Fan-out queries are the background searches ChatGPT runs in parallel before composing a response, and according to research compiled by Search Engine Journal (2025), the way those queries are evolving suggests OpenAI is actively filtering for higher-quality sources. For brands that rely on AI-generated answers to reach new audiences, the mechanics of that filtering process now matter as much as traditional search rankings.
What Are Fan-Out Queries and Why Do They Matter for Your Brand?
When a user’s question triggers a web search in ChatGPT, the model doesn’t search once. It deconstructs the prompt into multiple background searches, called fan-out queries, runs them in parallel using RAG (retrieval-augmented generation, a technique that pulls live web content into the AI’s answer-building process), and synthesizes the results into a response. According to Search Engine Journal’s analysis, every word and operator ChatGPT puts into those queries reveals what kinds of sources the model expects to find and trust. Monitoring fan-out queries over time is, the article argues, one of the clearest windows into how OpenAI is tuning its search behavior with each model update.
How Is ChatGPT Using the site: Operator to Filter Sources?
One of the most significant patterns identified in Search Engine Journal’s research is ChatGPT’s growing use of the site: operator in its fan-out queries. The site: operator (a search command that restricts results to a specific domain, as documented in Ahrefs’ guide to advanced search operators) has long been standard in SEO research. According to Search Engine Journal’s analysis, ChatGPT is increasingly using this operator to scope searches directly to domains it judges as trustworthy, sometimes adding the word “official” to a query or pointing searches at specific platforms such as individual Reddit communities.
The article’s working theory: OpenAI is using the site: operator, alongside fan-out query design more broadly, as one method of reducing spammy or low-quality outputs from web retrieval. The author draws a parallel to how Google developed its E-E-A-T framework, but describes this as implemented “ChatGPT style.”
What Is the Difference Between Retrieved and Cited URLs?
Understanding this distinction is essential for anyone tracking AI-search performance. Retrieved and cited describe different stages of ChatGPT’s process, and according to Search Engine Journal’s synthesis of multiple industry datasets (2025), they are currently moving in opposite directions:
- Retrieved: a URL ChatGPT fetched while running its fan-out queries. The number of retrieved URLs per response is growing.
- Cited: a URL that appeared as a visible link in the final answer. The number of unique domains cited per response is falling over the same period.
The practical implication: more pages are being considered, but fewer are being credited in the visible answer. This means citation counts alone can undercount a page’s actual influence on what the model says, while also overstating how broadly any single piece of content benefits a brand across different user queries.
Which ChatGPT Version Is Actually Doing the Searching?
Not all ChatGPT users receive the same search experience. According to Search Engine Journal’s analysis, ChatGPT 5.6 ships in two variants: “Sol,” the standard version, and “Luna,” a cheaper variant that rolled out as the new default model for Free and Go users around the start of August 2025. This distinction matters when reading fan-out research, because studies may be measuring different model tiers without noting it.
According to a breakdown by Olivier de Segonzac published in Search Engine Land, more than 90% of ChatGPT’s weekly users are on the free plan. That makes Luna’s retrieval behavior the experience most ChatGPT users actually receive.
Search Engine Journal also notes that not every question triggers a web search at all. Free and cheaper models are more likely to answer from training data, since web retrieval costs more to generate. When retrieval does trigger, ChatGPT pulls from both external search engines and its own internal index, called Labrador.
What This Means for AI-Search Visibility
Here is the clearest way to frame this: ChatGPT is not running a keyword match. It is choosing which sources to consult, and that choice happens before it writes a single word of the answer.
Think of a site:-scoped search (a search targeting one specific domain by name) as the model saying, “I want to know what this particular source says about this topic.” That is a fundamentally different question from a broad web search. A brand can rank well in standard Google results and still get bypassed entirely if ChatGPT’s fan-out process never targets its domain directly. The two are not the same signal.
The retrieved-versus-cited divergence adds a second layer. A brand’s content may be consulted during retrieval without ever appearing as a citation in the visible answer. That is influence without attribution, which means standard citation-tracking tools may be measuring only part of the picture.
One observation the sources do not make explicit: the pattern of site:-scoped fan-out queries likely advantages brands with consistent topical authority across a domain, not just brands with one or two high-performing articles. When a model scopes a search to a domain by name, it has already decided that domain is a known reference for a topic. Earning that kind of recognition, across both training data and live retrieval, may matter more for AI-search visibility than optimizing individual pages for keyword density.
Hingewise’s assessment: fan-out query behavior has become one of the most informative signals available for understanding how AI models evaluate source credibility at inference time (the moment the model generates a live answer, separate from what it absorbed during training). Watching which domains a model chooses to scope its searches to reveals active trust decisions being made in real time, not just associations from historical training data. That is a meaningfully different and more actionable signal than training data analysis alone can provide.
What to Check Before Your Next AI-Search Audit
- Are you tracking fan-out queries, not just final citations? Tools including Peec AI, Profound, the Resoneo Chrome plugin, and FanoutFox surface the actual background searches ChatGPT runs before answering.
- Is your domain appearing in site:-scoped searches? That is a distinct and stronger signal than showing up in broad web results.
- Which model tier are you testing on? Sol and Luna may behave differently; Luna now reaches more than 90% of ChatGPT’s weekly free-plan users, per Search Engine Land (2025).
- Are your metrics measuring retrieved URLs or cited URLs? Know which one your reporting captures before drawing conclusions about AI-search performance.
- Does your site appear consistently across a topic cluster, or only for isolated keywords? Broad topical consistency across a domain may be more relevant to site:-targeted retrieval than single-page keyword rankings.
- Check Google Search Console for referral or crawl patterns that might indicate ChatGPT is fetching your pages during retrieval, even when no citation appears in the final answer.
The fan-out query layer sits between what a user types and what ChatGPT ultimately says. As OpenAI continues refining this process, the gap between brands that get retrieved and brands that get cited, and between brands that get broadly crawled versus brands whose domains get targeted by name, is likely to widen. The most useful development to watch over the coming months is whether Luna and Sol diverge further in their retrieval behavior, and whether the site: operator pattern expands to additional query types beyond where it is currently appearing.
Lam Nguyen, Hingewise
Sources:
“What We Can Learn From Evolving ChatGPT Fan-Out Queries,” Search Engine Journal, 2025. searchenginejournal.com
“Google Advanced Search Operators: The Complete List,” Ahrefs. ahrefs.com