Semrush study finds ChatGPT reasoning modes cite different sources

ChatGPT’s Thinking mode pulled a different citation stack than Instant mode in a Semrush study published June 30, 2026, and the gap was wide enough to challenge any marketer treating ChatGPT as one stable source of traffic. Semrush and Kevin Indig ran 100 prompts through GPT-5.2 in both modes, producing 200 responses across 20 buyer journeys in B2B SaaS, Finance, Consumer Tech and Health/Lifestyle, with prompts mapped to Problem, Exploration, Comparison, Validation and Selection stages.

Only 25.6 percent of cited domains overlapped between the two modes. Citation rates rose from 50 percent in Instant mode to 68 percent in Thinking mode, average sources per cited response increased from 2.6 to 4.5, and internal fan-out expanded 4.6 times. Thinking mode drew from 173 unique domains, compared with 127 in Instant, and 99 domains appeared only when reasoning was turned up. For visibility teams, that means a prompt test run in one mode can miss most of the source set that appears in the other.

The source mix changed as well. Reddit and other user-generated content sites lost roughly half their citation share in Thinking mode, while government, academic and official documentation sources gained ground. Semrush said that in 4 of the 20 buyer journeys, a brand held through the conversation from the first prompt to the last under high reasoning, while that kind of full-funnel persistence was rare in Instant mode. The pattern matters for brands trying to win both the first exploratory query and the final validation prompt, because the higher-reasoning path is closer to a research workflow than a simple answer engine.

OpenAI says ChatGPT search can provide fast, timely answers with links to relevant web sources and can automatically search the web based on what users ask. OpenAI also says GPT-5.2 is available in Instant, Thinking and Pro variants. Semrush’s wider AI Visibility Index, published in June 2026, was built on 126 million real U.S. AI search prompts across 22 industries and four AI platforms, underscoring how quickly visibility measurement has moved beyond a single-query snapshot. The practical takeaway is blunt: a brand that shows up in Instant mode can disappear in Thinking mode, so share-of-voice claims need testing across both.