ChatGPT’s Search Behavior Revealed: A Deep Dive into AI Information Retrieval and Its Implications for AI Visibility

The Search Engine Optimization (SEO) industry has been intensely scrutinizing ChatGPT’s mechanisms for searching and identifying its sources over recent months. To gain firsthand insight into how ChatGPT retrieves information and what this signifies for AI visibility, a comprehensive analysis was conducted on 66.6 million free, logged-out GPT-5.6 Luna responses and their associated 15 million fan-out queries from August and September. Free sessions are considered more representative of typical user interaction, as over 90% of ChatGPT’s weekly users utilize the free plan, according to Olivier de Segonzac, co-founder of RESONE0, as reported by Search Engine Land. Logging out also removes chat history and memory, allowing for an examination of ChatGPT’s default behavior, uninfluenced by individual user interaction patterns.

Key Findings

An important clarification was made regarding the initial data: 14.4% of responses contained fan-out queries, not that 14.4% of responses searched the web. This correction has been implemented, with apologies for any confusion. The underlying data and other findings remain unchanged.

Only 1 in 7 ChatGPT Responses Contained Fan-Out Queries

The analysis revealed that only 14.4% of the 66.6 million ChatGPT responses included fan-out queries. These responses featured an average of approximately 1.6 fan-out queries each. This implies that a webpage must align with one or two specific searches within these fan-outs to be considered by ChatGPT; otherwise, it will not be seen.

76% of ChatGPT’s ‘site:’ Searches Target a Brand Site

Out of the 15 million fan-out queries analyzed, 6.8% employed the site: operator to target a single website. Among the approximately 160 most frequently searched sites, three-quarters of these site: searches were directed at brand websites. Brand sites are defined as company-specific websites (e.g., apple.com) rather than retailers, government sites, forums, or review platforms. Retailers such as Walmart, Costco, and Amazon accounted for 11.9%, while Reddit and Facebook collectively represented only 1.9%. The terminology used within these fan-out queries further emphasizes a focus on official, primary information. Words like "official" and "price" were the most common, following "2026," and showed a notable month-over-month increase.

This trend is logical, as brand websites are the most direct source for product pricing and specifications. ChatGPT is more inclined to search beyond official sources when a question lacks a single definitive answer. In a separate test involving 20,000 U.S. prompts, 84.4% of factual questions regarding pricing, product features, or company details cited at least one official source. Conversely, for comparative questions like "A vs. B" or "which is cheaper," this figure dropped to 68.3%. The mix of cited sources also shifted: 63.4% of citations were to official websites for factual questions, compared to 46.4% for comparison questions, where a larger proportion of citations went to other types of sites. This indicates ChatGPT’s reliance on official sources for factual queries and a greater propensity to consult other sites for comparative analysis.

To ensure a brand’s website is discoverable for factual inquiries, it’s crucial that pricing, specifications, and support details are presented in plain text on accessible pages.

Consumer Brands Dominate ChatGPT’s Top 10 ‘site:’ Searched Domains

The most frequently targeted domains by ChatGPT using the site: operator are major consumer brands. While these rankings are based on the study’s dataset and prompts, the extensive data of around one million site: queries provides a strong indication of ChatGPT’s preferred domains. These are predominantly brands that sell products, making their official sites the natural source for detailed information. Government domains also feature prominently. Sites like gov.br, gov.uk, France’s service-public.fr, and U.S. agencies such as NHTSA, FDA, and SEC appear within the top 150. This aligns with observations by Lily Ray, Founder of Algorythmic, that ChatGPT favors official sources for Your Money or Your Life (YMYL) questions. Notably, Reddit remains the ninth most site: searched domain, consistent with findings by Suganthan Mohanadasan, co-founder of Snippet Digital, suggesting ChatGPT still accesses Reddit for authentic user opinions, despite citing it less frequently.

ChatGPT Rarely ‘site:’ Searches OpenAI’s Partners, Yet Their Citation Share Has Doubled

Among the 17 publishers and platforms with current licensing agreements with OpenAI, Reddit is the sole entity within the top 150 most site: searched domains. One potential explanation is that ChatGPT may have partner content pre-loaded and readily accessible. While site: searches indicate what ChatGPT consults during query fan-out, they do not reflect what it ultimately cites. The cited domains present a different narrative, with partner publications appearing frequently. In Brand Radar’s analysis, partner titles among ChatGPT’s 200 most-cited domains saw their combined citation share more than double, from 7.5% in March to 15.8% in July, and maintained 12.5% in late September. Securing coverage in these partner publications is therefore advisable, as ChatGPT is increasingly citing them.

"Price" and "Review" Fan-Outs Increased by Up to 47% in a Month

The terms incorporated into ChatGPT’s fan-out queries offer clues about the types of pages it seeks. Between August and September, there was a rise in fan-out queries related to prices, product specifications, and reviews, with a slight decrease in those concerning definitions. A gain of 1-2 percentage points might seem minor, but these terms started from a low base, and their relative increase signifies significant growth. For instance, "Price" increased from 6.04% to 8.88% of fan-outs, representing a nearly 47% relative rise in one month. "Reviews" grew by a third, and "specifications" entered the top 50 for the first time. Conversely, "Meaning" and "definition" each declined by about a fifth, though from small shares, suggesting a dip rather than a sustained trend.

It is recommended to assess which of a brand’s pages align with these growing query types. Tools like Ahrefs Brand Radar can help identify prominent fan-out query types within an industry by filtering the AI responses report for ChatGPT and a specific market, then examining the "Fan-out queries" column for terms like "price," "specifications," or "reviews."

ChatGPT’s Reddit Citations Dropped 88% in Six Months

In addition to fan-out data, an analysis of ChatGPT’s citation sources was conducted. This analysis drew from a separate dataset: Brand Radar’s ChatGPT citations in the U.S. from March to September 2026, covering 2.4 to 3.8 million answers monthly. Snippet Digital reported an 86% drop in Reddit citations in August, and data from Aleyda Solís indicated that social, reference, and news sites experienced the largest citation share decline between July and August. Brand Radar data corroborates this trend across Reddit, Wikipedia, and YouTube. In March, approximately one in five ChatGPT answers cited Reddit, a figure that fell to about one in 43 by late September. Similarly, Wikipedia’s citation share decreased from one in seven answers to one in 35, and YouTube’s from about one in 370 to one in 500 over the same period. While these sites remain among ChatGPT’s most cited domains, their shrinking share warrants attention for entities that rely heavily on them for AI visibility.

ChatGPT is Shifting from Summarized Information to Original Sources

Citations lost by Reddit, Wikipedia, and YouTube have been redirected to research bodies, regulators, brands, and hands-on testers, while explainer and aggregator sites have seen a decline. A chart illustrating significant rank changes among ChatGPT’s 200 most-cited domains between March and September reveals this shift. PubMed climbed from 64th to 4th, and the FDA entered at 8th, while health sites like WebMD and Verywell Health, which summarize research, experienced a drop. This pattern was consistent even when analyzing identical prompts from both periods. The data suggests ChatGPT is increasingly prioritizing original sources over those that summarize information. Content that primarily synthesizes existing publications may be less likely to be cited, with ChatGPT favoring original research, hands-on testing, and direct product facts.

How to Analyze Fan-Out Queries in Your Industry

Analyzing fan-out queries in bulk provides insights into what ChatGPT consults before recommending a brand within a specific industry, guiding content creation and coverage acquisition strategies. This analysis can be performed using tools like Ahrefs Brand Radar. By filtering the AI responses report by topic, platform (ChatGPT), and market, and then examining the "Fan-out queries" column, users can identify patterns. For instance, an analysis of "running shoes" revealed that "best" appeared 206 times, "2026" 160 times, "review" 117 times, "vs" 105 times, and "official" 65 times. This indicates that for running shoes, ChatGPT primarily seeks recent, third-party reviews and comparative analyses, a different pattern than the overall data suggests. For a running shoe brand, review coverage and comparison pages would therefore be more impactful than product pages alone.

Tools like Ask Ahrefs can automate the grouping and counting of fan-out queries, saving significant time. Users can input prompts such as "Pull the ChatGPT fan-out queries from Brand Radar for my tracked prompts/prompts about [your category]. Group them by intent (pricing, specs, reviews, comparisons, how-to), count the most common terms and site: domains, and tell me which groups my site answers and which it doesn’t." This allows for regular tracking of evolving query types. The analysis can reveal the breakdown of fan-out queries by topic, compare a brand’s site: presence against competitor domains, and identify content gaps.

Comparison with Other Studies

The current data aligns with other research indicating ChatGPT’s trend towards prioritizing official sources, site: searches, and a reduction in citations. However, the scale of fan-outs and site: searches per prompt in this study is broader, stemming from real search demand across diverse topics and countries, unlike some studies that focused on commercial "best X" prompts that naturally trigger brand lookups. The tracking of free, logged-out sessions also shows less searching than paid or reasoning models. These findings represent a snapshot of a specific model between August and September 2026, acknowledging that OpenAI continuously updates ChatGPT’s information retrieval methods. Regular tracking of fan-outs is crucial for identifying changes as they occur.

Final Thoughts

The majority of tracked ChatGPT responses did not involve fan-out queries. When searches are performed, ChatGPT increasingly directs them to the owners of the information, such as the brand setting a price. Consequently, entry into ChatGPT’s answers is achieved either by being sufficiently well-known to appear directly or by being the original source that ChatGPT identifies during its searches. Pages that merely summarize existing content are currently experiencing a decline in citations, making pages that offer unique content, such as original research or specific product facts, more valuable. However, given the dynamic nature of AI model updates, continuous monitoring of a brand’s presence in fan-out queries and AI responses is essential.