Counting only returning users is a brave choice for day one, and it already paid off. It surfaced three problems that a signup chart would have hidden: guests with nowhere to come back to, a nudge job that ran every hour and sent nothing, and a returning user who got a silent rejection on the photo step.

The second one is worth a check of its own. A job that runs successfully and does nothing looks healthy in every log. A simple daily count of emails sent, with an alert when it is zero for real accounts, would have caught it on the first day.

The design choice of letting the model write only words while tested code builds the page is a good one too. At Ascendra Ventures we use the same split on AI features: the model fills content, deterministic code owns layout and anything a user can break, and generated output is treated as untrusted. It makes failures much easier to find.

Now that sign-up comes first, are you seeing fewer sites made overall, or about the same number with more of them coming back?