The Cold Start Problem: How to Start and Scale Network Effects

by Andrew Chen

like a simple app—after all, the premise was always to hit a button and get a ride. But underneath its deceptively basic user interface was a complex, global operation required to sustain the business. The app sat on a vast worldwide network of smaller networks, each one representing cities and countries. Each of these networks had to be started, scaled, and defended against competitors, at all hours of the day.

A Network of Networks It was my first year at the company, and although many companies have weekly reviews, Uber’s were different. First, in the discussion about each city, the level of detail surprised me. For San Francisco, the group began to discuss the surge percentages in the city’s seven-mile by seven-mile center, versus East Bay, versus the Peninsula. This was a senior group of executives, but the granularity and level of detail was incredible. But this was a requirement to run a complex, hyperlocal network like Uber where supply and demand depended on the dynamics of popular neighborhoods and frequent “lanes”—like Marina and the Financial District—that tended to be poorly served by other transportation options.

In the weekly dashboard, each row represented a city—yes—but more important, each city was an individual network in Uber’s global network of networks that needed to be nurtured, protected, and grown. It was deeply and uniquely ingrained in Uber’s DNA to talk about metrics at the hyperlocal network level. In my several years there, it was unusual to ever hear about an aggregate number—like total trips or total active riders—except as a big vanity milestone at a company all-hands. Those aggregate metrics were regarded as mostly meaningless.

What are network effects, really? How do they apply to your business? How do you know if your product has them—and which other products don’t? Why are they so hard to create, and how do you create them? Can you add a network to your product after the fact? How do they impact your business metrics, at the tactical level? Is Metcalfe’s Law actually right, or should you apply something else to your strategy?

A successful network effect requires both a product and its network, and that was true in the age of the American Telephone & Telegraph Company, and true today. For Uber, the “product” is the app that people run on their phones, and the “network” refers to all the active users at any given time who are connecting with Uber to drive or ride. (There’s no physical wiring, in this case!) In contemporary parlance, the product is typically made of software whereas the network is typically made of people.

The Allee effect → The Network Effect Allee Threshold → Tipping Point Carrying capacity → Saturation

With these key features, “Slack” was picked to mean “Searchable Log of All Conversation and Knowledge.”

Each of these beta customers formed an atomic network—a stable, self-sustaining group of users who can drive a network effect. Once an atomic network was formed in one of their beta testers, Slack would continually add users, become more useful, ramp up engagement, and ultimately become the de facto method of communication within their workplaces. The minimum number of people to be defined as a team, even today at Slack, is three. As long as you have three people, it can be stable. But it’s even better if there’s a larger team of fifty people in an organic unit—like a department—or an entire company of thousands that joins the product. Then it will just grow and grow.

Like many of the examples I’ll visit later on, Slack is a network of networks.

In later years, when the company added an enterprise sales team, they would target large customers who had adopted Slack across multiple parts of their company and ask if they wanted to go “wall-to-wall,” bringing the entire company onboard in exchange for added security/enterprise features as well as company-wide pricing. Enterprise sales became a huge accelerator a few years into the company’s launch.

Eventually, Tiny Speck rebranded as Slack Technologies, and its new flagship product was launched to the world. In August 2013, when Slack debuted, 8,000 companies signed up to the wait list to try the product. Within two weeks, this grew to 15,000. By the next year, Slack would have 135,0000 paying subscribers and 10,000 new users signing up per day. Soon after that, a million daily active users, then two, then three, and so on. Stewart and the company—now 100 percent focused on Slack—would raise their next round in April 2014, fully completing their transition.

It’s a myth that network effects are all powerful and positive forces—quite the opposite. 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.

These networks often have “sides,” whether they are buyers and sellers, or content creators and consumers. Generally one side of the network will be easier to attract—this is the easy side of the network. However, the most important part of any early network is attracting and retaining “The Hard Side” of a network—the small percentage of people that typically end up doing most of the work within the community. For example, most of Wikipedia was written by a tiny percentage of prolific editors. A small group of drivers, about 5 percent of Uber’s users, carry most of the load within the rideshare marketplace—riders are numerous but engage less frequently and deeply. To attract the hard side, you need to “Solve a Hard Problem”—design a product that is sufficiently compelling to the key subset of your network. Tinder did this for the most attractive users in its network, which I’ll use as an example.

For Uber, you can’t use the service until there are enough drivers, who won’t drive until there are enough rides. This first stage of creating a network is the hardest, and some will call it a “chicken and egg” situation or needing to “bootstrap” a community. I call it the “Cold Start Problem.”

The Cold Start Problem is the first challenge that new networks must face. Beating this initial challenge is very hard. Just look at the metrics: there are tens of thousands of social, communications, and marketplace startups but just a few dozen ones that have become large, independent companies. It’s already hard to start a new company, but there is even more difficulty within the winner-take-all dynamics of these product categories.

“Based on our experience of which companies stuck with us and which didn’t, we decided that any team that has exchanged 2,000 messages in its history has tried Slack—really tried it,” Butterfield says. “For a team around 50 people that means about 10 hours’ worth of messages. For a typical team of 10 people, that’s maybe a week’s worth of messages. But it hit us that, regardless of any other factor, after 2,000 messages, 93% of those customers are still using Slack today.”9

How many users does your network need before the product experience becomes good? The way to answer this is for companies to do an analysis on the size of their networks (on the X-axis) plotted against a set of important engagement metrics (on the Y-axis). For Uber, this chart showed that more drivers generally meant a lower waiting time, and thus more users—at least up to some point. Eventually, the difference between two minutes to get a car or one minute to get a car becomes diminishing returns. Facebook’s famous growth maxim, “10 friends in 7 days,” is an expression of the same idea.

Cofounder Nate Blecharczyk is highly quantitative and had determined that 300 listings, with 100 reviewed listings, was the magic number to see growth take off in a market.11

Get ETAs down under 3 min on average as quickly as possible while covering the whole city.

Communication apps can be 1:1, so the network is small, and you can plan accordingly. Contrast that to products that are highly asymmetrical, with content creators and viewers, or marketplaces with buyers and sellers—these are likely to require a much bigger number to hit the threshold, and require a much bigger effort to get started. The size of an initial network helps determine a launch strategy.

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.

If you study the launch of products with network effects, you’ll see that one of the most common threads is that they often start small, in a single city, college campus, or in small beta tests at individual companies—like Slack’s story. Only once they nail it in a smaller network do they build up over time to eventually conquer the world. If you can create one stable, engaged network that can self-sustain—an atomic network—then likely you can build a second network adjacent to the first one. And if you can build 1, and then 2, you can probably build 10 or 100 networks. Copy and paste many times, and you can build a huge interconnected network that spans the entire market.

Credit cards have network effects for the same reasons that marketplaces do: they aggregate consumers, merchants, and other financial institutions as a multi-sided network. Everyone in the network benefits, particularly the consumer, who can go shopping without carrying physical cash. Merchants and banks are happy, too. And the bigger the network gets—meaning more consumers, more places where credit cards are accepted, etc.—the more useful the network. This in turn drives new merchants and consumers to adopt it. Bank of America invented the credit card, and picked Fresno, California, as its first test site. It would just focus on this single city—why? Fresno was chosen as the first test site partly because of its size—with a population around 250,000, it offered the critical mass the bank thought necessary to make a credit card work—and partly because a staggering 45 percent of Fresno’s families did some business with the Bank of America.15

On September 18, the bank mailed 60,000 Fresno residents a BankAmericard. There was no application process. The card simply arrived in the mailbox, ready to use. Credit card fees for merchants were set at 6 percent and consumers received between $300 and $500 in instant credit. There was a certain brilliance behind the 60,000-person drop: On day 1, cardholders simply existed. This permitted Bank of America to sign up all merchants who didn’t already have proprietary credit card programs. BofA focused on fast-moving, small merchants like Florsheim Shoes, not giants like Sears. More than 300 Fresno merchants signed up. Within three months, Bank of America had expanded its customer base to Modesto in the north and Bakersfield to the south. Within a year, it added San Francisco, Sacramento, and Los Angeles. Thirteen months after the initial Fresno drop, the bank had issued 2 million cards and onboarded 20,000 merchants.

Whether for credit cards, multiplayer games, or business collaboration software, the “atomic network” is the smallest network needed that can stand on its own. It needs to have enough density and stability to break through early anti-network effects, and ultimately grow on its own.

In Slack’s case, it was their incredible buzz within the early-adopter startup community, and their invite-only launch. There are other famous examples: The $5 referral fee that lit up the original PayPal network, or Dropbox’s demo video on Hacker News that created a huge line of folks excited to try the magical cloud-storage product. Or the “Uber Ice Cream” promotions letting people order soft-serve ice cream, on demand, via the rideshare app. In the early days, local newspapers and social media would often cover the ice cream promotion, which helped build out the rideshare network. Each of these growth hacks gave an important, quick lift that established an atomic network, kicking off future growth.

Networked products often look like toys, and moreover, toys for a weird niche. This is why they are easy to underestimate. The atomic networks are forming in niche audiences like teens or gamers, and picking up a lot of buzz, but it’s not clear yet that it will be interesting to the mainstream. But that may soon change, as the network gets built out. In the meantime, you’re not in the target audience, and that’s okay.

The first step to launching an atomic network is to have a hypothesis about what it might look like. My advice: Your product’s first atomic network is probably smaller and more specific than you think. Not a massive segment of users, or a particular customer segment, or a city, but instead something tiny, maybe on the order of hundreds of people, at a specific moment in time.

It should be about building the smallest possible group.

The more users you need to get to an atomic network, the harder it is to create.

The Power of Atomic Networks Growing city by city, campus by campus, or team by team is a surprisingly powerful strategy. It leads to dense, organic connections of users that strengthen network effects across multiple dimensions: Engagement goes up, because users are more likely to find other relevant users. Viral growth goes up when prospective users of a product see that their friends and colleagues are all using the service. In the earlier example of Bank of America’s credit card launch, focusing on one city at a time made it more likely that customers would find local merchants that accepted this novel new method of payment. Contrast that to when you peanut-butter your efforts across a whole industry or geography—the active parts of the network rapidly dissipate as anti-network effects kick in, because a network of 1,000 random users of Slack will have less retention than 1,000 users all inside the same company.

Even at the start of an atomic network, there is an important and surprising dynamic at play that only increases over time: there is a minority of users that create disproportionate value and as a result, have disproportionate power.

This the “hard side” of your network. They do more work and contribute more to your network, but are that much harder to acquire and retain. For social networks, these are often the content creators that generate the media everyone consumes. For app stores, these are the developers that actually create the products. For workplace apps, these are the managers that author and create documents and projects, and who invite coworkers to participate. For marketplaces, these are usually the sellers and providers who spend their entire day attracting users with their products and services.

Needless to say, acquiring the hard side of the network and keeping them happy is paramount to standing up an atomic network. To understand the motivations of these users, let’s describe one of the largest networked products ever built—Wikipedia.

Why is there a hard side at all? Hard sides exist because there are tasks in any networked product that just require more work, whether that’s selling products, organizing projects, or creating content. Users on the hard side have complex workflows, expect status benefits as well as financial outcomes, and will try competitive products to compare. As a result, their expectations are higher, and it’s difficult to engage and retain them.

Because the hard side is so critical, it is imperative to have hypotheses about how a product will cater to these users from day one. A successful new product should be able to answer detailed questions: Who is the hard side of your network, and how will they use the product? What is the unique value proposition to the hard side? (And in turn, the easy side of the network.) How do they first hear about the app, and in what context? For users on the hard side, as the network grows, why will they come back more frequently and become more engaged? What makes them sticky to your network such that when a new network emerges, they will retain on your product? These are difficult answers, and require a deep understanding of the motivations of your users. The motivations of the hard side depend on the product category—content creators have different goals than marketplace sellers. Those who use workplace collaboration tools have yet another set of motivations. Understanding these diverse points of view makes it easier to serve them.

This is often called the “1/10/100” rule, and it’s no surprise that the 1 percent of highly engaged users is extremely valuable.

The social feedback loop is a core concept because the creator/viewer network is so ubiquitous as a network structure. We also see this structure in categories beyond content sharing. We can see these motivations among individuals who organize group chats in WhatsApp, events on platforms like Eventbrite, writers who publish email newsletters, reviewers who love to curate their favorite restaurants on Yelp, and so on. Within these platforms, just ask yourself, “If a piece of content was created, and no one saw it, would the creator be disappointed?” If the answer is yes, then social feedback is a key value. The combination of tools, aggregation of audience, and a networked product is what is needed to unlock the hard side of these networks—it’s all about the content creators.

The hard side of a network is important to understand, not just for Wikipedia, but for any new product seeking to launch its atomic network. Without this critical group, an atomic network will struggle to get off the ground. You might go so far as to say they are the most important group of users to start with, and it’s important to have a thesis for why your product will appeal to them starting on day one.

Whereas men tend to swipe right (that is, to indicate interest) on about half of women’s profiles—about 45 percent to be exact—the ladies in the product swipe on only 5 percent of profiles they see.

Recently, the opposite trend has emerged—products like Uber and email company Superhuman, have started at the top of the market as a luxury product, and worked their way down.

Zoom didn’t have more features per se, but in fact had the most important feature of all: the “it works” feature. Zoom’s value proposition reinforced the network effects within a team and between companies, by enabling frictionless meetings. It allowed attendees to join with a single click of a link, rather than entering meeting codes and dialing numbers.

And the most important features in networked products often revolve around how users find and connect with each other, whether that’s photo tagging, sharing permissions, or “People You May Know.”

When the product concept and value is simple to describe, it makes them easier to spread from user to user, much like the “meme” coined by noted biologist Richard Dawkins in one of my favorite books, The Selfish Gene. You can copy and paste a Zoom link—it’s just that easy.

Yet many early investors balked at the fact that Zoom was tackling what they perceived as a solved problem. The commonality between all these networked products is that they offered novel ways for people to interact, and over time, the network became the defensibility.

In other words, the ideal product to drive network effects combines both factors: The product idea itself should be as simple as possible—easily understandable by anyone as soon as they encounter it. And at the same time, it should simultaneously bring together a rich, complex, infinite network of users that is impossible to copy by competitors.

I wanted Zoom to be free, at least for the basic experience, so that people could see why it was so much better. I first thought, maybe it should be limited based on participants. Maybe 3 attendees could join, but once there were 4, you’d have to pay. But that didn’t feel right. I studied Dropbox’s pricing strategy and wondered, why did they start charging at 2 gigabytes instead of 1? As I thought about it, I realized, it gave you time to use Dropbox more, and the more you used it more likely you will hit the cap and starting paying. I wanted Zoom to be the same way, so I set the limit to be 40 minutes per meeting but you could get the full experience of the product. That way, if the quality was good and you liked it, then you’d eventually pay. Even from the early days, these key product decisions made a big impact on Zoom’s trajectory. The product combined an easy-to-use app that allowed meetings to form quickly, had a value proposition that was simple to describe, and was free to use, at least initially.

Several other colleges in the area were the next set of customers. The product engaged customers and they would reach out wanting to pay. Ever since then, there’s been a lot of leads. I didn’t have a marketing team for the first 4 years! This strategy of making Zoom a freemium business meant it was easier for the network to grow.

This is a recurring theme across business models because of the Cold Start Problem. Charging customers directly is a straightforward way to generate revenue, but it adds friction for every new user to join the network. It’s hard enough to build an atomic network; why make it even harder by erecting barriers? Without being able to build out a network quickly, growth channels like virality are more muted.

Freemium is an essential part of Zoom’s appeal and ability to grow. And once Zoom had the right ingredients—a simple, killer product—its business model then provided the revenue and business model to leverage viral growth.