A ChatGPT citation is the visible end of a much larger selection process. Before a source can appear beside an answer, the system may decide whether to search, choose a retrieval pipeline, rewrite or expand the query, fetch candidate pages and select which evidence deserves a citation.
That layered process explains why repeated prompts can produce different source lists without any underlying page changing. It also changes how publishers should interpret AI visibility: one observed answer is a sample of a variable system, not a definitive ranking.
A citation is the output of several hidden decisions
The source cards visible to users do not disclose the full route that produced them. According to the CrushPress.AI report, research by Chris Green and Suganthan Mohanadasan identified internal source-selection labels including Labrador, Bright, Oxylabs and SERP. These labels appeared behind the answer rather than in its public citations.
This creates several distinct opportunities for a page to be excluded. ChatGPT may classify the prompt as not requiring web search. If it does search, the selected retrieval source may not surface the page. The system may then fetch the page but decline to cite it, or it may use the page for a narrow factual claim while relying on another source for the broader answer.
The practical distinction is important. A missing citation does not, by itself, show that a page lacks authority or relevance. It may reflect an earlier routing, retrieval or parsing decision that is invisible in the final response.
Repeated prompts expose pipeline-level variability
Green examined 1,000 prompts, running each as many as 10 times, and recorded 9,946 completed searches, as reported by CrushPress.AI. Labrador was the primary search source in 88.1% of those runs, followed by Bright at 9.9%, Oxylabs at 1.7% and SERP at 0.3%.
Most prompts remained on one primary source, but 11.6% switched sources across repeated runs. For prompts that switched, reported URL overlap declined from 0.273 to 0.149, while domain overlap declined from 0.265 to 0.155. Green characterized those changes as approximately 45% less URL overlap and 42% less domain overlap.
Those overlap figures measure consistency between result sets; they should not be read as a page’s probability of earning a citation. Their significance is structural: a change in retrieval route can materially change the pool of domains and URLs available to support an answer.
Mohanadasan observed a different distribution while examining two days of raw network traffic from one logged-in Pro account. His sample contained about 1,240 source records from a few dozen searches. Although he found the same four result-source values, Bright had a larger role in his sample, particularly for commercial, shopping, finance, weather and local queries. SERP appeared mainly with news-oriented results, while Labrador included established publishers and reference sites; Bright and Oxylabs were associated with their namesake data providers.
The differing distributions are not necessarily contradictory. The studies used different prompts, observation methods, sample sizes and account contexts. Together, as presented in the source article, they suggest that no single observed pipeline mix should be assumed to represent every query class or user session.
Search can be skipped, rewritten or expanded
Pipeline selection matters only after the system decides to search. Mohanadasan reported that ChatGPT first classified some requests through a turn-use-case field. Some apparently current prompts were categorized as text tasks and did not trigger a web search. When that happens, no current page can be fetched or cited, regardless of how well it is optimized.
Queries that received more extensive reasoning could travel in the opposite direction. The reported traces showed branching searches that included site-specific probes, pricing checks and searches for competitors the user had not named. Consequently, a publisher may be competing for retrieval against results generated from several machine-created subqueries, not merely the exact wording entered by the user.
This makes prompt-level visibility difficult to reduce to conventional rank tracking. The same surface question can lead to no search, a relatively direct search or a multi-step investigation. Each path creates a different candidate set before citation selection begins.
Fetched, cited and mentioned are different outcomes
Mohanadasan separated source participation into three useful states: fetched, cited and mentioned. A fetched page enters the system’s working context. A cited page is displayed as support for a claim. A mentioned brand may appear in the prose without its own site serving as visible evidence.
| Outcome | What it indicates | What it does not establish |
|---|---|---|
| Fetched | The page was retrieved for possible use. | That users saw it or that it supported a final claim. |
| Cited | The page was presented as evidence for part of the answer. | That it was the only source consulted or the preferred source in every run. |
| Mentioned | The brand or entity appeared in the response. | That its own website was retrieved or cited. |
The source article illustrates the distinction with a small commercial-query sample. Reddit and YouTube were both fetched frequently, but Reddit received citations while YouTube did not. Mohanadasan attributed the difference to accessible text: Reddit threads exposed usable copy, whereas YouTube search results often supplied metadata rather than full transcripts. Because the sample was limited, this should be treated as an observed pattern rather than a universal rule about either platform.
Source roles also varied by claim type. Vendor pages supported first-party facts such as prices and specifications, while third-party pages were more likely to support comparative recommendations. In some cases, ChatGPT appeared to seek an official pricing page but use a third-party source when the official information was hidden behind JavaScript or otherwise difficult to parse.
The broader implication is that citation eligibility depends on both relevance and usability. A page can contain the right information yet lose the visible citation if the information is inaccessible, ambiguous or less suitable for the particular claim than another source.
A better framework for measuring ChatGPT visibility
Because routing and search behavior can change between runs, citation monitoring should emphasize distributions rather than isolated answers. Repeated tests can show how often a domain appears, whether the cited URL changes, which claim types attract first-party or third-party support, and how volatile the results are. The studies reported here do not establish a universal number of repetitions, so testing depth should be documented instead of presented as a fixed standard.
Measurement should also keep brand inclusion separate from source attribution. Citation share, mention share and fetched-page data answer different questions. Combining them into one visibility score can conceal whether a brand is absent from the answer, present without evidence from its own site, or retrieved but not shown to the user.
Key takeaways
- A citation is produced by a chain of classification, routing, retrieval and evidence-selection decisions.
- Repeated prompts are necessary to reveal variability; a single response cannot represent a stable source position.
- Search eligibility should be evaluated separately from citation performance because some prompts may not trigger web retrieval.
- Fetched pages, visible citations and uncited brand mentions should be tracked as distinct outcomes.
- Plain HTML, clearly labeled facts, accessible prices and specifications, and substantial text improve the chance that retrieved information can support a claim.
- First-party pages and independent coverage serve different evidentiary roles, so visibility work should account for both.
As AI search measurement matures, the most durable approach will be to record uncertainty rather than hide it. Publishers that make evidence easy to retrieve and interpret, while measuring performance across repeated runs and source types, will be better equipped to understand citation changes as the underlying pipelines evolve.
Leave a Reply