Loud churn and measured churn are rarely the same number

Claude subscribers spent the days after the watermark announcement posting cancellation screenshots. Anthropic's answer was that its own numbers show no significant increase in cancellations. That puts two churn figures on the table, and only one of them was measured.

Why both readings survive

The people who announce a cancellation and the people who quietly stop paying are mostly different people. A public backlash tells you a segment is angry and articulate. It says very little about revenue, because the loudest cohort is usually small and the silent one never explains itself. Anthropic holds the second number, and even that one is early: subscription cancellations bite at renewal, so a measured “no significant increase” days after launch is a reading taken before the tide could come in.

None of that makes the anger illegitimate. It makes it a poor sole input to a roadmap.

What I'd do about it

When a feature triggers a public backlash, put a cohort number next to the noise before anyone reopens the decision, and pick one that moves inside the decision window. Retention will not; it sits still until the renewal cycle, long after the roadmap call is made. Watch cancellation-flow starts and auto-renew switch-offs in the affected cohort against the base rate, with paid retention as the lagging confirmation. If the early numbers move, you have a product problem. If they hold, you have a communications problem, and those get fixed with a post.

There is a second read here for anyone deploying AI under regulatory pressure. The watermark was never a feature Anthropic could trade away on user preference; it is a compliance artifact, shipped to meet the EU transparency Code it signed. Work of that kind ships regardless of how the customer base feels, which is the argument for scoping your compliance obligations well before they turn into a launch. Governance sets that boundary long before anyone writes the release note.