For the past several months, the search engine optimization industry has methodically taken ChatGPT apart to understand how the artificial intelligence searches the web, retrieves information, and selects its sources. To gain a clearer, independent picture of how information is retrieved and what it signifies for modern AI visibility, researchers recently undertook a massive empirical analysis. By examining 66.6 million free, logged-out GPT-5.6 Luna responses alongside their 15 million fan-out queries collected across August and September, data has emerged shedding light on the hidden operational mechanics of conversational search engines.
Free sessions offer the closest approximation to how everyday users interact with ChatGPT. According to Olivier de Segonzac, co-founder of RESONEO, more than 90 percent of the platform’s weekly users rely on the free tier. Furthermore, logging out strips away individualized chat history and memory functions, allowing analysts to observe ChatGPT in its default behavior rather than seeing responses shaped by a specific user’s past prompts and preferences.
The investigation revealed that only 14.4 percent of the millions of responses analyzed contained fan-out queries. Among those specific responses, the model averaged roughly 1.6 fan-out queries each. Because the margin for discovery is narrow, a publisher’s web page must directly match one of those few underlying searches, or ChatGPT will never encounter it.
When looking closely at how those searches are formulated, official and first-hand sources dominate. Of the 15 million fan-out queries examined, 6.8 percent utilized the site: operator to target a single, specific website. Among the roughly 160 most-searched domains, three out of four searches targeted a direct brand site. For the purposes of the analysis, a brand site was defined as a primary corporate web property like apple.com or ikea.com, distinct from broad retailers, government portals, online forums, or consumer review aggregates. Retailers such as Amazon, Walmart, and Costco accounted for 11.9 percent of these targeted searches, while community and social platforms like Reddit and Facebook captured just 1.9 percent combined.
The textual content inside the fan-out queries further underscores this heavy emphasis on official, primary information. After the year identifier "2026," words like "official" and "price" stood out as the most common terms, growing steadily month-over-month. For instance, "2026" appeared in 17.6 percent of queries in August and rose to 22.9 percent in September, while "official" climbed from 10.1 percent to 13.1 percent over the same timeframe. Similarly, inquiries containing "price" increased from 6.0 percent to 8.9 percent.
This behavior aligns with logical search intent. When a user seeks a specific product’s price or technical specifications, a brand’s official website is the most reliable destination. However, when a query lacks a single, definitive answer, ChatGPT is significantly more likely to look beyond official sources. In a separate test encompassing 20,000 prompts in the United States, 84.4 percent of factual questions regarding pricing, product features, or corporate details cited at least one official source. By contrast, for comparative prompts such as evaluating "A versus B" or determining which option is more affordable, that figure dropped to 68.3 percent.
The distribution of citations shifted accordingly. For straightforward factual questions, 63.4 percent of citations pointed to official websites, whereas comparative inquiries saw that figure drop to 46.4 percent, with the majority of citations flowing to third-party domains. These operational patterns indicate that ChatGPT relies most heavily on primary sources for baseline facts, while depending more on external commentary when it needs to weigh competing options against one another. Consequently, brands must ensure their core pricing, specifications, and support details are explicitly published in plain text on pages the system can easily crawl and index.
Consumer brands make up a dominant share of the domains most frequently targeted by ChatGPT’s site-specific searches. While these rankings stem from a specific dataset and set of prompts, tracking approximately one million site-specific queries offers clear visibility into the domains the model leans on most. Government domains also feature prominently, with international and domestic entities such as Brazil’s gov.br, the United Kingdom’s gov.uk, France’s service-public.fr, the National Highway Traffic Safety Administration, the Food and Drug Administration, and the Securities and Exchange Commission all placing within the top 150. This trend mirrors observations by Lily Ray, founder of Algorythmic, who noted that ChatGPT consistently favors official sources for Your Money or Your Life queries.
At the same time, Reddit remains the ninth most searched domain via the site: operator. This persists despite the platform being cited less frequently overall, matching findings from Suganthan Mohanadasan, co-founder of Snippet Digital, who observed that ChatGPT continues to fetch content from Reddit when seeking authentic user opinions. Interestingly, among the 17 publishers and platforms holding formal AI licensing deals with OpenAI, Reddit is the sole entity appearing in the top 150 most-searched domains via the site: operator. One plausible explanation is that ChatGPT already has content from its official publishing partners pre-loaded into its system architecture, reducing the need for real-time exploratory searches.
While site-specific searches reveal what ChatGPT looks up during query fan-out, they do not necessarily reflect what the model ultimately cites in its final answers. The cited domains tell a notably different story, particularly regarding OpenAI’s media partners. Across a broad tracking sample, partner publications included in the top 200 most-cited domains more than doubled their combined share of citations, rising from 7.5 percent in March to 15.8 percent in July, and maintaining a substantial 12.5 percent share in late September. This upward trajectory highlights the growing value of securing brand coverage within licensed partner publications.
The evolving nature of fan-out queries also highlights shifting preferences in content types. Between August and September, queries focusing on prices, product specifications, and reviews experienced notable growth, while inquiries regarding definitions and meanings saw minor declines. "Price" surged as a share of fan-outs, marking a relative increase of nearly 50 percent in a single month, while queries containing "reviews" grew by a third and "specifications" broke into the top 50 for the first time.
Simultaneously, a transformation has unfolded regarding which external websites ChatGPT chooses to cite. Historical datasets tracking U.S. citations between March and September reveal a dramatic contraction in visibility for major social, reference, and news aggregates. Research highlighted by Snippet Digital and Aleyda Solís, founder of Orainti, showed that Reddit citations within ChatGPT dropped precipitously over a six-month window. In March, roughly one in every five answers cited Reddit, but by late September, that ratio had fallen to approximately one in 43. Wikipedia experienced a similar decline, dropping from one in seven answers to one in 35, while YouTube saw a reduction from one in 370 answers to one in 500.
While these platforms remain among the most-cited domains overall, their relative share has shrunken rapidly as ChatGPT shifts its reliance away from sites that merely summarize information and toward the primary sources that generate it. This trend is further evidenced by dramatic rank changes among the 200 most-cited domains across millions of analyzed answers. PubMed surged from 64th to 4th place among ChatGPT’s most-cited domains, and the FDA debuted at 8th place. Conversely, health websites that traditionally summarize academic research—such as WebMD and Verywell Health—saw their citation shares decline.
To verify that these shifts were not merely artifacts of changing prompt selections, comparative studies analyzing thousands of identical high-volume prompts across multiple months confirmed a consistent pattern. ChatGPT consistently prioritizes original sources over secondary summaries. Consequently, content that relies purely on recapping information published elsewhere is increasingly vulnerable to being bypassed in favor of primary research, hands-on testing, and direct manufacturer data.
Ultimately, the empirical data demonstrates that the vast majority of ChatGPT responses operate without triggering any fan-out searches at all. When the model does search the web, it increasingly bypasses general aggregators to locate the definitive owner of the information, whether that is a regulatory body, a research institution, or a product manufacturer. For digital publishers and brands, securing long-term visibility within AI-generated answers requires either building enough generalized brand prominence to be held directly in the model’s parametric memory or serving as the authoritative, original source that the system discovers during its targeted web retrievals.
