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

by Andrew Chen

The “Economic Effect” is the ability for a networked product to accelerate its monetization, reduce its costs, and otherwise improve its business model, as its network grows. Workplace products, for example, often convert to higher tiers of pricing as the number of knowledge workers using them grows within a company.

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. If your goal is to grow 3x year over year, and sign-ups are way down, then it becomes clear how much churn has to be improved in order to still make the target—it’s just some simple math. Overlaying revenue is easy, too. You just add two more variables, multiplying the active users number with the average revenue per active user (ARPU).

Are newer users having a better experience over the first few weeks, compared to an older cohort that was using a buggier version of the product? These graphs—often called “cohort retention curves”—are the foundational method for understanding whether a product is working or not.

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. It’s usually only the networked products that can exceed these numbers. That’s because networked products are unique in that they often become stickier over time, which cancels out the inevitable customer churn.

Their unique ability to tap into the Engagement network effect lets them drive up retention over time—first, by creating new use cases as the network develops; then by reinforcing the core “loop” of the product; and lastly, by reactivating churned users. I’ll unpack how these levers work.

What starts as infrequent and noncommittal usage often deepens into daily usage. Luckily, nudging users into more frequent usage can be part of the product design. The key is to target relevant users with messaging or incentives, or otherwise to try out new use cases over time. This moves them from low engagement into high engagement.

If the network is too sparse, the loop is broken—not enough users will see a photo to reply with likes, and not enough buyers will see a listing to purchase a product. If a loop is broken, then the user churns, which further cascades the network problem.

It’s incredibly useful to lay out an engagement loop, one screen at a time, and brainstorm ways to increase each step—this method is at the heart of what I typically do when advising startups on creating higher stickiness.

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.

If you’re inactive, what kinds of notifications are you getting from other users, and are they compelling enough to bring you back? Almost always, churned users don’t receive any communication at all. You can boost reactivation success rate significantly just by sending a weekly digest of the activity in a user’s network, or “Your friend X just joined” notifications.

While it is tempting to throw money at the problem, without a scalable, repeatable source of new users, it is likely that budgets will get out of control and advertising channels will eventually tap out. Viral growth builds on the power of networks to acquire users, often free of charge.

Pay attention to the ratios between each set of users—1000 to 500 to 250. This ratio is often called the viral factor, and in this case can be calculated at 0.5, because each cohort of users generates 0.5 of the next cohort.

In this example, things are looking good—starting with 1,000 users with a viral factor of 0.5 leads to a total of 2,000 users by the end of the amplification—meaning an amplification rate of 2x. A higher ratio is better, since it means each cohort is more efficiently bringing on the next batch of users.

The real magic starts to happen as the viral factor starts to approach 1. After all, at a viral factor of 0.95, 1,000 users show up and then bring 950 of their friends, who will then bring 900, and so on—ultimately the amplification will be 20x. This is the mathematical expression of when a product “goes viral” and starts growing incredibly fast.

Once a metric like this has been defined, it becomes much easier to understand what changes in the product drive it higher. Retention is usually the strongest lever. In the example of PayPal, if a user continues to send money over weeks and months and years, each transaction might help bring in new users onto their platform. In other words, their viral factor grows over time, slowly inching toward a magical > 1.0 metric. On the other hand, if users are always one-and-dones, then they have to invite a ton of users in one big spammy blast in order for the product to propagate—which isn’t ideal.

The cornerstone to amplifying the Acquisition network effect is to understand how one group of users taps into their respective networks to bring in the next group of users.

The final force I’ll discuss is the Economic network effect, which is how a business model—including profitability and unit economics—improves over time as a network grows. This is sometimes driven by what are called data network effects—the ability to better understand the value and costs of a customer as a network gets larger.

But what happens when a product saturates the entirety of its market and can’t grow into the next one? If new products and segments aren’t layered on, growth inevitably slows. At the same time, the marketing channels that companies rely on to grow are degrading, which I deem “The Law of Shitty Clickthroughs.” This law describes how marketing channels inevitably become less effective over time—banner ads and email marketing are two good examples. If your product’s network effects depend on these channels—for example, people sending each other invites over email—growth will inevitably decline over time.

And finally, I describe how discovery of relevant people and content becomes hard—I call this the “Overcrowding” dynamic within networks. More users and more content mean that you need to bring in features like search, algorithmic feeds, curation tools, and a plethora of other tools to manage this. If you don’t solve this problem, then users will start to leave, potentially preferring competitive products that are smaller but more curated.

Neeraj Agarwal, a venture capitalist and investor in B2B companies, first calculated this growth rate by arguing that SaaS companies in particular need to follow a precise path to reach these numbers:64 Establish great product-market fit Get to $2 million in ARR (annual recurring revenue) Triple to $6 million in ARR Triple to $18 million Double to $36 million Double to $72 million Double to $144 million

SaaS companies like Marketo, Netsuite, Workday, Salesforce, Zendesk, and others have all roughly followed this curve. And the rough timing makes sense. The first phase, in which the team initially gets to product/market fit, takes 1–3 years. Add on the time to reach the rest of the growth milestones, and the entire process might take 6–9 years.

Of course, after year 10, the company might still be growing quickly, though it’s more common for it to be growing 50 perce...

In year 1 and 2, you might argue there’s probably zero revenue since the team will be focused on product development. Then year 3 becomes about solving the Cold Start Problem, and only in year 4 does it get to meaningful revenue. The company might hit $1 million/year in revenue that year. Then to extrapolate out, the marketplace product needs to grow from $1 million to $200 million over year 4 to 10—in other words, 266x over a year period. This is where the Rocketship Growth Rate starts to look formidable. It turns out that doubling a product’s revenue each year isn’t enough. Doubling over six years is 64x, and it wouldn’t hit the goal in time.

This is why hitting the ceiling is so dangerous. The consequences for a product that can’t reignite its growth are severe. In the ultra-efficient job market for star product designers, software engineers, and technology employees, everyone knows which products are on the rise and which ones are stuck. Defections to higher-growth, buzzier startups are common.

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. This phase of eBay’s story is not unique for fast-growing startups. What looks like an exponential growth curve is often, in reality, a series of lines layered quickly on top of each other. Uber’s impressive growth trajectory was a combination of launching in more and more cities each year, while simultaneously layering on new products—like carpooling and food delivery. While each line inevitably tapers as each market saturates, adding them in layers can counteract the slowing growth.

Our success was anchored on what I now call The Adjacent User Theory. The Adjacent Users are aware of a product and possibly tried using it, but are not able to successfully become an engaged user. This is typically because the current product positioning or experience has too many barriers to adoption for them. While Instagram had product-market fit for 400+ million people, we discovered new groups of billions of users who didn’t quite understand Instagram and how it fit into their lives.67

There is sometimes low-hanging fruit. Often, it’s as easy as looking at all the users who are signing up or who are active, and checking out their email domains to figure out which companies to sell into. Or maybe just asking these users for their company name and team size—and then start emailing them to sell. Another quick fix is to add a “Contact us” tier of service on the pricing page. At the same time, run a parallel effort that emphasizes content marketing, events, and other programs that drive more top of funnel leads. Build a growth team that can score individual accounts and add life cycle triggers so that the sales team sees when the product reaches a certain footprint inside an organization. In doing all of this, multiple channels can be combined into a broader strategy.

The Law of Shitty Clickthroughs is best countered through improving network effects, not by spending more on marketing.

The penalty for acquiring new members of the hard side but not setting them up to be successful is steep: they churn. Unlike the innovators and early adopters, they might not stick around just because they think the product is cool or fun. They’re doing it to solve a problem—often to earn a living—and if a networked product can’t deliver, they’ll leave.

The entire life cycle of Usenet’s rise and fall serves as a cautionary tale for when networked products hit scale—they suffer from the combined anti-network effects of spam, trolling and other bad behaviors, and most important, context collapse. These provide a strong natural counterbalance to the viral growth and engagement loops that make the network stronger, ultimately canceling these positive forces out. Given enough time, and left untreated, they can collapse the network entirely.

So how do you prevent context collapse? Products like iMessage or WhatsApp give us a clue. Messaging apps are resistant to context collapse. You talk to your dozen or so friends and family, and even if the network adds millions more people, it doesn’t change your experience. Slack channels offer a different model—as more and more people within a company join, people set up smaller spaces to interact with their close teammates. This allows people to split the company-wide network into team-wide networks, or even project-based networks.

Leveraging the network itself to combat abuse is one of the most scalable methods of fighting bad actors. As the network gets larger over time, there are more users who can moderate, and will do so without cost. By giving users the ability to report spam, flag malicious accounts, block bad content, and so on, it not only creates ways for users to customize their own experience but also provides the data that can be used to moderate in other ways. One of the simplest ways to do this is simply to allow users to upvote, downvote, and flag content.

Too many videos on YouTube is a specific case of a broader phenomenon of overcrowding, which can hurt network effects and ultimately make a product unusable. It’s what happens when there are too many comments, threads, and emails in your work inbox. It’s when you’ve followed too many people on a social media app, and there’s too much content to deal with.

It’s not that the existence of old money or old social capital dooms a social network to inevitable stagnation, but a social network should continue to prioritize distribution for the best content, whatever the definition of quality, regardless of the vintage of user producing it. Otherwise a form of social capital inequality sets in, and in the virtual world, where exit costs are much lower than in the real world, new users can easily leave for a new network where their work is more properly rewarded and where status mobility is higher.75

One notable example of this is the ever present “People You May Know” or “Friend suggestions” feature. Every social platform at scale has some kind of implementation of it for a reason: it works incredibly well.

Rarely in network-effects-driven categories does a product win based on features—instead, it’s a combination of harnessing network effects and building a product experience that reinforces those advantages.

In a network representing thousands of such diverse communities, there are always a few that get underserved. This is especially true when larger networks hit a ceiling because they aren’t able to keep the network discoverable, or maintain quality, or because of the other negative effects I explored in the past chapter. The parts of the network that are most affected by these negative factors are the ones most vulnerable to new, emerging competition.

The opportunity to unbundle these larger networks requires both building the necessary product features to support these splinter communities and also taking the direct action to message, advertise, or otherwise convince members of the larger horizontal community to shift over.

The Big Bang Launch is convenient for larger, more established companies as a method to launch new products because they often have distribution channels, huge engineering teams, and sales and marketing support. But counterintuitively, for networked products, this is often a trap. It’s exactly the wrong way to build a network, because a wide launch creates many, many weak networks that aren’t stable on their own. When companies don’t understand these nuances, it leads to disaster.

But here’s the paradox: To build a massive successful network effect, I argue that you must start with a smaller, atomic network. And use the success in the first set of networks to tip over the next set of small networks. I’m not convinced this step can be avoided.

When a networked product takes competition seriously, it has to collect metrics to figure out the comparative position of all the players in the market. This in turn allows product teams to experiment and execute, while keeping an eye on results—it allows them to set goals, not only against their product’s success, but also their competitors’ declines.

Uber’s NACS team invested substantially in understanding and tracking market share in every city. If they saw that they were behind rivals in a market, the team would quickly react. Not in a month, not next week, but rather, the goal was to flip the dynamics in the market as quickly as possible.

Another source came from email analytics companies that had access to the emails—and thus receipts—of millions of consumers, and could offer market share metrics down to specific geographies and trip types.

While these specific methods aren’t applicable to every networked product, there is an important idea at the core: any product that’s in a head-to-head race with competitors should track the outcomes—market share, active users, engagement, or otherwise—while they execute in the market, to put together cause and effect.

By understanding each competitor’s spending on incentives, combined with referencing funding announcements, Uber could estimate what was remaining of a competitor’s financial runway. As the runway got short, if pressure was applied at the right moments—using incentives and product improvements—Uber’s competitors would find it hard to show consistent growth.

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.

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.

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.

If web developers started to target toward the common standards between IE and Netscape Navigator, then it would be much harder for Netscape to build their own developer-driven network effects.