Shared AI chats landed in search results
Chat pages opened through Claude's share links, along with Artifacts, showed up in Google and Bing results over the weekend. The story was reported on Monday, July 27, and the results disappeared the same day. The technical cause is simple, and the consequence is not: a share link behaves like a page published on the open web.
Share links stayed open to search engines
Over the weekend users noticed that a site operator search would list chat pages hosted on Claude's share address. The results included not only chat transcripts but also the Artifacts that Claude generates. Examples described in the press ranged from work notes and code snippets to internal company documents, employee reviews and health related records. The story appeared in the trade press on Monday, July 27, and the results were reported as removed the same day.
Not a leak, a missing indexing directive
According to Anthropic, conversations are private by default and only the share links a user deliberately creates become publicly accessible. The company said those links are not guessable, that it does not hand search engines a directory or a sitemap, and that links reach search results only when someone posts them somewhere crawlers can see. Anthropic also said it had made changes so that shared conversation pages are no longer indexed.
Google's position runs along the same line: a search engine does not decide which pages are public, and site owners are given controls over crawling and indexing. The technical detail search specialists highlighted points to a common mistake. Blocking a page in robots.txt while also placing a noindex rule on it does not work, because the crawler never reaches the page and therefore never reads the noindex rule, so the page can stay in results. Google's own documentation describes exactly this behaviour. The correct approach is to let the page be crawled so the noindex rule is visible, or to block indexing of share addresses with a response header.
This is not the first case. As the coverage notes, roughly 600 Claude conversations were reported as indexed in 2025, and a similar picture emerged earlier for around 100,000 shared ChatGPT conversations. The issue is not one product's bug but the general design of the share feature.
The share button is really a publishing decision
Most people assume that sharing a conversation means sending it to one person. In practice the share link produces a page with no authentication in front of it. The moment that link is pasted somewhere, search engine crawlers can visit it just like any other browser. The difference is invisible in the interface, but visible in the outcome.
The second point is the nature of the content. Unlike a document repository, AI conversations are held in raw form: a customer name, a contract term, a key in a code snippet, patient information or an unannounced plan all end up in the same window. That means a single share link can carry more context than a classic file share. The third is persistence: a page removed from search results can remain reachable for anyone who already holds the link.
What it means for businesses in Türkiye
In Türkiye, AI tools entered most companies through employees' personal accounts rather than a formal rollout. That creates a concrete risk under the country's data protection law: a conversation containing customer or employee data turning into a share link can amount to a transfer that happens without the data controller's knowledge. So the first step is not technical but written: which tools may be used, which categories of data never enter a chat window, and when a share link may be created at all.
The second step is oversight. Enterprise plans can offer admin controls to restrict sharing and visibility into usage logs; free personal accounts do not. The third is a scan: if your own domain hosts a similar type of page, a quick site operator check in search and, where needed, blocking indexing with a response header takes a few minutes. Finally, a shared link cannot be undone; it can be revoked, but content already viewed does not come back. Prevention has to happen before the incident, not after it.
The UNALSOFT take
This incident exposes the layer most often skipped when AI goes to work: data boundaries. In the Agentic AI systems we build, which data reaches the model, where each output is stored and which link is exposed outward are decided as part of the design. The chat window gets chosen because it is easy, but in enterprise work ease alone is not enough; you need a flow whose access and trail can be controlled.
AI moves fast, data boundaries stay.
Let's define the boundaries and the flow for AI use inside your company.