Study Analyzes 15 Million ChatGPT Fan-Out Queries and Retrieval Shifts
Research reveals direct-to-brand searching patterns and major citation declines for aggregators.
1Analyzing Millions of Retrieval Paths
Ahrefs examined 66.6 million free, logged-out GPT-5.6 Luna responses alongside 15 million fan-out queries during August and September 2026. This setup isolated ChatGPT's default operational behavior from personalized user chat histories. Researchers sought to decode how the platform retrieves live web data to construct factual conversational answers.
2Low Overall Web Search Triggering
Only 14.4% of total tracked model responses actually initiated an external web search. When an answer required external information, the system generated roughly 1.6 fan-out search queries per prompt. This restrained retrieval suggests OpenAI prioritizes internal parametric knowledge over live fetching to control massive infrastructure and computational expenses.
3Heavy Preference for Brand Domains
Among generated fan-out searches, 6.8% utilized site-specific search operators across roughly 116,000 distinct domains. Over 75.9% of these targeted individual brand properties rather than third-party aggregators or forums. Retailers captured only 11.9% of targeting, demonstrating an algorithmic preference for first-party data sources when pulling specifications or pricing details.
4Factual versus Comparison Question Behavior
A separate test across 20,000 prompts revealed distinct retrieval patterns between query formats. Factual queries cited official brand sources in 84.4% of cases, with direct domain citations reaching 63.4%. Conversely, comparative queries saw official citations drop to 46.4%, forcing ChatGPT to gather third-party evaluations to weigh differing options.
5Official Entities Dominate Direct Lookups
Consumer manufacturers represented eight of the ten most targeted domains within direct site queries. Regulatory and governmental portals like the FDA, SEC, and UK public services appeared frequently within the top 150 lookups. This pattern confirms that AI retrieval engines heavily favor verified institutional databases for critical safety and corporate disclosures.
6Growth of Partner Publication Citations
While media partners holding OpenAI licensing deals rarely triggered explicit site searches, their overall citation visibility expanded dramatically. Partner publications accounted for 15.8% of top 200 cited domains in July, up from 7.5% in March. This indicates licensed content may sit pre-indexed within the core architecture before live querying occurs.
7Surge in Commercial Keyword Modifiers
Analyzing fan-out query language revealed shifting internal intent between August and September. Fan-out searches containing the modifier price increased relative volume by 47%, jumping to 8.88% of all queries. Search phrases seeking reviews grew by one-third, while abstract definition queries declined by approximately twenty percent over the observed window.
8Severe Decline in Reddit Mentions
Data tracking US citations between March and September 2026 uncovered dramatic declines for major aggregate platforms. Reddit mentions dropped from 1 in 5 responses down to 1 in 43, representing an 88% collapse. Similar downward trends struck Wikipedia and YouTube as retrieval pipelines systematically demoted broad user-generated discussion directories.
9Primary Sources Supplant General Aggregators
The citation share lost by general forums migrated directly to specialized primary sources. Scientific indexing repository PubMed climbed to the fourth most-cited domain, alongside major gains for regulator domains. Conversely, general consumer health summarizers like WebMD lost substantial visibility across 25,000 identical high-volume tracking prompts tested across identical timeframes.
10Topic Context Alters Retrieval Approaches
Vertical-specific analysis demonstrates that retrieval behaviors change based on commercial category. Examining queries around running shoes revealed heavy usage of third-party review terms and comparative keywords over brand terms. In contrast to broad trends, specialized performance products force the retrieval engine to pull third-party validation rather than solely first-party sites.
11Methodology Differences Explain Discrepant Metrics
Previous industry studies reported higher search frequencies per prompt because they relied on commercial transactional queries. Analyzing unauthenticated, general-demand user journeys revealed significantly lower search trigger rates across standard sessions. Continuous tracking of model retrieval updates remains essential as OpenAI frequently modifies core fan-out triggers and ranking heuristics without notice.
12Strategic Imperatives for Brand Content
Marketing teams face a bifurcated AI discovery channel requiring direct authority or primary ownership. Content that merely summarizes third-party data faces rapid extinction across generative search citations. Brands must publish machine-readable first-party product facts, proprietary research, and earn authentic third-party testing reviews to maintain presence across evolving generative answer engines.