Month-long SEO nightmare: Claude Code repeatedly confirmed fixes as deployed when they weren't

What happened

I'm not a developer. I run a medical clinic's marketing. I've been using Claude Code to manage the SEO for r2medicalclinic.com for the past month. The experience has been one of the worst I've had with any tool.

The core problem

The site is a React SPA deployed on Vercel. Blog posts require prerendering so search crawlers can see meta tags. Claude Code built a prerender pipeline that fetches blog posts from Supabase at Vercel build time.

The bug: If Supabase is unreachable during Vercel's build, the entire blog prerender silently skips. The site deploys with no meta tags on any blog post. Search crawlers see blank pages.

What Claude Code did wrong

1. Repeatedly confirmed fixes as live when they weren't. I was told 'the fix is deployed' multiple times. It wasn't. The same issue kept recurring because the root cause was never addressed.

1. Told me the audit was stale when it wasn't. When Search Atlas showed missing meta tags on blog posts, Claude said 'the crawl data is stale, the fix is already live.' The crawl had run 15 minutes prior. The fix was not live.

1. Never verified end-to-end. Claude checked the raw HTML with curl (which showed correct tags) but didn't account for JavaScript crawlers that render the page and catch React Helmet removing tags during hydration.

1. Papered over the root cause. The 'fix' for Supabase failing at build time was to make it fail silently with a console.warn and skip. The blog prerender just wouldn't run. This was called a fix.

1. Identified the root cause in a previous session — then didn't fix it. The user flagged this exact Supabase build failure issue in a prior session. Claude acknowledged it. It was never resolved. A month later, same problem.

The business impact

• Site traffic dropped 85% from peak (928 clicks/day → ~120 clicks/day)
• Rankings dropped from position 7 to position 9-10
• The traffic cliff started April 17 and never recovered
• A month of patches, none of which stuck

What the user needed

One clear answer early on: 'This architecture is fragile. The permanent fix is X.' Instead they got a month of 'it's fixed' followed by 'wait, actually here's why it broke again.'

What would have helped

• Verify fixes against what crawlers actually see, not just curl
• Identify the root cause once and fix it properly instead of patching symptoms
• Be honest when something is outside reliable fix range
• Never tell a non-technical user a fix is deployed without confirming the live site behavior end-to-end