ChatGPT Citations Change When Hidden Search Pipelines Switch

Introduction

Two new analyses found that ChatGPT's cited sources changed when its search traffic moved between hidden retrieval pipelines.

The findings

Research by Chris Green and Suganthan Mohanadasan adds a new layer of complexity to tracking visibility in AI answers: the final response does not reveal how ChatGPT chose its sources.

• Both researchers found internal labels for source selection, including Labrador, Bright, Oxylabs, and SERP.
• These labels sit in the background of the answer, not in the source cards shown to users.

Hidden pipelines changed sources

Green tested 1,000 prompts up to 10 times each and logged 9,946 completed search runs. Most prompts stayed on a single 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 their primary search source between repeat 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 this as roughly a 45% drop in URL overlap and about 42% in domain overlap.

The labels behind the results

Mohanadasan reviewed daily raw network traffic from a single connected ChatGPT Pro account and logged roughly 1,240 source records across several dozen searches.

• He found a result_source field attached to Web results, with four observed values: SERP, Bright, Labrador, and Oxylabs.
• He found that Labrador includes established publishers and reference sites; Bright is tied to Bright Data; Oxylabs is tied to Oxylabs; and SERP is the open-web baseline that appeared mainly in news-style results.
• Green's repeat-prompt test found Labrador dominated his dataset.
• Mohanadasan attributed a larger role to Bright in his sample, especially for commercial, shopping, finance, weather, and local queries.

Some prompts skipped search entirely

Mohanadasan also found that ChatGPT classified some queries before search using a turn_use_case field. Some prompts were classified 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 ChatGPT can split into dozens of searches, including site: checks, price checks, and competitor searches without a brand name.
• That behavior changes which pages can enter the answer process, because ChatGPT may search rewritten queries, direct site checks, or follow-up checks instead of the exact wording the user typed.

Fetched does not mean 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, and a brand can be mentioned without being the source of the claim.
• In a small sample of commercial queries, Reddit and YouTube were fetched often, but Reddit was cited and YouTube was not. He attributed the gap to text availability: Reddit threads contain text, while YouTube search results provide metadata instead of full video transcripts.
• Manufacturer pages were cited in their factual context, such as prices and specs. Third-party pages supported broader recommendations.

Why this matters

You cannot measure a single ChatGPT visibility outcome. Your page may never be checked if ChatGPT skips search, uses a different retrieval source, or finds a clearer third-party page that supports 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 and then fell back to third-party sources when prices were hidden behind JavaScript or hard to parse semantically.
• Green showed results where source routing changed which URLs and domains entered the answer set.
• Simple HTML, crawlable facts, clear prices and specs, strong third-party coverage, and text-rich pages mattered more when source selection depended on retrieval and readability.