The Cold Start Problem

Why Networks Fail Early

At GREE International in San Francisco, I watched social and mobile games pick up installs and still feel empty. If players could not find a match or an active group, the download count did not matter much. Andrew Chen’s 2021 book, The Cold Start Problem, gave me language for what I had seen: a network needs the right people to show up together before it feels useful.

An app can attract a crowd in its first week and still have no network. I would rather see a small group come back on its own than a large opening spike disappear.

Atomic Networks

Chen calls the smallest group that keeps a network useful an “atomic network.” It might be buyers and sellers in one city, one team using a workplace tool, or enough players in a region to get reliable matches. The test is whether the product works for that group.

Sellers bring inventory; creators give people something to watch; experienced players keep a mode alive. Chen calls this the “hard side” because it is often harder to recruit. At the beginning, those people deserve the recruiting effort before a team pays to bring in everyone else.

Test Assumptions

Hackathons in San Francisco made this practical for me. In two or three days, we could build a rough version and watch someone try it. Where they got confused, or surprised us, told us what to test next.

I am wary of treating any failed test as progress. A useful experiment has a question, a small cost, and a decision attached to its result. If nobody plans to change course, then the team has spent time learning something it will ignore.

Retention Shows the Truth

After launch, I want to know who came back after the initial traffic spike. I look at cohorts—people who joined around the same time—and ask whether each new group finds faster matches, better inventory, more useful content, or easier collaboration than the one before it.

Protect the Core Experience

An early product can be rough around the edges, but the main thing has to work. If checkout fails, a marketplace cannot learn whether people want its goods. If matchmaking fails, a game cannot learn whether people enjoy playing together. I would cut other features before compromising the interaction I am actually trying to test.

Repeat Before You Scale

Once one group works, I would try to repeat it in a second city, team, or game region. Referrals suggest the value can travel. Paid marketing can help it move faster, but it cannot create what the first group lacked.

The Order Matters

I keep coming back to the sequence: make one group feel the product’s value, keep the people who create it, then see who stays after launch attention fades. In games, I have watched a crowded opening turn into an empty room. I want evidence the room stays alive before paying to make it bigger.