You Should Never Go Viral With Your AI App

Victor Stepanov’s contrarian take on AI distribution: virality is a trap for AI products. The piece comes from someone who has run growth at Netflix and BuzzFeed, so the advice is “I have done this and it ruined my product” rather than armchair theorizing.

The core argument

For traditional consumer products, network effects are the holy grail — your product gets better as more people use it. For consumer AI, Stepanov argues, the deeper moat is relationship effects:

The memory, personalization, and trust that develops between user and AI — particularly with AI agents — creates switching costs that compound with use.

A viral spike floods the product with users who haven’t built that relationship yet. They churn, leaving zero trace. Worse, they consume capacity and force the company to optimize for the wrong metrics. Slow growth lets relationships compound.

Why it matters here

This piece introduces a category of switching-costs that didn’t exist before LLMs: per-user accumulated context. Every conversation deposits memory, preferences, and trust that competitors literally cannot replicate without time. It’s a moat made of relationship history.

This is a key piece of the chatgpt-pmf story: ChatGPT’s memory feature isn’t a quality-of-life improvement, it’s a deliberate moat. Once a user has a year of conversations with ChatGPT, switching to Claude means starting a new relationship from scratch.

See switching-costs, network-effects (and how relationship effects differ), and chatgpt-pmf.