INSIDE CHATGPT SEARCH: HOW WEB.RUN AND FAN-OUT QUERIES SHAPE AI VISIBILITY

Summary from the TLDR Marketing newsletter

New research reverse-engineering ChatGPT's internal browsing system found that a March model switch reduced cited domains per response by over 20%, concentrating visibility on fewer high-authority sources. The newer model distributes queries across 10 or more "fan-out queries" per response, each targeting specific trusted domains, while product searches now get individual retrieval commands per product rather than bundled calls. The study distinguishes between parametric visibility, built through training data including press coverage and Wikipedia presence, and dynamic visibility from real-time search, noting that a brand unknown to the model's training data won't surface even when search is enabled.