A large-scale analysis of ChatGPT’s search behaviour suggests the AI platform is increasingly turning to official websites, original research, and direct product information when retrieving sources for its answers, offering new insights into how businesses can improve their visibility in AI-generated responses.
The study examined 66.6 million responses from free, logged-out GPT-5.6 Luna sessions recorded in August and September, identifying approximately 15 million additional queries generated when ChatGPT searched for information. These follow-up searches, commonly known as query fan-outs, help the system explore different aspects of a user's original request.
The findings indicate that ChatGPT often answers questions without accessing the live web, while its search activity is increasingly focused on specific brands, product prices, specifications and reviews.
Most ChatGPT responses do not trigger web searches
Only 14.4% of the 66.6 million responses analysed resulted in a web search. When searching was necessary, ChatGPT generated an average of approximately 1.6 fan-out queries per response.
The figures suggest that a substantial proportion of answers rely on information already available to the model rather than newly retrieved web content.
The researchers noted that this behaviour has implications for businesses seeking visibility in ChatGPT. If the system does not conduct a search, a company's chances of being mentioned may depend more heavily on what the model already knows about the brand.
The study focused on free, logged-out sessions to examine ChatGPT's default behaviour without the influence of individual chat histories or saved memories. The article also cited industry commentary indicating that more than 90% of ChatGPT's weekly users are on the free plan.
Brand websites dominate targeted searches
The analysis found that 6.8% of fan-out queries used the site: search operator, which restricts a search to a particular website. These queries covered roughly 116,000 different domains.
Among the most frequently targeted sites, approximately 75.9% of site: searches focused on brand websites rather than retailers, government portals, discussion forums or review platforms.
Retailers such as Amazon, Walmart and Costco accounted for 11.9%, while Reddit and Facebook together represented just 1.9%.
This preference reflects the value of first-party information when ChatGPT needs precise details about a company's products or services. Official websites are often the most direct source for pricing, product specifications, and corporate information.
A separate test involving 20,000 US prompts reinforced the pattern. At least one official source appeared in 84.4% of citations for factual questions about prices, product features, and company details. For comparison-based queries, such as deciding which of two products is cheaper, that figure fell to 68.3%.
Official websites accounted for 63.4% of citations for factual questions, compared with 46.4% for comparison queries. The results suggest that ChatGPT draws more heavily on external sources when users need help evaluating competing options.
Consumer brands feature prominently in domain rankings
Eight of the ten most frequently targeted domains in the study belonged to major consumer brands, according to the analysis.
Government websites were also prominent, with domains associated with Brazil, the United Kingdom, France and US agencies appearing among the 150 most frequently targeted sites.
Their presence points to the importance of authoritative, first-hand information, particularly for topics involving health, regulation and other areas where accuracy is especially important.
Reddit remained the ninth most frequently targeted domain in site: searches, despite a significant decline in its share of ChatGPT citations. The distinction suggests that the platform may still retrieve discussions to understand user experiences and opinions, even as it becomes less likely to cite those discussions in its final answers.
Licensed publishers gain a larger share of citations
The research also identified a difference between the websites ChatGPT specifically searches and those it ultimately cites.
Among 17 publishers and platforms identified as having content licensing agreements with OpenAI, Reddit was the only one appearing in the top 150 most frequently targeted site: domains.
However, data from Brand Radar showed that licensed publishers were gaining a greater share of ChatGPT citations. Their combined share among the platform's 200 most-cited domains increased from 7.5% in March to 15.8% in July, before settling at 12.5% in late September.
The findings raise the possibility that some licensed material reaches ChatGPT through retrieval processes other than explicit site: searches. The study referenced observations by other researchers who have identified signs of pre-summarised publisher content and a separate retrieval source described as a possible licensed-content tier.
These observations offer potential reasons why some publishers receive citations frequently without appearing prominently in targeted website searches.
Searches for prices and reviews are increasing
The language used in ChatGPT's fan-out queries also changed between August and September, with searches increasingly focused on commercial information.
Queries containing the word “price” rose from 6.04% to 8.88% of fan-outs, a relative increase of approximately 47% in one month. Searches involving reviews increased by about one-third, while product specifications entered the top 50 terms for the first time.
By contrast, queries containing “meaning” and “definition” each declined by roughly one-fifth, although both terms represented relatively small portions of the overall dataset.
The shift suggests that ChatGPT's retrieval activity is placing greater emphasis on information that helps users assess products and make purchasing decisions.
For businesses, maintaining accurate and accessible product details, current pricing, and useful evidence of product performance may therefore be important to how their information is surfaced in AI-powered search.
Reddit citations fall sharply alongside Wikipedia and YouTube
A separate dataset tracking ChatGPT citations in the United States between March and September 2026 revealed a substantial decline in citations to several prominent online platforms.
The analysis, based on approximately 2.4 million to 3.8 million answers per month, found that Reddit's citation frequency dropped from roughly one in every five answers in March to one in every 43 by late September.
Wikipedia also experienced a steep decline, falling from approximately one citation per seven answers to one per 35. YouTube citations decreased from around one in every 370 answers to one in every 500.
Although these platforms remained among ChatGPT's most-cited domains, their shrinking share indicates a significant change in the composition of the sources appearing in responses.
The results broadly align with separate industry observations that social media, reference sites and news publishers lost citation share during parts of the period.
Original research and primary sources gain ground
The researchers found that some of the citation share lost by Reddit, Wikipedia and YouTube shifted towards research institutions, regulators, brands and sources offering direct, original evidence.
PubMed, for example, climbed from 64th place among ChatGPT's most-cited domains in March to fourth place in September. The US Food and Drug Administration also entered the rankings at eighth place, while health information publishers such as WebMD and Verywell Health declined.
To test whether the pattern was simply caused by changes in the prompts being tracked, the researchers compared approximately 25,000 prompts that appeared among the highest-volume queries in both March and September. The same general direction of movement appeared across the matched prompts.
While the mix of questions can influence which sources ChatGPT retrieves and cites, the findings suggest that original sources are gaining ground against websites that primarily summarise existing information.
What the findings mean for AI visibility
The research points to several considerations for businesses seeking to appear in ChatGPT responses.
First, a strong official website remains important because the platform frequently seeks direct sources for product facts, specifications, pricing and company information.
Second, businesses may benefit from publishing information that cannot easily be replicated elsewhere, including original research, first-hand testing results and detailed product documentation.
Third, brands should recognise that factual questions and comparison-based prompts can produce different citation patterns. Official websites may be especially influential when users request straightforward information, while independent sources can become more important when ChatGPT evaluates alternatives.
Finally, companies that rely heavily on citations from Reddit, Wikipedia, YouTube, or third-party explainers may need to monitor whether those sources continue to contribute as much visibility as before.