ChatGPT search now uses the site:operator at scale
Simon Willison’s Weblog Simon Willison
ChatGPT Search suddenly started using site: a lot more. It jumped from tiny use to 16–17% in one day, which hints OpenAI changed how search works.
Based on reporting by Simon Willison’s Weblog, Simon Willison — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Promptwatch, one of the companies trying to sell “GEO” tools for getting quoted by chatbots, has spotted a sharp shift inside ChatGPT Search. Their tracking shows the site: operator barely showed up for weeks, then leapt on August 8 to 16–17% of the fanout queries they were watching. Before that, the share sat around 0.3% to 0.5%, with a brief dip to 0.15% between August 3 and 5. That dip looks a lot like a staged rollout or a test before launch.
The timing lines up with OpenAI’s August 6 note about GPT‑5.6 Sol in Chat for Plus and Pro users, where the company said it was making answers more reliable on facts and more focused. OpenAI didn’t spell out how Search changed, though. That leaves outside observers doing the sort of digital archaeology now common in this corner of the AI world: watching prompts, comparing outputs, and trying to infer product changes from the crumbs left behind.
Promptwatch’s reports matter because they are based on automated checks across ChatGPT, Claude, and Gemini, and they publish aggregates as part of their own marketing. That doesn’t make them perfect, but it does make them useful. The pattern here is specific enough to suggest a real change in behavior, even if it only covers the prompts they track.
Simon Willison says ChatGPT’s search tool seems shaped more like search(query, recency, domains) than a direct invitation to use site:. That matches the broader story: OpenAI keeps the mechanics hidden, so everyone else is forced to guess from the output. In a follow-up on August 18, Promptwatch also said ChatGPT seemed to be using Reddit far less in those searches. Willison says his own checks have not confirmed any system prompt change that would explain that, and the best leaked prompt collection he knows of still shows nothing relevant.
My take — AI-written commentary, not fact-checked reporting
This is the new normal of closed AI products: the interface changes first, the explanation comes later if anyone can tease it out at all. OpenAI can call that product polish; everyone else has to call it guesswork with receipts. The real losers are the people trying to build around these systems without a window into how they work, which is a very on-brand way for the industry to keep pretending transparency is optional.
Read more about this at: Simon Willison’s Weblog