ChatGPT does not search the web through one pipeline. It quietly assigns each query to one of several backend retrieval sources, and the source it picks changes which pages get cited in the answer. That routing happens behind the citation cards users see, with no signal to the person asking the question. The practical result: the same prompt run twice can draw from two different slices of the web.
Chris Green and Suganthan Mohanadasan authored two separate analyses of the phenomenon, reported by Search Engine Land on July 8. Green tested each of 1,000 prompts as many as 10 times, producing 9,946 completed search runs in total. He found four internal source labels sitting behind ChatGPT’s answers: Labrador, Bright, Oxylabs, and SERP.
Most prompts stayed with one source. In Green’s dataset, Labrador was the primary source for 88.1% of runs, Bright for 9.9%, Oxylabs for 1.7%, and SERP for 0.3%. The dataset leaned heavily toward one default channel.
But 11.6% of prompts switched their primary search source across repeated runs. That switch was not cosmetic. When the source changed, URL overlap fell to 0.149 from 0.273, about 45% lower than when it held steady. Domain overlap similarly dropped to 0.155 from 0.265, roughly 42% lower.
Mohanadasan approached the question from a different angle. He examined raw network traffic captured over two days from a single ChatGPT Pro login. That log surfaced a result_source field carrying the same four values Green found. He described Labrador as leaning toward reference sites and established publishers, distinct from the open web. Bright, tied to the data-scraping vendor Bright Data, played a larger role in his sample for shopping, finance, weather, local, and commercial queries. The two datasets did not fully agree on which source dominated, which is itself evidence that routing depends on query type rather than a fixed default.
Mohanadasan uncovered a second filter working ahead of any retrieval step: a turn_use_case field that sorts prompts before ChatGPT decides whether to search at all. Some queries get routed as pure text generation and never trigger a web lookup, even when they read as timely news. A page that never gets pulled cannot show up as a citation, no matter how well it ranks anywhere else.
He also drew a distinction that most AI-visibility tools currently collapse into one metric. A page can get pulled into ChatGPT’s context, become the credited source for a specific claim, or simply get named as a brand without supplying any fact. Those are three different outcomes, and only the middle one shows up as a citation card. In his commercial-query sample, both Reddit and YouTube appeared frequently as fetched pages, yet Reddit routinely earned the citation and YouTube rarely did. He traced the difference to format, not brand strength: Reddit threads render as plain readable text, while YouTube’s snippets mostly return metadata rather than a transcript ChatGPT can quote.
The findings matter for GEO and AEO measurement, GEO being generative engine optimization and AEO being answer engine optimization. Both disciplines assume a citation audit reflects a stable target. Search Engine Land’s report shows the opposite: one audit captures a single retrieval path among several, not a fixed ranking. A brand can lose close to half its typical citation overlap between two runs of an identical prompt, with no change to its own content.
Search teams tracking AI visibility should treat a single ChatGPT citation check as a snapshot, not a scorecard. Over the next 90 days, rerun the same prompt set on a recurring schedule rather than auditing once. Log whether each result was fetched, cited, or only mentioned. Prioritize plain HTML pages with crawlable pricing and specs, since both researchers found readability, not authority, decided which pages survived the retrieval step.
This article is based on reporting by Danny Goodwin for Search Engine Land, published July 8, 2026, drawing on research from Chris Green and Suganthan Mohanadasan.
