If your brand tracks whether ChatGPT mentions your website, the answer may not mean what you think. Two recent analyses found that ChatGPT routes searches through several hidden retrieval pipelines, and switching between them changes which pages get fetched and cited.
The user-facing citation looks stable enough. Behind it, ChatGPT may be pulling from different data sources, using different search partners, or skipping search entirely. That instability matters for any business building visibility in AI search.
What the research found
Danny Goodwin at Search Engine Land summarized findings from Chris Green and Suganthan Mohanadasan, two researchers who studied ChatGPT’s web search behavior from different angles.
Green ran 1,000 prompts up to ten times each and captured 9,946 completed search runs. He mapped internal labels that describe where ChatGPT pulled its sources. Mohanadasan inspected two days of raw network traffic from a logged-in ChatGPT Pro account and logged about 1,240 source records.
Both found four pipeline labels hidden from users:
| Pipeline | What it appears to represent | Share in Green’s dataset |
|---|---|---|
| Labrador | Established publishers and reference sites | 88.1% |
| Bright | Bright Data-powered retrieval | 9.9% |
| Oxylabs | Oxylabs-powered retrieval | 1.7% |
| SERP | Open-web baseline, common in news-style results | 0.3% |
Labrador dominated Green’s repeated-prompt dataset. Mohanadasan saw Bright play a larger role, especially for commercial, shopping, finance, weather, and local queries. The two samples differ because they used different methods and query mixes, but both show the same underlying pattern: ChatGPT does not use one uniform web search.
Why pipeline switching changes citations
Most prompts stayed on the same primary source across repeat runs, but 11.6% did not. When the pipeline changed, the overlap between cited sources dropped sharply.
| Overlap metric | Same source | Source changed | Difference |
|---|---|---|---|
| URL overlap | 0.273 | 0.149 | ~45% lower |
| Domain overlap | 0.265 | 0.155 | ~42% lower |
That means a page cited in one session may not even be retrieved in another session with the same prompt. The citation card the user sees gives no hint that the back-end source changed. A single successful ChatGPT mention is not proof of durable visibility.
This is a real problem for measurement. Brands and agencies are investing in AI search visibility reports, but the signal they are tracking can shift for reasons outside their control. One day your page is the source. The next day, a different pipeline favors a different domain.
Sometimes ChatGPT skips search entirely
Mohanadasan also found that ChatGPT classifies queries before searching using a turn_use_case field. Some prompts that sounded current were filed as text queries and skipped web search completely. When that happens, no page is fetched, cited, or used as evidence.
Complex queries behave differently. ChatGPT can fan out into rewritten searches, site: probes, pricing checks, and searches for unnamed competitors. The user types one question; ChatGPT searches several variations. That changes which pages can enter the answer process at all.
If your content only matches the exact keyword the user typed, you may miss the rewritten or follow-up searches that actually build the answer.
Fetched is not the same as cited or mentioned
Mohanadasan separated three outcomes that brands often conflate:
- Fetched. The page enters ChatGPT’s context but may never appear to the user.
- Cited. The page is shown as the source behind a specific claim.
- Mentioned. The brand name appears without being tied to a source.
In his small commercial-query sample, Reddit and YouTube were both fetched often. Reddit was cited. YouTube usually was not. The difference was text availability: Reddit exposes text; YouTube search results tend to supply metadata rather than full 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.
That distinction matters for content strategy. A brand’s own site can win on specific facts. A review site, news outlet, or industry publication may be the one that wins on recommendation and comparison queries.
What brands should do about it
There is no single ChatGPT visibility score to chase. The goal is to show up consistently across multiple retrieval paths.
Publish clear, readable pages. Plain HTML, crawlable facts, explicit pricing and specs, and text-heavy content are easier for retrieval systems to use. If your pricing sits behind JavaScript or a login wall, ChatGPT may fall back to a third-party source that scraped or guessed your numbers.
Build strong third-party coverage. Because ChatGPT sometimes relies on third-party pages for recommendations, a brand needs accurate, independent mentions across reference sites, publishers, directories, and review platforms. Your own site can state facts. Other sites can vouch for you.
Diversify what you measure. Track mentions, citations, and source diversity, not just one query result. Test the same prompt multiple times. Look at whether ChatGPT cites the same domain or different ones. Look at whether the answer is built from search at all.
Plan for variability. A citation today does not guarantee a citation tomorrow. Build a content and authority base that can survive pipeline switches instead of optimizing for one visible answer.
The bigger picture
AI search is not a single algorithm. It is a stack of classification, routing, retrieval, and generation layers. Each layer can change the answer without warning. ChatGPT’s hidden pipelines are one more reminder that the citation the user sees is only the surface.
For businesses that depend on being found, the strategy is the same as it has always been: be accurate, be readable, and be present in enough places that no single system can overlook you.
