Two new analyses found that ChatGPT’s cited sources changed as its search traffic shifted between hidden retrieval pipelines.
The findings. Research from Chris Green and Suganthan Mohanadasan adds a new complication to AI visibility tracking: the final answer doesn’t reveal how ChatGPT selected its sources.
- Both researchers found internal source-selection labels, including Labrador, Bright, Oxylabs, and SERP.
- Those labels sit behind the answer, not in the citation cards users see.
Hidden pipelines changed sources. Green tested 1,000 prompts up to 10 times each and captured 9,946 completed search runs. Most prompts stayed on one retrieval source. Labrador accounted for 88.1% of primary search sources in his dataset, followed by Bright at 9.9%, Oxylabs at 1.7%, and SERP at 0.3%.
- But 11.6% of prompts changed primary search source across repeated runs.
- URL overlap dropped from 0.273 to 0.149 when the search source changed. Domain overlap fell from 0.265 to 0.155. Green calculated that as roughly 45% lower URL overlap and 42% lower domain overlap.
The labels behind results. Mohanadasan inspected two days of raw ChatGPT network traffic from one logged-in Pro account and logged about 1,240 source records across a few dozen searches.
- He found a
result_sourcefield attached to web results, with four observed values: SERP, Labrador, Bright, and Oxylabs. - He described Labrador as including established publishers and reference sites, Bright as tied to Bright Data, Oxylabs as tied to Oxylabs, and SERP as an open-web baseline that appeared mostly in news-style results.
- Green’s repeated-prompt test found Labrador dominated his dataset. Mohanadasan saw Bright play a larger role in his sample, especially for commercial, shopping, finance, weather, and local queries.
Some prompts skipped search. Mohanadasan also found that ChatGPT classified some queries before searching, using a turn_use_case field. Some prompts were filed as text and skipped web search entirely, even when they sounded current. In those cases, no page could be fetched, cited, or used as evidence.
- More complex “thinking” queries behaved differently. Mohanadasan found that ChatGPT could fan out into many searches, including
site:probes, pricing checks, and searches for unnamed competitors. - That behavior changes which pages can enter the answer process because ChatGPT may search rewritten queries, direct site probes, or follow-up checks instead of the exact phrase a user typed.
Fetched was not cited. Mohanadasan separated three outcomes: fetched, cited and mentioned.
- A page can be fetched into ChatGPT’s context without being shown to users. It can be cited as the source behind a specific sentence. Or a brand can be mentioned without being the source of the claim.
- In his small commercial-query sample, Reddit and YouTube were both fetched often, but Reddit was cited and YouTube was not. He attributed the gap to text availability: Reddit threads expose text, while YouTube search results often provide metadata rather than full video transcripts.
- Vendor pages were cited for their own facts, such as prices and specs. Third-party pages were more likely to support broader recommendation claims.
Why we care. There’s no single ChatGPT visibility result to measure. Your page may never be considered if ChatGPT skips search, uses another retrieval source, or finds a clearer third-party page to support the claim.
Readable pages were easier to use. Both analyses showed that ChatGPT’s source selection depended partly on what it could retrieve and read.
- Mohanadasan found cases where ChatGPT appeared to prefer official pricing pages, then fell back to third-party sources when prices were hidden behind JavaScript or otherwise hard to parse.
- Green’s results showed that source routing changed which URLs and domains entered the answer set.
- Plain HTML, crawlable facts, clear pricing and specs, strong third-party coverage, and text-heavy pages became more important when source selection depended on retrieval and readability.