Skip to main content
Leaked AI chats and shared document URLs do enter AI retrieval pipelines — at microscopic rates. David Konitzny analyzed the 2026 wave of indexed shared chats (Claude, DeepSeek, following ChatGPT in 2024) and traced shared URLs into retrieval (full article, Jul 2026):
  • The mechanism is old web behavior, not an AI flaw: share features generate public URLs, search engines index them — Google Docs and Dropbox links included. DuckDuckGo and Brave often keep them indexed longer than Google.
  • Shared Google Docs URLs are 0.000217% of all URLs in AI retrieval — statistically negligible, but present. Each engine favors a file type: Google AI Overviews surfaces Documents, ChatGPT Presentations, Perplexity Spreadsheets.
  • Reuse differs sharply by engine (Jul 2026 data). For 68.9% of affected prompts the Docs link is a one-off random hit. But Perplexity’s reuse rates are 5–8× higher than other engines — one in eight affected Perplexity prompts reuses the same Docs URL in over half its chats (median 76.9%). Google AI Overview shows the opposite: broadest discovery, lowest stickiness (1.52%).
Takeaway: engines differ not just in what they retrieve but in how consistently they reuse discovered sources. The DeepSeek-specific indexing case: DeepSeek shared links indexed. Source: Beyond AI chat leaks: The hidden risk of shared URLs in the age of AI, David Konitzny, LinkedIn, Jul 2026.