The problem shows up after the first click.
A prospective customer fills in a form, sees a success message and waits. Nothing arrives because the form was never connected to a delivery service. The page did exactly what the demo needed. It failed at the job the business depended on.
This is not an argument against building with AI. We use it. The responsibility is to verify the output against real tasks, including the parts that are invisible in a screenshot.
Test a journey, not just a page.
Walk through the site as a first-time visitor on a phone. Follow a service link, submit realistic test data in a safe environment, recover from an error and confirm what reaches the business. A button changing colour is not delivery evidence.
- Verify form delivery and the failure response.
- Visit a missing URL: return a real 404 status and a useful route back.
- Check loading and interactions on a slower mobile connection.
- Use the menu and form with a keyboard.
- Review access permissions, secrets and data handling before launch.
A PageSpeed result needs context.
Use lab measurements to find bottlenecks, then compare them with real-user data when available. The experience includes loading, responsiveness and layout stability; it is wider than how quickly a hero image appears.
Source: Understanding Web Vitals
Keep a person responsible for the outcome.
AI can help draft components, explore alternatives and speed up repetitive work. Someone still needs to understand the architecture, review the changes and decide whether the result is ready. Our definition of finished includes the actual enquiry, the error state and the next person who has to maintain the site.