Listen: Our OpenAI fanout data shows no rise in site: operator use

Recent reports suggested a massive spike in how often ChatGPT uses the "site" operator to limit its web searches to specific domains. According to one study, these targeted searches jumped from less than one percent to nearly seventeen percent on a single day in August. But a closer look at the data suggests a different story. By tracking nearly two hundred thousand search queries generated by GPT-5 models, researchers found no such spike. In fact, their August data showed targeted searches remaining well under one percent. The discrepancy comes down to two things: the version of the model being used, and the style of the prompt. Model updates play a massive role. Older versions of GPT-5 used the "site" operator in about five percent of their searches. However, the newest version, which handled the August traffic, used it less than one-fifth of one percent of the time. Prompt style is even more critical. Short, repetitive prompts used for daily tracking rarely triggered targeted website searches. On the other hand, longer, more detailed prompts designed to mine citations triggered those searches up to ninety times more often. Ultimately, how an AI searches the web depends heavily on how we talk to it. The consumer interface and the developer API behave differently, and the exact wording of a prompt remains the biggest driver of how an AI digs for information.