
How I Reverse-Engineered ChatGPT's Hidden Search Behavior with a Chrome Extension
When you ask ChatGPT a question with Browse mode enabled, something invisible happens. Before giving you an answer, it silently generates 3 to 12 hidden search queries , consults over a dozen web sources, and makes a decision about which ones deserve to be cited. You never see any of this. I built a Chrome extension that intercepts this invisible process in real time. Here's how, and what I discovered. The Hidden Architecture of AI Search ChatGPT doesn't use a simple fetch() to search the web. It uses a complex streaming architecture based on Server-Sent Events (SSE) combined with JSON Patch operations (RFC 6902). Here's what actually happens when you ask ChatGPT "What are the best SEO tools in 2026?": Your browser sends a POST request to the /conversation endpoint The response comes as an EventStream , chunks of data streamed over time Each chunk contains a JSON Patch operation that modifies a running response object Hidden inside these patches are the search queries, source URLs, and
Continue reading on Dev.to Webdev
Opens in a new tab




