Claude has a privacy problem. Thousands of shared conversations turned up publicly indexed on Google and Bing this week. API keys. Resumes with real names and addresses. What looked like social security numbers. Deeply personal chats nobody thought a stranger would ever read.
Here’s how it happened. Claude’s share feature creates a public link. Anthropic tried to keep those links out of search using robots.txt alone. No noindex tag. And robots.txt does not do what a lot of people think it does.
So Wired covered the story. And in doing so, got the actual fix wrong too.
Robots.txt and noindex are not the same tool
Wired’s advice was to block the pages with robots.txt and add a noindex tag. Sounds sensible. It’s backwards.
If a page is blocked by robots.txt, Google’s crawler never reads the page. Which means it never sees the noindex tag sitting inside it, because the tag lives on the page itself. Block the crawler and you’ve blinded it to the one instruction that would have kept the page out of search.
Google says as much, in bold, in red, at the top of its own documentation. This isn’t obscure. Glenn Gabe said on X that he’d wished the journalists involved had spoken to an SEO first. He wasn’t being unkind. He was right.
This keeps happening, and that’s the real story
Claude’s leak isn’t a one off. It’s the third version of the same mistake in two years.
ChatGPT did this in July 2025. Shared conversations, including one naming a senior consultant at a named firm alongside their age and job description, sat fully indexed because OpenAI made the same robots.txt assumption.
Claude’s own artifacts hit a quieter version of it before that. Pages got indexed, noindex got added afterwards, and Google simply never recrawled the old ones to notice. Search “site:claude.ai/public/artifacts” today and plenty are still there.
Three separate incidents. Same root cause each time. And a major tech publication still got the fix wrong while reporting on the third one. If Wired can’t get this right, don’t assume your web developer has it covered either.
Why indexation control matters more now than it ever did
We already know AI companies will take data from anywhere they can reach it. The story about print books being bought up, scanned, and destroyed to feed training runs made that obvious enough. And honestly, we all know by now that “anywhere they can reach” has never been limited to what’s been freely handed over.
Which is exactly why what does and doesn’t get indexed isn’t just a rankings question any more. It’s about what becomes permanently discoverable, and increasingly, what quietly ends up as someone else’s training data.
I saw a version of this play out at a previous SaaS role. A pricing PDF, sent out to prospects, ended up indexed on Google after a client synced their machine to something like Scribd. The price in that PDF was six times out of date by then. And because the site had no dedicated pricing page of its own, that stray, outdated PDF was outranking the homepage for the company’s own pricing queries. Nobody put it there deliberately. Nobody blocked it either.
The actual lesson
Get an SEO involved before you publish anything you don’t want sitting in a search index forever. Doesn’t matter if you’re Anthropic, Wired, or a business nobody’s heard of.
Indexation control isn’t a checkbox to tick at the end of a project. It’s the difference between controlling what people find about you and letting Google, or the next AI model, decide that for you.
If you’re mapping out what should and shouldn’t be crawlable on your own site, this piece on authority signals in the AI era is a good next stop. And if a redirect or migration is on your roadmap, the redirect audit nobody does until it’s too late covers the technical side of getting that right the first time.
References
- Search Engine Land, Barry Schwartz: “Google indexed Claude chats because Anthropic didn’t block your private chats from search engines”
- Google Search Central documentation, robots.txt and noindex interaction
- Search Engine Land: “Your ChatGPT conversations may be visible in Google search” (2025)
Rhys Turton 