depends on the connection with the other telephone and increases with the number of connections.
Tinder would work with Justin’s younger brother to throw a birthday party for one of his popular, hyperconnected friends on campus, and use it to promote Tinder. The Tinder team would do all the work to make it an incredible party. The day of the party, students from USC were getting bused to a luxurious house in LA, where everything had been set up to pull you inside. Sean described how it worked: There was one catch with the party: First, you had to download the Tinder app to get in. We put a bouncer in the house to check that you had done it. The party was great—it was a success, and more importantly, the next day, everyone at the party woke up and remembered they had a new app on their phone. There were attractive people they hadn’t gotten to talk to, and this was their second chance. The college party launch tactic worked. For the Tinder team, this one party created the highest ever one-day spike of downloads, however modest it might seem in retrospect. It’s not just the number that matters here, but that it was “500 of the right people”—Sean would explain to me later. It was a group of the most social, most hyperconnected people on the USC campus, all on Tinder at the same time. Tinder started to work. Matches began to happen, as the students who met each other from the previous night started to swipe through and then chat. Amazingly, 95 percent of this initial cohort started to use this app every day for three hours a day. The Tinder team built one atomic network, but soon figured out how to build the next one—just throw another party. And then another, by going to other schools, and throwing even more parties. Each network was successively easier to start. Tinder quickly reached 4,000 downloads, then 15,000 within a month, and then 500,000 just a month after that—first by replicating the campus launch, but then letting the organic viral growth take over.
Tinder would work with Justin’s younger brother to throw a birthday party for one of his popular, hyperconnected friends on campus, and use it to promote Tinder. The Tinder team would do all the work to make it an incredible party. The day of the party, students from USC were getting bused to a luxurious house in LA, where everything had been set up to pull you inside. Sean described how it worked: There was one catch with the party: First, you had to download the Tinder app to get in. We put a bouncer in the house to check that you had done it. The party was great—it was a success, and more importantly, the next day, everyone at the party woke up and remembered they had a new app on their phone. There were attractive people they hadn’t gotten to talk to, and this was their second chance. The college party launch tactic worked. For the Tinder team, this one party created the highest ever one-day spike of downloads, however modest it might seem in retrospect. It’s not just the number that matters here, but that it was “500 of the right people”—Sean would explain to me later. It was a group of the most social, most hyperconnected people on the USC campus, all on Tinder at the same time. Tinder started to work. Matches began to happen, as the students who met each other from the previous night started to swipe through and then chat. Amazingly, 95 percent of this initial cohort started to use this app every day for three hours a day.
Uber’s app was initially outsourced to Mexico, so that when later engineers joined the company, they needed to be issued Spanish-to-English dictionaries to understand the comments and source code. In these cases it isn’t until later, as the product hits scale, that the engineering teams are upgraded.
Sometimes the army is built on people with excess time, but sometimes it is built on people with underutilized assets as well.
Uber had to get creative to unlock the hard side of their network, the drivers. Initially, Uber’s focus was on black car and limo services, which were licensed and relatively uncontroversial. However, a seismic shift occurred when rival app Sidecar innovated in recruiting unlicensed, normal people as drivers on their platform. This was the “peer-to-peer” model that created millions of new rideshare drivers, and was quickly copied and popularized by Lyft and then Uber. Jahan Khanna, cofounder/chief technology officer of Sidecar, spoke of its origin: It was obvious that letting anyone sign up to be a driver would be a big deal. With more drivers, rides would get cheaper and the wait times would get shorter. This came up in many brainstorms at Sidecar, but the question was always, what was the regulatory framework that allows this to operate? What were the prior examples that weren’t immediately shut down? After doing a ton of research, we came onto a model that had been active for years in San Francisco run by someone named Lynn Breedlove called Homobiles that answered our question.22 It’s a surprising fact, but the earliest version of the rideshare idea came not from an investor-backed startup, but rather from a nonprofit called Homobiles, run by a prominent member of the LGBTQ community in the Bay Area named Lynn Breedlove. The service was aimed at protecting and serving the LGBTQ community while providing them transportation—to conferences, bars and entertainment, and also to get health care—while emphasizing safety and community. Homobiles had built its own niche, and had figured out the basics: Breedlove had recruited, over time, 100 volunteer drivers, who would respond to text messages. Money would be exchanged, but in the form of donations, so that drivers could be compensated for their time. The company had operated for several years, starting in 2010—several years before Uber X—and provided the template for what would become a $100 billion+ gross revenue industry. Sidecar learned from Homobiles, implementing their offering nearly verbatim, albeit in digital form: donations based, where the rider and driver would sit together in the front, like a friend giving you a ride. With that, the rideshare market was kicked off.
The networked product should be launched in its simplest possible form—not fully featured—so that it has a dead simple value proposition. The target should be on building a tiny, atomic network—the smallest that could possibly make sense—and focus on building density, ignoring the objection of “market size.” And finally, the attitude in executing the launch should be “do whatever it takes”—even if it’s unscalable or unprofitable—to get momentum, without worrying about how to scale.
The solution to the Cold Start Problem starts by understanding how to add a small group of the right people, at the same time, using the product in the right way. Getting this initial network off the ground is the key, and the key is the “atomic network”—the smallest, stable network from which all other networks can be built.
Small, sub-scale networks naturally want to self-destruct, because when people show up to a product and none of their friends or coworkers are using it, they will naturally leave. What solves this? “The Atomic Network”—the smallest network where there are enough people that everyone will stick around.
Bundling eventually stopped working for Microsoft. After the antitrust investigation, the company maintained its dominance on the PC operating systems market, but it lost control of many other markets. Eventually the industry jumped from PC to mobile. Microsoft tried to exactly replicate the network effects it had before—an ecosystem of hardware manufacturers who paid a licensing fee to run Windows Mobile, and app developers and consumers to match—but this time it didn’t work. Instead, Google gave away its Android mobile OS for free, driving adoption for phone makers. The massive reach of Android attracted app developers, and a new network effect was built, derived from a business model where the OS was free but the ecosystem was monetized using search and advertising revenue. Microsoft has also lost the browser market to Google Chrome, and is being challenged in its Office Suite by a litany of startup competitors large and small. It continued to use bundling as a strategy, adding workplace chat via Teams to its suite—but it hasn’t achieved a clear victory against Slack. If bundling hasn’t been a sure thing for Microsoft, it’s an even weaker strategy for others. The outcome seems even less assured when examining how Google bundled Google+ into many corners of its product, including Maps and Gmail, achieving hundreds of millions of active users without real retention. Uber bundled Uber Eats across many touchpoints within its rideshare app, but still fell behind in food delivery versus DoorDash. Bundling hasn’t been a silver bullet, as much as the giants in the industry hope it is.
Contrast this with the teams that eventually succeeded in competing with Facebook where Google+ failed. Snap famously grew within the high school segment before breaking out into the mainstream, and the ephemeral photos captured a whole unique set of content that had never been published—casual, unposed photos that were meant for communication. Early on, with fewer than 10,000 daily active users, Snapchat was already hitting 10 photos/day/user, several orders of magnitude more than equivalent services—showing it had mastered the hard side of the network. Twitch, Instagram, and TikTok innovated in a similar vector, giving creators new tools and media types to express themselves.
Anti-Network Effects Hit the Google+ Launch A charismatic executive from one of the most powerful technology companies in the world introduces a new product at a conference. This time, it’s June 2011 at the Web 2.0 Summit, where Google vice president Vic Gundotra describes the future of social networking and launches Google+. This was Google’s ambitious strategy to counteract Facebook, which was nearing their IPO. To give their new networked product a leg up, as many companies do, it led with aggressive upsells from their core product. The Google.com homepage linked to Google+, and they also integrated it widely within YouTube, Photos, and the rest of the product ecosystem. This generated huge initial numbers—within months, the company announced it had signed up more than 90 million users. While this might superficially look like a large user base, it actually consisted of many weak networks that weren’t engaged, because most new users showed up and tried out the product as they read about it in the press, rather than hearing from their friends. The high churn in the product was covered up by the incredible fire hose of traffic that the rest of Google’s network generated. Even though it wasn’t working, the numbers kept going up. When unengaged users interact with a networked product that hasn’t yet gelled into a stable, atomic network, then they don’t end up pulling other users into the product. In a Wall Street Journal article by Amir Efrati, Google+ was described as a ghost town even while the executives touted large top-line numbers: To hear Google Inc. Chief Executive Larry Page tell it, Google+ has become a robust competitor in the social networking space, with 90 million users registering since its June launch. But those numbers mask what’s really going on at Google+. It turns out Google+ is a virtual ghost town compared with the site of rival Facebook Inc., which is preparing for a massive initial public offering. New data from research firm comScore Inc. shows that Google+ users are signing up—but then not doing much there. Visitors using personal computers spent an average of about three minutes a month on Google+ between September and January, versus six to seven hours on Facebook each month over the same period, according to comScore, which didn’t have data on mobile usage.86 The fate of Google+ was sealed in their go-to-market strategy. By launching big rather than focusing on small, atomic networks that could grow on their own, the teams fell victim to big vanity metrics. At its peak, Google+ claimed to have 300 million active users—by the top-line metrics, it was on its way to success. But network effects rely on the quality of the growth and not just its quantity
When examined through the lens of Meerkat’s Law and the central framework of this book, it is obvious why the resulting networks generated by big launches are weak. You’d rather have a smaller set of atomic networks that are denser and more engaged than a large number of networks that aren’t there. When a networked product depends on having other people in order to be useful, it’s better to ignore the top-line aggregate numbers. Instead, the quality of the traction can only be seen when you zoom all the way into the perspective of an individual user within the network. Does a new person who joins the product see value based on how many other users are already on it? You might as well ignore the aggregate numbers, and in particular the spike of users that a new product might see in its first days. As Eric Ries describes in his book The Lean Startup, these are “vanity metrics.” The numbers might make you feel good, especially when they are going up, but it doesn’t matter if you have a hundred million users if they are churning out at a high rate, due to a lack of other users engaging. When networks are built bottom-up, they are more likely to be densely interconnected, and thus healthier and more engaged. There are multiple reasons for this: A new product is often incubated within a subcommunity, whether that’s a college campus, San Francisco techies, gamers, or freelancers—as recent tech successes have shown. It will grow within this group before spreading into other verticals, allowing time for its developers to tune features like inviting or sharing, while honing the core value proposition. Once a new networked product is spreading via word of mouth, then each user is likely to know at least one other user already on the network. By the time it reaches the broader consciousness, it will be seen as a phenomenon, and top-down efforts can always be added on to scale a network that’s already big and engaged. If Big Bang Launches work so poorly in general, why do they work for Apple? This type of launch works for Apple because their core offerings can stand alone as premium, high-utility products that generally don’t need to construct new networks to function. At most, they tap into existing networks like email and SMS. Famously, Apple has not succeeded with social offerings like the now-defunct Game Center and Ping. The closest new networked product they’ve launched is arguably the App Store, but even that was initially not in Steve Jobs’s vision for the phone.87 Most important, though, you aren’t Apple. So don’t try to copy them without having their kinds of products.
Finding the Competitive Levers When there’s a battle between two networks, there are competitive levers that shift users from one into the other—what are they? The best place to focus in the rideshare market was the hard side of the network: drivers. More drivers meant that prices would be lower, attracting valuable high-frequency riders that often comparison shop for fares. Attract more riders, and it more efficiently fills the time of drivers, and vice versa. There was a double benefit to moving drivers from a competitor’s network to yours—it would push their network into surging prices while yours would lower in price. Uber’s competitive levers would combine financial incentives—paying up for more sign-ups, more hours—with product improvements to improve Acquisition, Engagement, and Economic forces. Drawing in more drivers through product improvements is straightforward—the better the experience of picking up riders and routing the car to their destination, the more the app would be used. Building a better product is one of the classic levers in the tech industry, but Uber focused much of its effort on targeted bonuses for drivers. Why bonuses? Because for drivers, that was their primary motivation for using the app, and improving their earnings would make them sticky. But these bonuses weren’t just any bonuses—they were targeted at quickly flipping over the most valuable drivers in the networks of Uber’s rivals, targeting so-called dual apping drivers that were active on multiple networks. They were given large, special bonuses that compelled them to stick to Uber, and every hour they drove was an hour that the other networks couldn’t utilize. There was a sophisticated effort to tag drivers as dual appers. Some of these efforts were just manual—Uber employees who took trips would just ask if the drivers drove for other services, and they could mark them manually in a special UI within the app. There were also behavioral signals when drivers were running two apps—they would often pause their Uber session for a few minutes while they drove for another company, then unpause it. On Android, there were direct APIs that could tell if someone was running Uber and Lyft at the same time. Eventually a large number of these signals were fed into a machine learning model where each driver would receive a score based on how likely they were to be a dual apper. It didn’t have to be perfect, just good enough to aid the targeting.
To learn why bundling sometimes works, and other times doesn’t, I went to the source. I asked Brad Silverberg, who in his decade at Microsoft headed up some of the company’s most important product efforts—including the much-celebrated release of Windows 95, accelerating the franchise from $50 million to $3.5 billion, as well as all the early releases of Internet Explorer. He’s been a mentor of mine for years, having served on the board of a startup I founded years back. I interviewed Brad for The Cold Start Problem over videoconference; he was mostly retired and spending time with family in Jackson Hole, Wyoming. But his experience from the 1980s and ’90s has made him the definitive authority on this topic, and perhaps surprisingly, he’s skeptical of the power of bundling: Bundling a product is not the silver bullet everyone thinks. If it were that easy, the version 1.0 for Internet Explorer would have won, by simply bundling it with Windows. It didn’t—IE 1.0 only got to 3% or 4% market share, because it just wasn’t good enough yet. Bing is another example, when Microsoft wanted to get into search. It was the default search engine across the operating system, not just in Internet Explorer but also MSN and everywhere Microsoft could jam it. But it went nowhere. The distribution advantages don’t win when the product is inferior.91 Even if bundling gets you a lot of new users trying out a product, they won’t stick around if there’s a huge gap in features.
For Microsoft’s productivity applications, the break came when the world transitioned from text-based DOS applications to graphical user interfaces, in the mid-1980s. But as the industry shifted from text to graphical interfaces, it created an opening, as every application needed to be rewritten to support the new paradigm of dropdown menus, icons, toolbars, and the mouse. While Microsoft redesigned and rethought their applications, their competitors were too stuck in the old world, and so Word and Excel leapfrogged their competitors. Then in an ensuing stroke of product marketing genius, it was combined into the Microsoft Office suite, which promptly became a colossus. Much effort was put toward making each application within the suite work with each other. For example, an Excel chart would be embedded within a Microsoft Word document—this was called Object Linking and Embedding (OLE)—which made the combination of the products more powerful. In other words, the product really matters, and bundling can provide a huge distribution advantage, but it can only go so far. It’s an echo of what we now see in the internet age, where Twitter might drive users to its now-defunct livestreaming platform Periscope, or Google might push everyone to use Google Meet. It can work, but only when the product is great. This is part of why the concept of bundling as been around forever—the McDonald’s Happy Meal was launched in the 1970s, and cable companies have been bundling TV channels since their start. But at the heart of these bundling stories are important, iconic products that reinvent the market.
Over the years, Facebook has executed an effective playbook that does exactly this, at scale. Take Instagram as an example—in the early days, the core product tapped into Facebook’s network by making it easy to share photos from one product to the other. This creates a viral loop that drives new users, but engagement, too, when likes and comments appear on both services. Being able to sign up to Instagram using your Facebook account also increases conversion rate, which creates a frictionless experience while simultaneously setting up integrations later in the experience. A direct approach to tying together the networks relies on using the very established social graph of Facebook to create more engagement. Bangaly Kaba, formerly head of growth at Instagram, describes how Instagram built off the network of its larger parent: Tapping into Facebook’s social graph became very powerful when we realized that following your real friends and having an audience of real friends was the most important factor for long-term retention. Facebook has a very rich social graph with not only address books but also years of friend interaction data. Using that info supercharged our ability to recommend the most relevant, real-life friends within the Instagram app in a way we couldn’t before, which boosted retention in a big way. The previous theory had been that getting users to follow celebrities and influencers was the most impactful action, but this was much better—the influencers rarely followed back and engaged with a new user’s content. Your friends would do that, bringing you back to the app, and we wouldn’t have been able to create this feature without Facebook’s network. Rather than using Facebook only as a source of new users, Instagram was able to use its larger parent to build stronger, denser networks. This is the foundation for stronger network effects. Instagram is a great example of bundling done well, and why a networked product that launches another networked product is at a huge advantage. The goal is to compete not just on features or product, but to always be the “big guy” in a competitive situation—to bring your bigger network as a competitive weapon, which in turn unlocks benefits for acquisition, engagement, and monetization. Going back to Microsoft, part of their competitive magic came when they could bring their entire ecosystem—developers, customers, PC makers, and others—to compete at multiple levels, not just on building more features. And the most important part of this ecosystem was the developers.
Most products these days are low technical risk—meaning they won’t fail because the teams can’t execute on the engineering side to build the products—but they are generally also low defensibility. When something works, others can follow—and fast.
First, does the product have a network? Does it connect people with each other, whether for commerce, collaboration, communication, or something else at the core of the experience? And second, does the ability to attract new users, or to become stickier, or to monetize, become even stronger as its network grows larger?
The “effect” part of the network effect describes how value increases as more people start using the product.
Larger competitors are often able to copy the product, but find it difficult to capture the network.
