ChatGPT search does not fetch the web itself. It subcontracts, and which subcontractor picks up a query changes which URLs get cited. Two independent studies, reported by Search Engine Land on 2026-07-08, put numbers on that. Chris Green ran 1,000 prompts up to ten times each, for 9,946 completed search runs. Suganthan Mohanadasan logged the raw network traffic of a Pro account over two days, about 1,240 source records. The same prompt, asked twice, can come back routed through a different provider and carrying a different citation list. It matters this week because that is the run-to-run variance sitting under every AI-visibility number reported this month.
Credibility: MEDIUM. Two independent researchers, published methodologies anyone can open, corroborating each other. Neither is peer-reviewed, the pipeline names are inferred from network traffic, and OpenAI has not confirmed the architecture. Bias named: both authors work in SEO, a mild lean toward "visibility is measurable."
- In Green's 9,946-run dataset, a pipeline labeled Labrador supplied 88.1% of primary search sources; Bright 9.9%, Oxylabs 1.7%, SERP 0.3%.
- 11.6% of prompts changed their primary search source across repeated identical runs.
- When the source switched, URL overlap between runs dropped from 0.273 to 0.149 (roughly 45% lower); domain overlap fell from 0.265 to 0.155 (42% lower).
- Mohanadasan's two-day traffic capture (~1,240 source records across a few dozen searches) independently confirms multiple retrieval providers behind ChatGPT search.
Stay ahead of AI search
The changes that move your rankings and AI citations — and the exact move to make — before they cost you. Free, three times a week.
For anyone who signs off on an AI-visibility report, this dataset puts a number on the randomness: check a prompt once, and roughly one reading in nine is an artifact of which pipeline happened to answer.
A tracker still tells you direction, and direction is what a tracker is for. The sampling method is the part that has to change. A single run is a reading taken while the machine was rolling dice over which retrieval provider to use, and nobody, including the person paying for ChatGPT Pro, is told which one answered. So when a client's citation appears one week and vanishes the next, the first honest question is whether anything about the client changed at all. Only repeated sampling separates routing variance from a real, content-driven move. Anchor the direction against scoreboards that do not route through hidden pipelines: Search Console, and your own first-party citation tests.
- Re-run every tracked prompt at least 5 times per engine before you record a citation win or loss in the prompt-tracking sheet, and report the rate (3 of 5 runs), never a single-run yes or no.
- Annotate the monthly AEO report with the run count behind each data point, and treat any week-over-week swing smaller than the roughly 12% source-switch rate as unconfirmed until a second sampling window repeats it.
- Record why you chose that run count and that threshold, so the calibration survives the next change of engine or tracker.
Search Engine Land: ChatGPT citations change when hidden search pipelines switch, Search Engine Land, 2026-07-08 — https://searchengineland.com/chatgpt-citations-change-hidden-search-pipelines-481843
Chris Green: What can we learn from ChatGPT search?, chrisgreenseo.substack.com — https://chrisgreenseo.substack.com/p/what-can-we-learn-from-chatgpt-search
Suganthan Mohanadasan: How ChatGPT picks sources, suganthan.com — https://suganthan.com/blog/how-chatgpt-picks-sources/