By Gareth Hoyle, Managing Director and Founder at AI and search marketing agency, Marketing Signals
Hundreds of private conversations people had with Claude, Anthropic’s AI chatbot, turned up searchable on Google late last month, including medical records and company documents. In some cases, even children’s names and phone numbers. The chats had been shared using Claude’s built-in ‘share’ feature, which is meant to let one person send a conversation to another.
The reason this keeps happening to AI chatbots and not your online banking is pretty simple. A shared link is just a webpage with a web address, and Google doesn’t know or care whether it came from a chatbot. Post that address anywhere a search engine can see it – a forum, a tweet, wherever – and it ends up in results the same as any other page on the internet would. I see the same pattern on client sites constantly.
A few things worth actually checking if your team uses these tools.
-
Start with a review of what you’ve already shared. Most AI tools have a settings page listing every chat or document you’ve ever made public. Go through it. Delete what you don’t need out there anymore.
-
Don’t leave links switched on longer than they need to be. A link doesn’t stay private just because you haven’t handed it to anyone directly. Once it’s posted somewhere online, even somewhere small, Google can eventually find it. So once whoever needed to see it has seen it, unshare it.
-
Ask yourself this before you share anything: would you put it on your website with no password? Whether it’s client details, contracts, or anything with a name or phone number attached, think before you share.
-
Put it in writing. Most businesses already have a rule about what can go in an email or a shared file. AI chat needs the same treatment – otherwise, your team will default to whatever’s fastest.
This isn’t the first time it’s happened to Claude. Google indexed hundreds of shared conversations last September, and Anthropic said at the time it had blocked crawlers to stop it happening again. That fix was apparently still in place when this second lot showed up anyway. ChatGPT and Grok have both had near-identical incidents.
This doesn’t mean don’t use these tools. They’re genuinely useful, and that’s not going to change. But what it does mean is that the settings and defaults on these platforms shift constantly, so somebody in the business should actually be checking in on this every so often, not setting a policy once and moving on. I go back and check our own team’s sharing settings every few months for exactly this reason.
