ChatGPT’s Thinking Mode Cites Different Brands

A study published by Semrush, conducted in collaboration with growth strategist Kevin Indig, has produced one of the most practically significant findings in AI search visibility research to date. When ChatGPT shifts from minimal reasoning to high reasoning mode, nearly three in four cited sources change entirely.

The study tested 100 prompts across 20 buyer journeys in four categories: B2B SaaS, finance, consumer tech, and health and lifestyle. Each prompt ran once in minimal reasoning and once in high reasoning. The analysis tracked citation rates, cited domains, and the number of sub-queries ChatGPT ran before forming each answer.

The full report is available on the Semrush blog.

The Core Finding: 25.6% Source Overlap

Across the full test set, only 25.6% of cited domains appeared in both reasoning modes for the same prompts. That figure means a brand visible in ChatGPT’s standard responses has no guarantee of appearing when a user activates Thinking mode on the same question.

This matters because Thinking mode is increasingly the default for complex or high-stakes queries. Users researching software purchases, financial decisions, or health-related questions are more likely to engage high reasoning than users asking quick factual questions. For brands targeting buyers at the consideration and decision stages, Thinking mode is the environment where presence is most commercially relevant.

How High Reasoning Changes Citation Behaviour

The mechanism behind the shift is the number of sub-queries ChatGPT runs before constructing its answer. In minimal reasoning, the model ran 245 web searches across the full 100-prompt test set. In high reasoning, that figure rose to 1,130.

Each sub-query is an opportunity to surface a different source. High reasoning does not simply retrieve more of the same content. It searches more specifically, pulling from documentation pages, government sources, academic references, and official support content that minimal reasoning does not reach.

Citation rates reflected this shift directly. Minimal reasoning cited sources in 50% of responses, with an average of 2.6 citations per answer. High reasoning cited sources in 68% of responses, with an average of 4.5 citations per answer.

Which Source Types Won and Lost

The shift in reasoning depth produced a clear reordering of which content types get cited.

Reddit and user-generated content lost ground significantly. Reddit’s citation share fell from 15% in minimal reasoning to 7% in high reasoning. Review sites and community-generated content dropped from 14.3% to 6%.

Official and institutional content moved in the opposite direction. Government and academic sources rose from 1.9% to 8.8%. Official documentation and support pages grew from 12.4% to 17.5%.

The implication for content strategy is direct. Brands whose AI visibility depends on community mentions, review aggregators, or informal discussion threads are building on a foundation that weakens as users engage more deeply with AI tools. Brands with strong official documentation, structured product pages, and third-party institutional references are better positioned in high reasoning environments.

Comparison Queries Drive the Most Sub-Searches

The study broke down reasoning behaviour by buyer journey stage. At the comparison stage, high reasoning averaged 24 sub-queries per prompt, versus 5.5 for minimal reasoning. Average citations at this stage reached 9.8 per high-reasoning response, versus 5.8 for minimal reasoning.

The study offers a concrete illustration of why. A CRM comparison query in high reasoning mode triggers separate searches for pricing pages, integration documentation, security certifications, support content, and customer references before ChatGPT assembles an answer. Each of those sub-queries can surface a different source.

Brands with comprehensive, accessible content across these dimensions, not just marketing copy but pricing transparency, integration guides, and support documentation, are more likely to appear across multiple sub-queries and accumulate citation frequency within a single response.

Early Citations Carry Further in High Reasoning

One of the more significant findings concerns citation persistence across the buyer journey. In four of the 20 journeys tested, a brand cited at the problem awareness stage was still present at the selection stage in high reasoning. Minimal reasoning showed no instances of full-journey persistence.

High reasoning also reused the same domains more often within individual responses. The same domain appeared multiple times in 51 of the 100 high-reasoning responses, versus 26 of the 100 minimal-reasoning responses.

Taken together, these findings suggest that earning a citation early in a buyer journey, at the point where a user first frames the problem, increases the probability of appearing at later, higher-intent stages when high reasoning is active.

Category Differences Are Significant

The citation lift from minimal to high reasoning varied considerably across the four categories tested.

Finance saw the largest increase, with citation rates rising 28 percentage points in high reasoning. Health and lifestyle rose 24 points. B2B SaaS gained 16 points. Consumer tech moved only 4 points, despite high reasoning running more sub-queries for consumer tech prompts than for any other category.

The consumer tech finding is the most counterintuitive in the study. More searches did not produce more diverse citations. For established consumer categories where brand recognition is already high, high reasoning appears to converge on the same sources regardless of how many sub-queries it runs. The implication is that newer or less dominant brands in consumer tech face a harder path to AI visibility through content alone, while brands in finance and health, where institutional credibility matters more, have clearer levers to pull.

What This Means for AI Search Visibility Strategy

The Semrush study identifies a practical gap in how most brands currently approach AI search visibility. Optimising for appearance in ChatGPT’s standard responses is not the same as optimising for appearance in Thinking mode. The two surfaces behave differently, draw on different source types, and reward different content investments.

Brands that want consistent visibility across both modes need content that performs at the sub-query level, not just at the level of the primary question. That means thorough documentation, structured product and service information, accessible pricing and integration content, and third-party references from institutional sources that high reasoning is more likely to surface.

The 25.6% overlap figure is the number worth keeping in mind. For most brands, the majority of their AI visibility work is currently aimed at an environment that accounts for a shrinking share of the queries that matter most.