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

by Andrew Chen

when you open a social app, engaging and entertaining content is there on your feed.

It was an overnight success that took years.

The real cost of a zero is not just at the moment where it’s experienced, but rather, the lingering destructive effects afterward. Users who get “zeroed” often churn and worse, they come to believe the service isn’t reliable. It’s not possible to sustain a strong network when a large percentage of users are churning, but unfortunately, by definition new networks default to having many zeroes.

Amazingly, 95 percent of this initial cohort started to use this app every day for three hours a day.

At Tufts University, less than a year after the launch, more than 80 percent of the Greek system had signed into the service and 40 percent of the entire undergraduate population was on it.

The key to Tinder’s start was discovering a repeatable strategy that took them from USC to other colleges, then absorbing the metropolitan areas, and then country to country.

If the hard side of the network isn’t yet activated, a team can just fill in their gaps themselves, using the technique of “Flintstoning”—as Reddit did, submitting links and content until eventually adding automation and community features for scale.

There are people like Bill Gates who are at the top of the professional hierarchy. He gets more requests for intros than he can deal with, and everyone who knows Gates will be asked for intros to him. At the launch, LinkedIn wouldn’t have made sense for people like Bill Gates. But there’s a mid-tier of successful people who are still building and hustling, who get fewer requests for intros but will actually take the meetings. This middle rank of people is where LinkedIn really worked.

On the first week of LinkedIn’s launch, employees and investors of the company could invite as many people as they wanted, but you couldn’t sign up from just the website. We intentionally seeded the network with the mid-tier of successful professionals that wanted to take time to connect.

Invite-only mechanics amplify a networked product that is already useful for the first few dozen users. Past this initial set, the invites will attract a dense network that will grow and grow.

When a new product carefully curates a network, followed by implementing invites so that it can copy and paste similar networks, then it can grow to take over the market.

Invite-only is a powerful strategy. When executed well, the people in an initial atomic network become a magnet for even more users. It allows a network to copy and paste itself many times over, attracting more and more adjacent networks over time.

Just as creators of new products spend endless hours designing the experience, creators of networked products have an additional task: curating the right people so that the experience of a new member joining the community, marketplace, or other network is just right.

A good product designer wouldn’t allow a random set of feature ideas to be added to the final version of a new app, and in the same way, a mindful designer of networks wouldn’...

Come for the tool, stay for the network” is one of the most famous strategies for launching and scaling networks. Start with a great “tool”—a product experience that is useful even for one user as a utility. Then, over time, pivot the users into a series of use cases that tap into a “network”—the

RJ Metrics analyzed data from Instagram’s APIs and concluded that 65 percent of users were not yet following other people on the network. Instead, the engagement was oriented completely around photo editing, noting that “Instagram’s 2.2 million users upload 3.6 million new photos per week (or 6 photos per second).”38 In other words, Instagram was being used first as a tool—a free Hipstamatic with a better design. The network would come later.

“Come for the tool, stay for the network” circumvents the Cold Start Problem and makes it easier to launch into an entire network—with PR, paid marketing, influencers, sales, or any number of tried-and-true channels. It minimizes the size requirement of an atomic network and in turn makes it easy to take on an entire network.

Amazon too lost money in its first seventeen straight quarters.

Coupons were invented in 1888 by John Pemberton and Asa Candler, cofounders of the Coca-Cola Company. The early coupons for Coca-Cola showed the classic cursive logo in the center, with the headline “This card entitles you to one glass of Coca-Cola” and along the sides, it encouraged you to go to any dispenser to redeem the coupon. This is one of the first nationwide campaigns to brand something that historically had many regional varieties. The campaign was a huge success, and within the first two decades of its existence 8.5 million free drinks for one in nine Americans had been redeemed.

I devised a plan for making Van Camp’s Milk familiar. In a page ad, I inserted a coupon, good at any store for a ten-cent can. We paid the grocer his retail price. For three weeks we announced that this ad would appear. At the same time we told the story of Van Camp’s Evaporated Milk. We sent copies of these ads to all grocers, and told them that every customer of theirs would receive one of these coupons. It was evident that they must have Van Camp’s Milk. Every coupon meant a ten-cent sale which, if they missed it, would go to a competitor. . . . The result was almost universal distribution, and at once.41

“If you have a chicken and egg problem—buy the chicken.”

over time structures like “Do 10 trips and get an extra $1 per trip” (called internally, DxGy) were built.

Hundreds of people on the Marketplace team—composed of data scientists, economists, engineers, and others—would manage these levers to shift the balance of supply and demand in the hundreds of markets around the world.

“Flinstoning” is a metaphor for this car, except in software, where missing product functionality is replaced with manual human effort.

Early product releases often go into beta while lacking simple features like account deletion, content moderation tools, referral features, and many others. In lieu of these features, the product might simply offer a way to contact the developers who will handle it manually for you, using tools they have in the back end. Once they get enough inquiries, eventually the feature gets built out and users can do it themselves. In the meantime, a Flintstoned product launch lets the developers get the app out into the market and get feedback from customers.

The goal is just to manually fill in critical parts of the network, until it can stand on its own.

The downside of Flintstoning is that it feels like it’s overly manual. You start by throwing people at the problem, but can it scale? I argue that it scales further and longer than you might think. Flinstoning can be thought of as a spectrum:

Flintstoning can be phased out over time. In that way, it can be thought of as a close cousin to the “come for the tool, stay for the network” approach.

it’s important to have an exit strategy with Flintstoning. With the former, a networked product must be designed to switch users from single player to multiplayer mode. With the latter, the product must switch from manual (and company-supported) to automated.

When this happens, do you try to fix the loopholes or add more controls, potentially impacting the usability of the product? Is it even a good idea to scan through the contents of your users’ folders? Or do you embrace it but nudge usage in the right direction over time? These are hard questions.

Perhaps because of the success of the self-serve approach, the sales team was small. To give you a sense of the company culture, in the early days when the sales team got overloaded, it was standard procedure to simply remove the sales team’s email address from the website, so that users couldn’t contact them. Eventually, the team realized it was better to hire more people in sales. But making money was still viewed as a “why would we care about that?” topic in the geeky, MIT engineering-oriented nature of the company. Revenue wasn’t considered “Dropboxy.”

the ones who used Dropbox for collaboration and sharing—the network features—became significantly more valuable over time. Dropbox’s users could be divided into High-Value Actives (HVAs) and Low-Value Actives (LVAs), which was useful as a quality indicator.

Originally, we thought our mission was trying to serve “everyone on the internet” but we realized that we shouldn’t be fighting every war.

Dropbox, for example, employed 2,000+ full-time highly paid designers, engineers, and marketers, doubling or tripling the employee base each year leading up to the 2018 IPO.

While you may only need a small handful of employees to achieve product/market fit—famously, Instagram had thirteen employees and 30 million users when it was bought by Facebook—you need a significant coordinated effort to scale a product to its full potential.

the network effect is not one effect. Instead, the network effect is a broader umbrella term that can be broken down into a trio of underlying forces: the Acquisition network effect, the Engagement network effect, and the Economic network effect.

The “Acquisition Effect” is the ability for a product to tap into its network to acquire new customers.

Here is what’s often called the “Growth Accounting Equation,” which shows how these key metrics relate for active users: Gain or loss in active users = New + Reactivated − Churned

Then based on the delta of each time period, you can figure out if you’ll gain or lose active users: This month’s actives = Last month’s actives + gain or loss

It’s become a best practice to take this equation and build dashboards out of its inputs, so that in any given month you know how the underlying components are trending.

Overlaying revenue is easy, too. You just add two more variables, multiplying the active users number with the average revenue per active user (ARPU).

The central inputs into a networked product’s growth equation will improve on their own, as a function of the network as opposed to the features of the product—creating an accumulating advantage over time. This is the magic of network effects.

Retention is the most critical metric in understanding a product, but most of the time, the data is not pretty.

“Nearly 1 in 4 people abandon mobile apps after only one use.”

Of the users who install an app, 70 percent of them aren’t active the next day, and by the first three months, 96 percent of users are no longer active.

comScore, revealed that people spend 80 percent with just three apps

As a rough benchmark for evaluating startups at Andreessen Horowitz, I often look for a minimum baseline of 60 percent retention after day 1, 30 percent after day 7, and 15 percent at day 30, where the curve eventually levels out.

Users need to trust the loop to rely on it. If the network is too small or too inactive and the loop breaks, then users will be less likely to use it in the future. After all, if you text a friend on a new messaging app and they don’t respond, or if you share a document at work but don’t get a reply, then trust goes down. But in the positive case, if a network scales and the connections get denser, then the loop gets tighter—content

up to 75 percent of users are inactive at any given point, most of whom will never come back. The ability to reactivate users provides a powerful counterweight to churn,

One important insight is that the Acquisition Effect can exist independently of the Engagement or Economic effect. In other words, you can acquire a lot of customers but still have a network that ultimately isn’t sticky.