Product: subscription site listing public bid opportunities for small contractors. Stage: live, paying customers. Engagement: free QA pass of the public site and signup, from the outside. Status: sent to the founder.
How it came up
The founder posted about his growth on Reddit. I went through the public site and the signup flow the way a prospective customer would, looked at the code the site sends to the browser, and wrote up what I found. I didn't have the server code or sign up for a paid plan.
The real data wasn't loading
On the main web address, the browser blocked the site's own request to its bid data. That's a configuration problem (the data service didn't list that address as allowed), and it's a common one.
What made it serious was what happened next. The site didn't show an error. It quietly fell back to a second source plus a set of bids written directly into the code, and displayed the result as the live feed. A visitor had no way to tell they weren't looking at current opportunities.
When the real data fails, showing sample data as if it were real is worse than showing nothing.
The feed itself had problems
- Garbled entries. Some titles were mangled, some bids were listed twice, and sorting didn't work.
- Impossible dates. Due dates thousands of years in the future.
- Bids filed under the wrong trade.
- A headline number that was written into the page rather than calculated from the data.
Pricing didn't match signup
The pricing page advertised a plan that signup didn't offer, the annual price for one trade didn't match the others, and two plans had confusingly similar names. Choosing a "just browse for now" option left the card requirement and the checkout button in place.
None of these are hard to fix. Together, they're the difference between a visitor trusting the numbers and a visitor leaving.
What the founder said
I sent the top three findings by message, and he asked for the full report.
Time spent: a short pass, no code review. Status: sent to the founder.
What this says about AI-built apps
Fallbacks are a common AI-coding pattern: if the real thing fails, use something that keeps the page from breaking. That's a reasonable instinct for a demo, and a dangerous one in production, because it hides the failure from everyone, including you. Any fallback in a production app should either be obviously labeled or should tell someone it fired.