Lighthouse gave my site a green score. That was the problem.
Why a green score doesn't mean your AI-built site is fine, and what I built about it.
AI made shipping a website absurdly fast. You describe what you want, Cursor or Claude Code or Lovable builds it, and by Sunday night it's live.
I've been writing software for 14+ years, including a few years as a CTO, and I love this shift. But I kept seeing the same pattern when I looked at sites built this way.
The pattern
Someone ships a site with an AI agent. Before launch they do the responsible thing: they open Chrome DevTools, run Lighthouse, and see mostly green. Performance fine, accessibility 90-something, SEO fine. They ship.
And the site still has problems nobody checked:
Google Analytics or a Meta pixel firing before the visitor ever sees the cookie banner (or there's no banner at all).
Missing security headers, so the site can be framed by anyone or has no HSTS.
Accessibility issues Lighthouse doesn't catch. Automated tools only find part of WCAG problems, and Lighthouse runs a subset of those checks. A green score is not "accessible".
Dark patterns the AI copied from the internet: pre-checked newsletter boxes, "No thanks, I hate saving money" buttons, cancel flows buried three pages deep.
None of this shows up in Lighthouse, because Lighthouse was never built to check it. It's a great tool for what it does. It's just not the whole picture.
The AI agent doesn't catch it either. It builds what you ask for. Nobody asks for "make sure trackers wait for consent".
And this isn't rare. WebAIM's 2026 analysis of the top million homepages found 95.9% had detectable WCAG failures. That's the web before the flood of AI-built sites.
What I built
So I built Sitejar. It scans a live page across six areas: accessibility, security, privacy, dark patterns, technical quality and SEO.
The part I care about most: it doesn't just list problems. It gives source-level fixes written so you can paste them straight into your AI coding agent. The agent made the mess, the agent can clean it up.
It also runs as an MCP server, so Claude Code, Cursor or VS Code can call it directly while you build. Each MCP scan checks one page.
What it's not: it's not an overlay widget, and it will never tell you you're "fully compliant". No scanner honestly can. It tells you what it found and how to fix it.
Where I am, honestly
Solo founder, bootstrapped, building from Johannesburg. Since going live, Sitejar has run 60+ guest scans and 100+ scans from registered users. I launched on Product Hunt last week: 10 upvotes, 30 sites scanned, a few free signups. Zero paying customers so far. I'll share the real numbers here as they move.
A side note from that launch: within hours my inbox filled with people offering to sell me upvotes. I ignored every one. If a product only ranks because votes were bought, the ranking means nothing, and I'd rather have 10 real people than 500 fake ones.
What's next
Tomorrow, September 22, I'm launching the Sitejar MCP server on Product Hunt, the version that plugs straight into Claude Code, Cursor and VS Code. If you build with AI agents, I'd love for you to check it out and tell me honestly what you think. There's 80% off your first month for everyone who comes from the launch, the code is on the Product Hunt page: https://www.producthunt.com/products/sitejar?utm_source=indiehackers&utm_medium=post&utm_campaign=ih_intro_post
What I'd love from you
If you've shipped something with AI recently, what did you check before launch? Just Lighthouse, or more?
And if you want, drop your URL in the comments. I'll run a scan and reply with the top 3 issues and how to fix them.
