The quietest users carry the loudest signal

Every product has three audiences: the champions who write posts about you, the critics who write tickets about you, and the silent middle who tried you once and never came back. The first two show up in every review. The third is where the hypothesis is hiding.

The signals from the silent middle are almost never captured by NPS or feature votes. They show up as truncated sessions, unclicked emails, abandoned onboarding at step three of five, a single question asked in a support chat and never followed up. Individually, each looks like a data point of a distracted user. In aggregate, they describe a very specific moment where your product's promise stopped matching the user's model of it.

The practice worth building is a weekly ritual of reading — not counting — the traces of drop-off users. Ten of them. By hand. What were they doing right before they left. What did the last page tell them. What did it not tell them. Machines can cluster the shape of drop-off; they cannot yet feel the small breach of expectation that causes it.

The hypothesis you want is rarely 'add feature X.' It is almost always 'reduce the gap between what we said and what they saw.'