ChatGPT's hidden routing changes which brands get cited

ChatGPT can return different sources for the same question asked twice, and new research shows why: it silently routes queries through several backend search pipelines, and which one runs changes the pages it retrieves and can cite. For brands, that means a citation can appear or vanish between two identical prompts with no change to your content, and AI-visibility tools that sample a query once are measuring that routing as much as they are measuring you.

What happened

Two independent researchers, working separately, found that ChatGPT’s web search sits on top of at least four backend retrieval pipelines, labelled internally Labrador, Bright, Oxylabs and SERP. SEO researcher Chris Green ran the larger test: 1,000 prompts, each repeated up to ten times, for 9,946 completed runs. Labrador handled 88.1 percent of primary search sources, Bright 9.9 percent, Oxylabs 1.7 percent and SERP 0.3 percent.

The instability is the finding. Green reports that 88.4 percent of prompts kept the same primary source across every repeat, but 11.6 percent switched. When the source switched, the overlap between the URLs ChatGPT pulled fell by about 45 percent, and the overlap between domains fell by about 42 percent. Roughly one prompt in nine returns a materially different set of citable pages depending on which pipeline happens to run.

Independently, researcher Suganthan Mohanadasan inspected raw network traffic from a logged-in ChatGPT account and identified the same four pipeline labels, with different pipelines favouring different source types. He also flagged a gap between what ChatGPT fetches and what it credits: some sources were pulled far more often than they were ever cited.

Why it matters

This reframes AI visibility as partly a measurement problem. If the same prompt can surface or drop your brand purely because ChatGPT swapped retrieval pipelines between runs, then a single check tells you little. A brand can look present on Monday and absent on Tuesday with no content change on either side, and a tool that queries once per keyword will report that noise as a trend.

It also widens the gap between being fetched and being cited. Mohanadasan found sources that ChatGPT retrieved heavily but rarely credited, which means appearing in the retrieval set is necessary but not sufficient. As he put it, a brand can be mentioned without being cited, and cited without being mentioned.

What this means for multi-location brands

For a central team tracking AI visibility across hundreds of locations, the practical response is to measure like a statistician, not a spot-checker. Sample each important query several times and read the distribution, not a single result, so a pipeline switch does not register as a win or a loss. Where telemetry allows, record which retrieval source produced each answer, as Green recommends, so shifts in your numbers can be traced to routing rather than mistaken for ranking.

The content response does not change, and that is the reassuring part. The way to be cited across every pipeline is the same first-party discipline that already governs ranking in AI search results: consistent business listings across every platform, clear structured data, and location pages that answer questions directly. Track how your brand and locations actually surface in AI answers with a tool built for it, such as Places AI, and treat any single reading as one sample, not the truth.

“Search source should become another dimension recorded during AI visibility studies wherever telemetry makes it available.”

Chris Green, ChatGPT search research

The bottom line

ChatGPT’s citations are less stable than a one-off check suggests, and part of the movement brands see is backend routing, not their own performance. The brands that read AI visibility correctly will be the ones sampling repeatedly and controlling their first-party data, so they show up whichever pipeline runs.