In early August OpenAI changed how ChatGPT retrieves information by switching to a site: operator that does not start from broad open‑web crawling. As a result, some sources saw abrupt shifts in how frequently they were cited in ChatGPT responses.
Measurements and concrete figures
Third‑party data from PromptWatch shows the site: operator’s share of all ChatGPT queries jumped from 0.37% to 16.8% overnight. Searches per response increased from 1.08 to 1.83. These numbers indicate the system now performs more targeted, site‑focused searches to locate sources.
One visible effect: Reddit, which had been cited in roughly 4.5% of ChatGPT responses, fell to about 0.5% after the change.
Why this matters
OpenAI characterized the shift as a technical optimization. Functionally, the change means the system is building an internal “map” of preferred destinations for particular kinds of questions rather than treating the entire open web as an equal starting point. That internal map determines which sites are consulted first.
The same architecture that deprioritized Reddit also allows structured signals—such as corporate or product data—to be injected into the same retrieval layer. At the VivaTech conference in June, L'Oréal announced a partnership to feed “enhanced signals” into ChatGPT. Given the timing, some observers note that one source lost visibility while another announced paid integration into the system.
Transparency and paid prioritization concerns
Because the retrieval map is internal and not publicly visible, external parties cannot inspect how destinations are ranked or which entities might pay for preferential positioning. OpenAI has not closed the web entirely, but it is selecting where to go for each query based on an internal architecture. The rapid reweighting of sources documented by third‑party measurements raises questions about transparency, competition, and whether paid signals can influence what users see.
Conclusion
Modifying ChatGPT’s retrieval mechanism can be presented as a technical optimization, but it has tangible consequences: some websites’ citation rates dropped sharply while the system became more receptive to structured, potentially paid inputs. The shift increases pressure for clearer disclosure of how source priorities are determined and whether commercial arrangements affect what the model surfaces.
