ChatGPT source shifts: what GEO must do now

The question of how ChatGPT selects and cites sources is becoming increasingly important for SEO and GEO teams. Two recent analyses now show that displayed sources can change when the system switches internally between different retrieval pipelines. For publishers, brands, and in-house teams, this means visibility in AI answers can no longer be evaluated solely through classic rankings or isolated citation screenshots.

What the studies reveal about ChatGPT search

The evaluations by Chris Green and Suganthan Mohanadasan provide a rare look behind the interface. Both observed internal labels such as Labrador, Bright, Oxylabs, and SERP. These labels are not shown in visible citation cards, yet they influence which documents enter the selection process at all. This is the operational core for GEO: not only the final answer text matters, but also the upstream source-routing logic.

Green tested 1,000 prompts with up to ten repetitions and captured 9,946 completed search runs. Labrador clearly dominated his dataset, while Bright, Oxylabs, and SERP held smaller shares. More important than distribution alone is volatility: for a meaningful portion of prompts, the primary source changed across repeated runs. As a result, overlap across URLs and domains dropped significantly.

Why pipeline shifts are critical for visibility

• An identical prompt can trigger different source paths.
• Citation cards do not reflect every internal step.
• Single-answer measurements underestimate variance.
• Domain presence can fluctuate sharply by pipeline.

Mohanadasan analyzed raw ChatGPT network traffic and also documented the same four source labels. His analysis further showed that certain query types more often route through specific data paths. Especially for commercial, local, and transactional queries, the probability shifts for which sources enter the final answer process. For companies, this means content must be planned not only for relevance, but also for retrieval suitability.

When ChatGPT skips web search

Another key finding concerns pre-search classification. Some prompts were treated as a text use case and answered without web retrieval. In those cases, a page cannot be crawled or cited, even if its content is strong. This creates a blind spot in AI visibility assessment: missing mentions do not automatically mean low quality, but may result from an upstream routing decision.

At the same time, more complex “thinking” queries behaved differently. They triggered multiple search branches, including site: probes, price checks, and additional comparison searches. While this broadens the search space, it also increases unpredictability. Pages optimized for the exact user query can still be excluded when ChatGPT prioritizes rewritten or follow-up search paths internally.

Fetched, cited, mentioned: three layers instead of one metric

Especially relevant in practice is the distinction between “fetched,” “cited,” and “mentioned.” A URL can be loaded into context without appearing as a visible source. A brand can be mentioned without providing the evidential basis for a claim. And a source can support narrow factual points while broader recommendations rely on third-party pages. For reporting and KPI design, a single metric is therefore insufficient.

• Fetched measures technical accessibility and retrievability.
• Cited measures visible evidence function in the answer.
• Mentioned measures brand presence without mandatory source status.

In the described examples, text-rich sources such as forums were cited more often than media-heavy results with limited extractable text. This reinforces a known but often underestimated GEO rule: machine readability directly affects whether content is used as reliable evidence. If critical information is hidden behind hard-to-render scripts, visibility can drop at precisely the moment AI models are selecting sources.

Implications for SEO and GEO strategies

The results point to clear operational priorities. Pages with clean HTML, clearly separated fact blocks, consistent product data, and easy-to-parse pricing or specification details have better chances to be usable across multiple retrieval paths. At the same time, strong third-party coverage remains important, because generative systems often rely on external framing for broader recommendations.

• Structured, clearly readable content reduces interpretation errors.
• Facts should be present directly in HTML, not only in interactive widgets.
• Comparison and context pages strengthen semantic positioning.
• Monitoring must account for repeats and pipeline variance.
• GEO analysis should evaluate prompt classes separately.

For editorial and marketing teams, this also requires organizational adjustments. Instead of relying only on classic visibility indices, teams should monitor the path from retrieval to citation in a more granular way. Systematically optimizing content for readability, evidence density, and topical clarity improves not only mention potential, but also the chance of being treated as a reliable source in AI-generated answers.