Inside How the Best AI products Grow

Balfour says ChatGPT’s rapid rise came from a strikingly differentiated experience that prompted word of mouth, social sharing, and media coverage. He cautions that this was an exceptional case, not a repeatable playbook. As novelty-driven acquisition slows, AI products need to improve activation and retention; he says generative AI products have retained users worse than non-generative AI products.

Balfour describes input friction as the effort needed to get a useful result, and output friction as the effort needed to apply it. He contrasts ChatGPT’s demanding prompts with GitHub Copilot, where code suggestions fit directly into the user’s workflow. Custom GPTs can lower the effort required to get value by packaging instructions into a ready-to-use experience, which may help engagement and retention.

Balfour says Discord gave Midjourney access to an audience that overlapped with its early adopters, plus public examples and easy user-to-user learning. That community helped people learn a product that was easy to start using but difficult to master, while letting Midjourney focus on its image-generation model. He says the approach involved trade-offs, but Midjourney reached hundreds of millions of dollars in ARR with fewer than 20 employees and appears focused on advanced users.

Balfour prefers Perplexity for many informational searches because it gives a direct answer and lets him inspect its sources, avoiding low-quality SEO content, pop-ups, and gated pages. He notes that some searches still require original assets, such as real product roadmaps, and that search contains many distinct markets. He sees word of mouth and partnerships that place Perplexity in new entry points as possible growth paths, but says its longer-term strategy is unclear.

Balfour says LinkedIn uses AI-generated outlines and introductions as a starting point, then invites users to contribute material that can improve search rankings and refresh articles. Badges provide an incentive to participate, though he says much of the resulting content is poor. He also describes one-click AI assistant prompts placed in existing parts of the feed as low-friction triggers for trying the feature.