ChatGPT Reasoning Modes Alter Brand Citations and Search Behavior

ChatGPT's high-reasoning mode functions as a distinct search surface, influencing brand visibility by citing different domains compared to its minimal reasoning mode, according to a Semrush analysis involving Kevin Indig. This high-reasoning mode also initiated nearly five times more web searches than the minimal reasoning mode. Specifically, only 25.6% of cited domains overlapped between the two reasoning modes for identical prompts, indicating a substantial shift in information sources. Nearly three-quarters of sources changed when ChatGPT transitioned from instant answers to its "Thinking" style.

The "Thinking" mode demonstrated increased source utilization, with citation rates rising from 50% in minimal reasoning to 68% in high reasoning. Furthermore, cited answers in high-reasoning mode incorporated more individual sources, averaging 4.5 citations per response, up from 2.6 in minimal reasoning. Across the test set, high reasoning executed 1,130 web searches, contrasting sharply with the 245 searches performed by minimal reasoning. User-generated content platforms like Reddit saw a decline in citation share, dropping from 15% to 7% when high reasoning was active. Similarly, user-generated content and review sites experienced a decrease from 14.3% to 6% in citations. Conversely, government and academic sources saw their citation share grow from 1.9% to 8.8%, and official documentation and support pages increased from 12.4% to 17.5%.

Comparison prompts were a significant driver of the increased search activity in high-reasoning mode. During comparison stages, high reasoning averaged 24 sub-queries per prompt, compared to 5.5 sub-queries for minimal reasoning. Average citations also peaked during these comparison scenarios, reaching 9.8 citations per high-reasoning response, versus 5.8 for minimal reasoning. For instance, a prompt comparing CRM systems could trigger separate searches for pricing, integrations, security, support pages, and documentation before ChatGPT compiles its final answer.

High-reasoning mode also exhibited greater persistence in carrying brand mentions throughout a user's research journey. In four out of 20 tested journeys, a brand cited at the initial problem-identification stage remained present in the final selection stage. This full-journey persistence was not observed with minimal reasoning. Additionally, high reasoning demonstrated a tendency to reuse the same domains more frequently within a single answer. The same domain appeared multiple times in 51 out of 100 high-reasoning responses, compared to 26 instances in minimal reasoning responses.