A single ChatGPT visibility score is not a dependable GEO measurement if it combines, assumes, or observes only one reasoning mode. In a Semrush and Kevin Indig analysis, just 25.6% of domains cited for the same prompts overlapped between minimal and high reasoning. Put differently, nearly three quarters of cited sources changed when the mode changed. semrush.com
That result changes the unit of analysis for GEO work. The question is no longer simply whether a brand is cited for a prompt in ChatGPT. It is whether the brand is cited for that prompt in a particular reasoning environment, and whether it remains visible as the user moves through the questions that make up a buying journey. The study does not establish a universal rule for all prompts, industries, or future model behavior. But it gives strong reason not to treat an observation from one mode as a proxy for the other.
The 25.6% figure is a measurement warning, not just a citation statistic
The underlying test ran 100 prompts through GPT-5.2 twice: once in minimal reasoning and once in high reasoning, producing 200 responses. The prompts covered 20 buyer journeys across B2B SaaS, finance, consumer tech, and health and lifestyle, with five stages per journey: problem, exploration, comparison, validation, and selection. semrush.com
Within ChatGPT’s interface, Semrush describes minimal reasoning as Instant mode and high reasoning as Thinking mode. Instant is the default fast-response experience; Thinking is intended for more complex, multi-step tasks. semrush.com The practical issue is straightforward: a brand can look visible in the fast experience while disappearing from the research-heavy one, even when the user’s wording does not change.
Search Engine Land characterizes the high-reasoning mode as a different search surface for brand visibility, based on the same Semrush analysis. searchengineland.com That is the useful framing for GEO teams. A mode is not a cosmetic response setting when it alters the source pool, citation frequency, and amount of web research behind an answer.
There is one reporting inconsistency worth noting. The primary Semrush extract says the study used 100 prompts, while ContentMarketing.ai describes Semrush as having analyzed “hundreds of prompts.” semrush.com contentmarketing.ai The supplied sources do not resolve that difference. For interpreting the detailed figures, I would use the primary report’s stated 100-prompt methodology and avoid extrapolating its findings into a population-level citation rate.
Why one-mode testing produces false confidence
High reasoning did not merely swap a few citations. Citation rate rose from 50% under minimal reasoning to 68% under high reasoning, while average citations per cited response rose from 2.6 to 4.5. High reasoning also used 1,130 web searches across the test set, compared with 245 for minimal reasoning. searchengineland.com
That means a single-mode test can fail in two different ways. Testing only Instant may understate the number of opportunities available when ChatGPT performs deeper research. Testing only Thinking may overstate how often a brand appears in the default fast-response experience. Aggregating the two into one score hides both differences.
The underlying source mix matters as much as volume. Reddit’s citation share fell from 15% in low reasoning to 7% in high reasoning. User-generated-content and review sites fell from 14.3% to 6%, while government and academic sources rose from 1.9% to 8.8%; official documentation and support pages rose from 12.4% to 17.5%. semrush.com
For a GEO practitioner, this makes a source-level report essential. “We are cited” does not answer the operational question. You need to know which of your own pages, which third-party references, and which source classes are present in each mode. A strategy dependent on community discussion and review pages may perform differently from a strategy supported by official documentation, reference material, and authoritative external sources. The study shows that the distribution changes; it does not prove that any one asset type will win every prompt.
Mode-level testing needs to be paired with query-level testing
The mode gap is not uniform across buyer-journey stages. At the problem stage, high reasoning’s citation rate was 35 percentage points higher than minimal reasoning. By validation, the gap narrowed to five points. semrush.com A dashboard that averages all prompts can therefore conceal precisely where a brand lacks visibility.
Comparison prompts are especially important. In the test, high reasoning averaged 24 sub-queries per comparison prompt, versus 5.5 in minimal reasoning. It also averaged 9.8 citations per response at that stage, compared with 5.8 under minimal reasoning. searchengineland.com The reported CRM example illustrates why: a comparison can lead to research into pricing, integrations, security, support pages, and documentation rather than just the literal comparison query. searchengineland.com
This does not mean optimizing for an imagined list of sub-queries is a substitute for testing. It means the original user prompt is an insufficient measurement unit. The answer may depend on several research paths, and a page that never appears for the headline comparison could still support visibility through one of those paths. Conversely, a brand that appears in a quick answer may not have the documentation or third-party corroboration that deeper research pulls into the final response.
What a defensible ChatGPT GEO measurement design looks like
The study supports separating observations, not collapsing them. For a working program, I would structure testing around a stable prompt set drawn from real category, comparison, validation, and selection questions, then run that same set in Instant and Thinking modes. Report the results as separate environments before calculating any roll-up.
- Track citation presence by prompt and mode. A brand-level percentage alone cannot show whether the gap is concentrated in early research, comparisons, or selection questions.
- Capture cited domains and pages, not only brand mentions. The central finding is domain turnover between modes; source evidence is therefore part of the measurement, not an appendix.
- Compare competitors in the same prompt-mode cell. A competitor’s apparent lead may exist only in one mode, while a combined score can disguise that distinction.
- Review source-type dependence. If visibility is concentrated in UGC or reviews, test whether it holds when Thinking shifts toward government, academic, and official documentation sources. semrush.com
The appropriate output is a matrix: prompt, journey stage, mode, citation occurrence, source/domain, and competitor presence. That is less tidy than a single AI-visibility number, but it represents the behavior the study actually measured.
Early-funnel visibility may be more consequential in Thinking mode
The study also found a continuity effect that makes stage-level testing more important. In four of 20 journeys, a brand cited at the problem stage in high reasoning still appeared at selection. Minimal reasoning showed no such full-journey persistence. searchengineland.com Within individual responses, the same domain appeared multiple times in 51 of 100 high-reasoning answers, versus 26 of 100 minimal-reasoning answers. searchengineland.com
The evidence is limited: four journeys are not a general guarantee that awareness-stage citations drive selection-stage visibility. But it is enough to challenge the habit of measuring only bottom-funnel recommendation prompts. If Thinking mode can carry a cited brand across parts of a journey in the tested cases, a GEO program that ignores problem and exploration queries may miss where later visibility begins.
Industry differences reinforce the same point. Finance had the largest reported citation-rate increase under high reasoning, at 28 percentage points; health and lifestyle rose 24 points, B2B SaaS 16 points, and consumer tech four points. searchengineland.com No universal mode adjustment is justified. Your category’s prompt mix must be tested rather than inferred from an overall average.
The practical conclusion: ChatGPT is not one measurement surface
The 25.6% overlap result does not tell us that every ChatGPT answer is volatile, that citations are random, or that GEO measurement is impossible. It says something narrower and more useful: in this test, changing reasoning depth materially changed the cited domains for the same prompts. semrush.com
That makes single-mode measurement unreliable as a representation of ChatGPT visibility. GEO teams should preserve the distinction between Instant and Thinking, test the same queries in both, and locate the gaps by journey stage and source type. A single score may still be convenient for reporting upward. It should not be the diagnostic system used to decide what content, documentation, or third-party authority work to fund.