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. The more workers that adopt a product, the more advanced features they might want to upgrade into, particularly when the features are collaborative in nature—like Slack charging for the ability to search messages from all users across the organizations. Similarly, app stores and other marketplaces will grow their average revenue per user as the number of listings increases. If customers have more choices, they often have a better chance of finding exactly what they’re looking to buy. Then, conversion rates increase.

I use Engagement, Acquisition, and Economic network effects as the core taxonomy for the reason that they map to the key outputs that product teams care most about: active users and revenue, and the leading indicators to these metrics.

Every product can be thought of in this way, and it’s the product team’s goal to increase each of these metrics. However, networked products are special in how they can leverage their networks to drive up each of these variables—something that traditional products can’t.

“Nearly 1 in 4 people abandon mobile apps after only one use.” The authors looked at data from 37,000 users to show that a large percentage of users would quit an app after just a single try. Unfortunately, I’ve found similar results.

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install an app, 70 percent of them aren’t active the next day, and by the first three months, 96 percent of users are no longer active.

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.

At Andreessen Horowitz, we have channels like #2030 about cool technology trends that might affect our near future, or #books and #movies-tv to share our favorite reads and Netflix specials. Each of these new channels signifies a new use case—one might be around company announcements, or socializing, or working on projects together. The more people on the Slack network, the more likely these additional use cases will develop.

It might be tempting to force every user to then use the product this way, but unfortunately correlation does not mean causation. You don’t want to study fire departments and fires and conclude that the former cause the latter!

For a social or communication product, the loop often starts with a content creator posting or sending content. The content is then sent to everyone they are connected to, and depending on the size of the network, they get a nice stream of likes and comments back. That’s the payoff that keeps them going.

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.

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.

Some of the most viral products ever created—like WhatsApp—have been able to generate over 1 million installs per day, without paid marketing.

Networked products are unique because they can embed their viral growth into the product experience itself. When a product like Dropbox has a built-in feature like folder sharing, it can spread on its own. PayPal’s badges and core user-to-user payments accomplishes the same. This is the Product/Network Duo at work again, where the product has features to attract people to the network, while the network brings more value to the product. Workplace collaboration products like Slack ask you to invite your colleagues into your chat, and photo-sharing apps like Instagram make it easy to invite and connect to your Facebook friends. They can tap into your phone’s contacts, integrate with your company’s internal employee directory, or tap into the sharing widgets built into your phone. This is software, not just building a buzzy, shareable video.

Just as the Engagement network effect can be thought of as a step-by-step loop, there is an equivalent framework for the Acquisition Effect as well. For example, consider this process: A new user hears about a service, signs up, finds value in it, and shares the product with their friends/colleagues, who also sign up. These new users then repeat the same steps—this is the viral loop. This loop is fundamentally created within the product experience by software engineers writing code, which makes it different from a fun, viral video—because it’s software, it can be measured, tracked, and optimized to make it more effective. This makes the Acquisition Effect such a potent force.

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. The viral factor can also be above 1, in rare cases, but this typically can’t last for long—eventually market saturation and changing user demographics start to drag down the metrics.

Measuring and optimizing viral growth in this way may make it feel like a spreadsheet project, but I assure you it is much more copywriting, user psychology, and product design. The teams working on growth must be aware of what’s worked in the past—there have been viral loops built on birthday alarm clocks, sending sheep emoji, comparing personality test results, building photo collages where you tag your friends, and much more. Some of these ideas are based on user psychology, which doesn’t change, and can be tweaked and iterated upon for any new product.

Acquisition Can’t Exist without Engagement 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. I’ll use a historical example to make this point.

Ever since the ancient times, people have been lending each other money. Just look at the Code of Hammurabi, one of the oldest deciphered writings in the world—carved thousands of years ago in 1754 BC to govern commercial interactions via fines and punishments. Law 88 says: If a merchant has given corn on loan, he may take 100 SILA of corn as interest on 1 GUR; if he has given silver on loan, he may take ⅙ shekel 6 grains as interest on 1 shekel of silver.55

What does a merchant do when a potential customer walks into a store and wants to purchase a ton of goods on credit? A solution was offered by the “The Society of Guardians for the Protection of Trade against Swindlers and Sharpers,” established in 1776. This society pooled data from 550 merchants to collect information on the reputation of customers. This would make it much harder for a bad customer to defraud multiple merchants. Its key principle: “Every member is bound to communicate to the Society without delay, the Name and Description of any Person who may be unfit to trust.” In other words, this was the beginning of credit scores as a means to assess the trustworthiness of a customer for loans—no swindlers or sharpers allowed.

This Society of Guardians was not the only credit bureau—thousands of similar small organizations were formed over the years, collecting individual names and publishing books with various comments and gossip. Modern giants Experian and Equifax grew from these small, local bureaus. Experian started as the Manchester Guardian Society in the early 1800s, eventually acquiring other bureaus to become one of the world’s largest. And Equifax grew from a Tennessee grocery store in the late 1800s, where the owners started compiling their own lists of creditworthy consumers. These bureaus tended to combine into larger bureaus over time because of what’s often described as a “data network effect.” When a bureau works with more merchants, it means more data, which means the risk predictions on loans will be more accurate. This makes it more attractive for additional merchants to join, who contribute even more data, and so on.

Being able to accurately assess lending risk allows the rest of the network to function—consumers can borrow to get the goods they want, merchants can sell their products profitably, and banks can help underwrite the loans. This network is held together by credit bureaus like Equifax and Experian, who centralize consumer data. But improvements in lending risk aren’t ...

As I’ve discussed earlier, launching a new network often requires subsidies for the hard side, which are paid back over time. These might be structured as up-front payments to content creators and influencers, to get them to participate on the platform. For example, when Microsoft introduced a new livestreaming service to compete with Twitch, it guaranteed Ninja, a streamer with millions of followers, a deal worth tens of millions. Elsewhere, in the streaming content world, there is an ongoing multibillion-dollar battle between Netflix, Hulu, Amazon, and others to sign exclusive content for their subscription content services. As they build up a base of subscribers, aggregating niche audiences compelled by teen horrors, international documentaries, and more, they are able to more efficiently support the up-front funding of content. This can become a huge advantage that increases over time as the networks get bigger.

But the Economic network effect states that for networked products, conversion can go up over time as the network grows.

It might seem that premium pricing is a bad thing, but for many networks like marketplace companies, cryptocurrencies, and payment networks, the users of the network actually win as well. If eBay becomes the trusted, primary place to trade collectibles, then higher conversion rates and higher prices will benefit the sellers. They will make more money, and build their own businesses. When startups like Patreon and Substack create the ability for creators to earn a living by creating content on YouTube or via premium email newsletters, all parties benefit.

However, this trio of network effects does not create permanent invincibility in the market. While a large network often enjoys years of uncontested dominance during the scaling phase, eventually it gets harder. So hard, in fact, that growth might slow to zero.

But earlier on, in 2010, the future was not so clear. The predecessor to Twitch—called Justin.tv—had grown to millions of users but had hit a ceiling. The original idea concept for the product was to focus on streaming video of all types, not just gaming. It had grown well, but plateaued, and the team was getting restless. CEO and cofounder Justin Kan described the situation:

Up to that point, Justin.tv had been a general video streaming network. It had a colorful founding story, with its CEO Justin Kan, walking around with a camera mounted on a baseball hat, broadcasting his life on a laptop connected to multiple cellular networks inside a backpack. Justin was the first streaming video creator on the platform, and the viewers consisted of mostly tech insiders who’d watch his life—this was the service’s first atomic network.

It was through watching Justin.tv that I got to know Justin, along with his cofounders Emmett Shear and Kevin Lin.

We did a lot different with Twitch than Justin.tv. The biggest thing was to focus on streamers, whereas originally it was more about the audience. This meant we worked on tools for streamers, which we improved over time. Making money was important to streamers, even small amounts, so we added tipping features. This was a big deal, because Justin.tv gave streamers some social status by having lots of viewers, but it was a big deal to even be able to make an extra $50/month. We also redesigned the whole website to allow streamers to be discovered based on which game they were playing, sorted in a way that rewarded the most popular streamers. By focusing on games, we could make all of these changes in a way that served our streamers and audience much better.58

The skills needed to be entertaining in real time are different than the skills needed to edit and upload videos.

The new features and functionality were built on the observation that the atomic network for Twitch could be as few as just one streamer and one viewer watching them.

Playing video games with even one Twitch viewer is way more fun than playing by yourself. If they’re watching and chatting while you play, there’s a human connection that you want to come back for.59

The real magic starts to happen once [you] have enough followers on Twitch, and you consistently have viewers every time you start streaming. Then every session on Twitch becomes fun, since there’s always an audience. But it’s even more fun to make money. Once there’s enough viewers, then you’ll eventually make your first dollar. This is a real aha moment—our streamers talk about how making even $20 or $50 a month is an eye-opener. But then build enough of an audience, and eventually, it’s possible to “go pro” and just work by streaming full-time.

This life cycle of a streamer meant that it wasn’t long after Twitch’s launch that the top streamers began to make $300,000+ per year.

Within a month after launch, Twitch had 8 million unique viewers, and within a year after that, it had doubled to 20 million. And then doubled again, and so on, so that today it is one of the most highly trafficked websites in the world. Individual streamers can have over 5 million followers, and make millions in revenue per year. The early code name “Xarth” also lives on, as the name of the main boardroom at Twitch offices today.

In the United States, there are roughly 6 million new businesses started each year, of which only a small number are a strong fit for venture capital investments—estimated to be in the tens of thousands. These startups are then referred to about 1,000 active venture capital firms, and each firm might evaluate a few thousand investment opportunities per year. Of those, just a dozen or two are selected by a firm for investment after multiple meetings, pitches, and hours of time together. Across the entire industry, there are about 5,000 venture capital investments per year available to new and early-stage startups.

These outsized returns are why Stanford researchers have ascribed 57 percent of the value of the US stock market to companies initially backed by venture capital.63

This is the Rocketship Growth Rate—the precise pace at which a startup must grow to break out. How do you calculate this rate of growth? First, by setting a goal of exceeding a billion dollars of valuation—thus being in a position to achieve an IPO—and working backward. Hitting a $1 billion valuation generally requires at least $100 million in top-line recurring revenue annually, based on the rough market multiple of 10x revenue. You’d want to hit that in 7–10 years, to sustain the engagement of the key employees and also reward investors who often work in decade-long time cycles. These two goals—revenue and time—work together to create an overall constraint.

The good news is that networked products often have more tools to counteract this plateau than products that lack network effects. For example, take a new consumer goods brand that sells clothing online—they will see declines in their marketing efficacy as it scales. However, this product category lacks network effects. As their social media advertising costs go up, the team will try to optimize creatives, media buying strategies, and product features—but it won’t be enough. It’s a tough problem to try to double revenues while trying to keep marketing costs the same, all without network effects. The outcome is an early plateau on growth, which is part of why consumer goods companies are often good businesses but not necessarily great ones that can be worth tens of billions.

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.

This provides another way to think of Facebook’s famous “7 friends in 10 days” heuristic. Getting to 7 connections is great, but what about 14—is that better? Definitely. But is it 2x better? Probably not. And if you take it to its logical extreme, every person having 10,000 friends won’t lead to 1,000x the engagement—in fact, it might start to lead to less engagement, as overcrowding effects take over.

When I started at Instagram, the Adjacent User was women 35–45 years old in the US who had a Facebook account but didn’t see the value of Instagram. By the time I left Instagram, the Adjacent User was women in Jakarta, on an older 3G Android phone with a prepaid mobile plan. There were probably 8 different types of Adjacent Users that we solved for in-between those two points.

To solve for the needs of the Adjacent User, the Instagram team had to be nimble, focusing first on pulling the audience of US women from the Facebook network. This required the team to build algorithmic recommendations that utilized Facebook profiles and connections, so that Instagram could surface friends and family on the platform—not just influencers. Later on, targeting users in Jakarta and in other developing countries might involve completely different approaches—refining apps for low-end Android phones with low data connections. As the Adjacent User changes, the strategy has to change as well.

The cheat code for large companies is simply to buy startups that have hit Escape Velocity, and integrate them into a preexisting network. This is exactly what eBay did with PayPal, which must be one of the best acquisitions in the technology industry. It turned out to be a great idea, as PayPal would eventually be worth more than its parent. But these days, acquisitions are hard and expensive. It’s easier said than done in a world where bad acquisitions are rampant, government antitrust concerns plague large networked products, and startups have become super expensive. Yet hitting the ceiling is an inevitable outcome of success, and companies need a response, lest they slow down to a crawl.

Four of our then-clients placed ad banners as part of that first campaign, MCI, Volvo, Club Med and 1-800-Collect. (The other two advertisers were AT&T and Zima.) Keep in mind, this was 1994; the first graphical web browser, Mosaic, was less than a year old (soon to be replaced by Netscape Explorer). And Web access? Purely dial-up, 24.4kps if you were lucky, meaning these ads took a while to load. The online U.S. population? Two million, if that.68 The advertisers launched the first campaign that included a banner ad that asked the viewer, “Have you ever clicked your mouse right HERE? You will.” And with this one tiny banner, Hotwired kicked off an industry that now powers the businesses of Facebook, Google, and the world’s technology giants.

For example, in Twitch’s journey, the team deeply focused on creators, giving them better tools and monetization, which in turn caused them to become more active. More satisfied creators meant they would broadcast live video streams more often, bringing in more viewers, which drove further engagement and monetization. It might have been easy to think they should just double down on their marketing spend, but instead the team looked toward amplifying the network effects that engaged its streamers.

In Uber’s case, the power drivers represented the top 15 percent of drivers but constituted over 40 percent of our trips. They were also among the safest and most highly rated drivers—after all, it was their primary source of income.

Slack’s S-1 filing showed less than 1 percent of Slack’s total customers accounted for 40 percent of the revenue, and Zoom’s indicated that 30 percent of revenue came from just 344 accounts, again less than 1 percent of their customer base.

For the most part, this concentration is the result of healthy loops that drive a network toward higher quality. A good content creator gets likes, shares, and follows, and features like algorithmic feeds will distribute their content even more widely. A bad creator gets none of that boost, and will lower their engagement level into just being a passive viewer, or just churn. A good team organizer will create projects, post new content, and invite coworkers into workspaces that stay active. A bad one will create projects that don’t pick up engagement, and they will eventually churn or their peers will pick up where they left off. A good restaurant on a food-delivery platform will get five-star reviews and eventually generate enough money to help them invest in a specialized kitchen for delivery, and expand into more markets. A bad restaurant will get bad reviews and eventually shut or churn. All of these feedback loops create more concentration within a small number of players, but to the benefit of the entire network.