OpenAI<\/a> employs several distinct source pools, each serving a different function in the response generation process. The analysis, conducted by researcher Chris Green, involved running 1,000 prompts from various categories ten times each, totaling close to 10,000 individual query runs. The data reveals that ChatGPT primarily relies on four source pools, each with its own characteristics and use cases.<\/p>\nThe first and most dominant pool, named \u201cLabrador,\u201d represents a curated list of established publishers. This includes well-known news organizations such as Reuters and The Guardian. \u201cBright\u201d operates as a broader retrieval system, pulling from a wider set of URLs and a more diverse range of domains. \u201cOxylabs\u201d is observed infrequently but, when used, shows a surprisingly varied domain selection. \u201cSERP\u201d functions as a direct connection to the open web and appears almost exclusively in response to news-related queries.<\/p>\n
These pools are not interchangeable. Labrador tends to produce smaller, more focused sets of results. Bright generates more URLs, more citations, and a wider spread across different websites. Oxylabs, despite its rare appearance, offers a diverse domain mix. SERP is reserved for situations requiring the most current information from general web sources.<\/p>\n
Labrador Dominates Over 88 Percent of Queries<\/h2>\n
The distribution of these source pools is heavily skewed toward Labrador. According to the data, Labrador served as the primary search source in 88.1 percent of all cases. Bright followed at a distant 9.9 percent, Oxylabs at 1.7 percent, and SERP at just 0.3 percent. In 88.4 percent of all prompts, the primary source pool remained consistent across all ten runs.<\/p>\n
This dominance suggests that Labrador is the default retrieval system for most queries. The other pools are activated only under specific conditions. Bright, for example, is most frequently added alongside Labrador or Oxylabs, indicating that it is used as a supplementary source rather than a replacement. This hierarchical structure means that the vast majority of responses are generated from a curated set of established publisher content, with broader web access reserved for specialized needs.<\/p>\n
Source Switching Occurs in Over 11 Percent of Queries<\/h2>\n
While Labrador\u2019s dominance is clear, the stability of source selection is not absolute. In 11.6 percent of prompts, the primary source pool changed between runs. Furthermore, in 13.4 percent of cases, the combination of source pools used shifted, even if the primary source remained the same. This variability has a direct impact on the URLs and domains referenced in responses.<\/p>\n
When a query stayed within the same source pool across multiple runs, the overlap in URLs called to generate responses was 27.3 percent. When the source pool changed, this overlap dropped to 14.9 percent. A similar pattern emerged at the domain level, with 26.5 percent overlap when the source pool remained constant and 15.5 percent when it changed. This means that a user submitting the same prompt twice might receive answers grounded in substantially different sets of web content.<\/p>\n
The type of prompt influences the likelihood of a source change. Timeless, fact-based questions almost always route through Labrador. Queries involving current events, regulations, surveys, or explicit date references increase the probability that Bright, Oxylabs, or SERP will be used. This mechanism suggests that OpenAI has built a system that attempts to match the retrieval method to the nature of the information request.<\/p>\n
What Is the Difference Between ChatGPT\u2019s Source Pools?<\/h2>\n
The difference lies in the scope and curation of the content each pool accesses. Labrador is a curated list of established, high-authority publishers like Reuters and The Guardian, making it the default for most general knowledge queries. Bright is a broader retrieval system that draws from many more URLs and a wider variety of domains, making it suitable for queries requiring more diverse or less established sources. Oxylabs, used infrequently, provides access to a surprisingly wide range of domains when it is called upon. SERP, which stands for search engine results page, is a direct connection to the open web and is reserved almost exclusively for news-related and highly current queries. This tiered approach allows ChatGPT to balance authority, breadth, and timeliness depending on the specific information need.<\/p>\n
Two Critical Implications for Tracking AI Visibility<\/h2>\n
For anyone monitoring how their content appears in AI-generated responses, the findings point to two necessary adjustments. First, the source pool used by ChatGPT should be tracked as a separate metric. The current practice of measuring visibility at the aggregate level obscures the fact that different retrieval systems access different slices of the web. A piece of content might be highly visible in Labrador but absent from Bright, or vice versa. Understanding which source pool drives visibility for specific queries is essential for accurate measurement.<\/p>\n
Second, the differences between logged-out sessions and paid, logged-in accounts are likely more significant than previously understood. The analysis hints at A\/B testing involving features like shopping functions, though the sample sizes were too small to draw definitive conclusions. The existence of multiple parallel retrieval systems means that user state, account type, and other contextual factors could further influence which source pools are activated.<\/p>\n
The data also reveals that ChatGPT cannot be understood as a single, unified retrieval system. It operates as a collection of parallel retrieval mechanisms, each with its own access to different parts of the web. Source switching, while affecting only a minority of prompts, has a measurable and meaningful impact on the content used to generate responses. This variability is not a bug but a feature of a system designed to balance different types of information needs. For publishers and SEO professionals, adapting to this reality requires a more granular approach to tracking and analysis, one that accounts for the multiple systems working behind a single query.<\/p>\n","protected":false},"excerpt":{"rendered":"
ChatGPT does not generate answers from a single, stable source pool. Instead, it draws from multiple retrieval systems that can shift between queries, even when the same prompt is submitted repeatedly. This variability, documented through a large-scale analysis of nearly 10,000 query runs, introduces a new layer of complexity for anyone tracking visibility within AI-generated […]<\/p>\n","protected":false},"author":7,"featured_media":74433,"comment_status":"closed","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"fifu_image_url":"https:\/\/iili.io\/Cltj93B.jpg","fifu_image_alt":"ChatGPT Source Switching Makes Tracking Harder","footnotes":""},"categories":[31],"tags":[],"class_list":["post-62554","post","type-post","status-publish","format-standard","has-post-thumbnail","category-technology"],"fifu_image_url":"https:\/\/iili.io\/Cltj93B.jpg","fifu_image_alt":"ChatGPT Source Switching Makes Tracking Harder","_links":{"self":[{"href":"https:\/\/overcentral.com\/en\/wp-json\/wp\/v2\/posts\/62554","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/overcentral.com\/en\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/overcentral.com\/en\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/overcentral.com\/en\/wp-json\/wp\/v2\/users\/7"}],"replies":[{"embeddable":true,"href":"https:\/\/overcentral.com\/en\/wp-json\/wp\/v2\/comments?post=62554"}],"version-history":[{"count":0,"href":"https:\/\/overcentral.com\/en\/wp-json\/wp\/v2\/posts\/62554\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/overcentral.com\/en\/wp-json\/wp\/v2\/media\/74433"}],"wp:attachment":[{"href":"https:\/\/overcentral.com\/en\/wp-json\/wp\/v2\/media?parent=62554"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/overcentral.com\/en\/wp-json\/wp\/v2\/categories?post=62554"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/overcentral.com\/en\/wp-json\/wp\/v2\/tags?post=62554"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}