It's far more effective to test a wrong answer than it is to test no answer. Instead of researching toward a first draft, I deployed a viable site on day one and spent the rest of the week iterating it against real user data.

The client was heavily inspired by Apple, so my first prompt referenced Apple's layout patterns alongside product renders from our industrial designer. Beyond that, I deliberately gave Lovable little direction besides user points generated by Claude. Starting with a barebones prompt can give the AI a chance to unravel paths forward I hadn't considered. Once I had my first iteration, the question was now how do I test fast?

Recruiting the right test users normally takes weeks. The client was already setting up Meta Ads Manager, so I used it as a live testing pipeline: two domains, identical ads, real traffic, real metrics. I measured success on cost per lead, conversion rate, and click-through rate.

At this pointed we needed actual user data to target the ads properly. I went out and interviewed barbers in my city. Those conversations gave me user data for Ads Manager, and told me exactly which product features to prioritize on the page. For example, barbers loved thinner blade heads, so that feature went higher up in the site hierarchy.

I tested layout variations and, crucially, the level of detail on the page. Because the audience was professional barbers, they responded better to more technical detail, and leaned toward the sportier, edgier design directions. My goal here was balancing information without bogging down the funnel.

I setup Stripe to handle reservation, Supabase handle records, and Tally to handle emails. The funnel was intentionally minimal to reduce the chance of creating pain points we can't quickly test for. I let AI generate the payment and signup flows (coordinating credentials and account setup with the client), and they worked with only minor copy touch-ups.