ChatGPT Retrieval System Uses Index, Cache, and Live Pages

ChatGPT's system for retrieving and citing web pages operates through a three-tiered retrieval stack, according to an analysis conducted by RESONEO. This system involves a discovery index to find relevant pages, a reading cache to store full copies of fetched pages, and a mechanism for accessing a small set of pages live. Understanding these layers helps explain why certain pages are retrieved, read, and ultimately cited in ChatGPT's responses.

RESONEO's team captured and dissected 1,200 ChatGPT answers, examining 88,000 search results and 26,900 distinct web pages to map this retrieval process. Their investigation initially identified four internal pipelines used for search results: 'labrador', 'bright', 'oxylabs', and 'serp'. These were observed in a 'result_source' field within the raw data stream ChatGPT sends to the browser. However, this field was removed by OpenAI overnight on July 21, prompting RESONEO to develop a classifier that identifies these pipelines by their formatting signatures, such as snippet length and title shape, with approximately 98% accuracy. To further validate their findings, RESONEO utilized a fleet of accounts across different tiers and countries, replayed identical prompts through OpenAI's API, and published canary pages on their own domain with comprehensive server logs to track precisely what OpenAI's robots fetched.

The discovery index serves as the initial layer, responsible for identifying potential web pages that might contain information relevant to a user's query. Once pages are identified, they may be added to a reading cache. This cache stores complete copies of pages that ChatGPT has previously fetched. This caching mechanism likely allows for faster retrieval of information that has been accessed recently or frequently, reducing the need to re-fetch pages from the live web. The third layer involves accessing a small set of pages live, meaning ChatGPT directly fetches content from the web in real-time when needed. Each of these layers possesses its own distinct rules for operation, its own criteria for data staleness, and its own limitations or "blind spots" that can influence the information retrieved and cited.

This detailed breakdown of ChatGPT's retrieval stack provides insights into the AI's information-gathering process. By understanding how pages are discovered, cached, and accessed live, it becomes clearer why some web pages are selected for retrieval, why others are read in detail, and why specific pages ultimately appear as citations in ChatGPT's generated answers. The research highlights the complex interplay between internal indexing, cached data, and real-time web access that underpins the AI's ability to ground its responses in external information. The disappearance of the 'result_source' field suggests OpenAI's ongoing efforts to refine and potentially obscure the specifics of its search provider relationships, emphasizing the proprietary nature of its AI infrastructure.