The mechanic: why reasoning mode finds different sources
ChatGPT's standard fast mode and its Thinking (reasoning) mode are not two versions of the same answer. They are two different retrieval strategies that produce structurally different citation sets.
Fast mode runs a single or small number of web searches, synthesises quickly, and returns an answer. The sources that survive a single fast retrieval pass are typically the most-linked, most-visible pages for that query: listicles, well-indexed review sites, high-authority aggregators. That is the surface most generative engine optimisation practice has been built around.
Reasoning mode works differently. Before generating a response, it iterates through multiple sub-queries, checking each partial answer against additional sources, cross-referencing claims, and running follow-up searches where the first answer raises new questions. The July 2026 Semrush study by analyst Kevin Indig tested 100 prompts across 20 buyer journeys and found high reasoning ran 1,130 web searches across the test set versus 245 for minimal reasoning. That is roughly 4.6 searches per prompt in reasoning mode versus under 2.5 for fast mode, and the searches are qualitatively different: they are evidence-seeking, not just keyword-matching.
The practical consequence: reasoning mode has many more opportunities to find the primary source behind a claim rather than the aggregator that cited it. A brand that appears prominently in SEO-indexed listicles can appear in fast-mode answers and vanish from reasoning-mode answers simply because the primary-source layer is occupied by a competitor with better documentation.
This is distinct from the general GEO framework covered in our GEO post. That post covers the baseline signals: entity clarity, self-contained chunks, named statistics. This post covers what happens when the model has time to check those signals against a wider source set.
What the data shows: 25.6% overlap across 100 prompts
The Semrush and Indig study tested B2B SaaS, finance, consumer technology, and health and lifestyle categories. Each prompt ran once in minimal reasoning and once in high reasoning. The analysis tracked citation rate, cited domains, source type, and sub-query fan-out.
| Metric | Fast mode (minimal reasoning) | Reasoning mode (high reasoning) |
|---|---|---|
| Source overlap with same prompts | Baseline | 25.6% overlap with fast mode |
| Citation rate (% of responses citing any source) | 50% | 68% |
| Average citations per cited response | 2.6 | 4.5 |
| Total web searches across 100 prompts | 245 | 1,130 |
| Avg sub-queries at comparison stage | 5.5 per prompt | 24 per prompt |
| Avg citations at comparison stage | 5.8 per response | 9.8 per response |
| Full-journey brand persistence (problem to selection) | 0 of 20 journeys | 4 of 20 journeys |
| Same-domain reuse within a single response | 26 of 100 responses | 51 of 100 responses |
Three patterns stand out. First, the overlap figure is the headline: 25.6% means fast-mode and reasoning-mode citation sets are more different than they are alike, across the same prompts. A brand cannot assume its fast-mode visibility translates. Second, reasoning mode cites more sources and cites them more often, so the total available citation positions are larger. Third, brand persistence across a buyer journey only appears in reasoning mode. In minimal mode, no brand maintained presence from problem-awareness through to selection in any of the 20 tested journeys. Reasoning mode produced journey persistence in four. That is not a statistical guarantee, but it is a direction.
The comparison stage numbers are worth a separate note. At the point in a buyer journey where someone is comparing options, high reasoning averaged 24 sub-queries per prompt versus 5.5 for minimal. That is the moment a buyer is closest to a purchase decision. It is also the moment where reasoning mode diverges most sharply from fast mode in both search volume and source mix.
One caveat on the study: Search Engine Land is owned by Semrush, and the study was conducted by Semrush. That relationship does not invalidate the findings, but it warrants noting. The methodology (100 prompts, 20 buyer journeys, per-mode tracking of citation domains and query fan-out) is described in enough detail to be reproducible. We treat the cited figures as directionally sound practitioner evidence, not as platform-published specifications.
Which source types gain and which lose in reasoning mode
The citation shift is not random. Reasoning mode changes the mix toward the types of sources that survive cross-verification. Review sites and user-generated content lose ground. Primary sources gain.
| Source type | Fast mode share | Reasoning mode share | Direction |
|---|---|---|---|
| Government and academic | 1.9% | 8.8% | Up +6.9 pp |
| Official documentation and support pages | 12.4% | 17.5% | Up +5.1 pp |
| User-generated content and review sites | 14.3% | 6.0% | Down -8.3 pp |
| 15.0% | 7.0% | Down -8 pp |
The government and academic figure going from 1.9% to 8.8% is the sharpest relative shift. A 4.6x increase in a source type's citation share is meaningful. The mechanism is intuitive: more sub-queries give the model more opportunities to reach .gov, .edu, and peer-reviewed sources that a single fast retrieval pass would bypass in favour of more-linked pages.
Reddit's drop from 15% to 7% is the most practically significant shift for consumer and B2B brands that have invested in Reddit community presence as a GEO signal. User-generated content performed well in fast-mode citation studies partly because it is highly indexed and extensively linked. Extended reasoning reduces that advantage: the model has more time to find the original claim behind the Reddit thread, rather than citing the thread itself.
Official product documentation tells a cleaner story. A brand whose help centre, API docs, or methodology pages are well-structured and publicly indexed is better positioned for reasoning-mode citations. Documentation is primary source material by definition. The AI visibility optimisation work we do at leapbuzz has always included structured entity documentation as a core deliverable, and this data reinforces that priority.
Finance, health, and B2B SaaS see the biggest citation shift
The Indig study tracked citation rate changes by category. Consumer technology barely moved at plus 4 percentage points. Finance moved 28 percentage points. Health and lifestyle moved 24 percentage points. B2B SaaS moved 16 percentage points.
The category variation is not surprising once you think about what reasoning mode is doing. For a consumer tech query like "best wireless headphones," a fast retrieval pass already finds the canonical review sites. Reasoning mode runs more sub-queries but frequently lands on the same domains because those are genuinely the most-cited, most-verified sources for that query type. The citation surface is already settled.
Finance and health are different. A buyer researching a financial product or a health condition is running queries where the right answer depends on regulation, clinical evidence, or jurisdiction. Fast mode may pull an aggregator. Reasoning mode has enough sub-query budget to reach the regulator, the clinical study, or the official product disclosure instead. That is where financial services brands with a solid documentation and compliance-content layer have an unexpected advantage over competitors whose content is better-written but less-primary.
The B2B SaaS category result is a practical note for Singapore, Australia, and North American SaaS marketers. A 16-percentage-point increase in citation rate when moving from fast to reasoning mode means reasoning-mode answers are significantly more likely to cite someone. The question is who. B2B buyers making software purchase decisions are the buyers most likely to use reasoning mode, because the complexity of the decision warrants it. The category is competitive to be cited in, and the stakes per citation are high.
Content properties that correlate with reasoning-mode citations
OpenAI does not publish a source-selection specification for reasoning mode. What follows is a combination of what the Semrush data implies, practitioner observation from visibility work we do with clients in Singapore, Australia, and North America, and adjacent findings from generative engine optimisation research. These are observations, not platform-confirmed ranking signals.
- Primary-source attribution, on the page. Content that names a specific study, regulatory body, or vendor documentation as the source of a claim gives a reasoning model something to cross-check. Content that asserts a fact without attribution is a dead end for an extended reasoning pass. The model cannot verify it by running a sub-query; it can only take it on trust or bypass it. Verified claims survive; asserted claims often do not.
- Named methodology sections. The Semrush study itself is a useful template: it describes sample size (100 prompts), scope (20 buyer journeys), categories (B2B SaaS, finance, consumer tech, health), and the specific variables measured (citation rate, domain overlap, sub-query count). A content piece that describes how a finding was reached, not just what the finding is, gives a reasoning model enough structure to evaluate whether the finding is trustworthy relative to other sources it finds.
- Table extractability. Data in a properly structured HTML table with named columns, a caption, and a unique identifier is easier to include in a multi-source synthesis than the same data in a paragraph. The Semrush study's citation numbers appear in this post as a table. A reasoning model running sub-queries on the topic will be able to extract those numbers as discrete, citable data points. Buried in prose, they become approximate.
- Self-contained 100-150 word chunks. This is the standard GEO recommendation from Princeton's 2023 research (arXiv:2311.09735), and it holds for reasoning mode as well. The difference is that reasoning mode has more budget to find and assemble multiple chunks. A brand with five well-structured, primary-source-backed chunks across its content is better positioned than one with fifteen loosely attributed paragraphs.
- Entity consistency across your own site. If your brand name, product names, and category terms appear consistently across your documentation, blog content, and schema markup, a reasoning model that finds your brand through one sub-query can confirm and reinforce that signal through subsequent sub-queries. Inconsistent entity naming breaks that confirmation loop.
The comparison-stage sub-query count (24 per prompt in reasoning mode) is worth returning to here. A buyer comparing two financial products or two software vendors will trigger 24 sub-queries in reasoning mode. Each of those sub-queries is a separate opportunity for your content, your documentation, or a third-party reference to you to surface. Building content that maps to each stage of the buying journey, not just the "what is X" stage, is the practical implication.
Our related post on the agentic shortlist economy covers the downstream commercial consequence: AI agents making shortlist decisions for buyers operate with a fixed citation budget, and the brands that built a reasoning-mode presence earlier pay a lower cost per shortlist inclusion as that market matures.
How to test your own citation presence across both modes
The measurement infrastructure for AI citation share has improved significantly since mid-2026. Microsoft Clarity now tracks AI citation share directly (we cover that in detail in the Clarity citation reporting post). But Clarity only measures Microsoft Bing and Copilot surfaces. For ChatGPT mode-specific testing, you need a manual prompt-set approach.
Step-by-step: test your citation presence across ChatGPT fast and reasoning modes
What you need: a ChatGPT account with access to the Thinking/reasoning toggle (available in ChatGPT Plus and above as of July 2026), a spreadsheet, and roughly three hours for the initial run.
- Build a prompt set that covers your full buyer journey. Map prompts to at least three stages: problem-awareness ("what causes X"), comparison ("best tools for X vs Y"), and selection ("which X is right for [specific use case]"). Ten to twenty prompts per stage is a reasonable starting set. Use the actual phrases your buyers type, pulled from your search console data or keyword research, not invented phrases.
- Run each prompt in standard mode first. Record the full response text, every domain cited, and whether your brand or domain appears. Log the number of cited sources per response. Do not use browser extensions or API access for this first pass; use the standard ChatGPT interface to match what your buyers experience.
- Enable the Thinking toggle and run the same prompts again, in a fresh session. Do not continue from the same conversation thread. Record the same fields: response text, cited domains, brand appearance, source count. Also note whether the response structure feels different (more sub-topic sections, more explicit source attribution in-text).
- Compare domain sets, not just brand mentions. A brand can appear in a reasoning-mode answer without being directly cited, if it is referenced in a cited third-party source. A domain-level comparison tells you whether your content infrastructure is being reached, not just your homepage.
- Identify the gap prompts. Where does your brand appear in fast mode but not in reasoning mode? Those are the pages or content types that are well-indexed but lack primary-source depth. Where do you appear in reasoning mode but not fast mode? Those are pages that have strong attribution but may need better SEO to surface in quick retrievals.
- Run the comparison-stage prompts specifically at high-reasoning. These produce 24 sub-queries on average. A competitor that appears at the comparison stage in reasoning mode is reaching your buyer at their most-deliberate, most-likely-to-convert moment. If you are absent, that is the highest-priority gap to address.
- Track once a month, not continuously. The citation picture shifts as content is indexed, de-indexed, and updated. Monthly snapshots give you enough signal to measure direction without drowning in noise. Use a consistent prompt set month-over-month so changes are attributable to content changes, not query drift.
What to do with the gap: reasoning-mode gaps are almost always a content depth problem, not a technical SEO problem. The pages exist and are indexed. The issue is that extended-reasoning scrutiny finds them thin: assertions without attribution, claims without methodology, statistics without a named source. The fix is content-side, not tag-side. Analytics and insights work we do with clients in this space starts with exactly this audit to prioritise which pages to deepen first.
One operational note: ChatGPT's reasoning mode behaviour may change as OpenAI updates the model. The Semrush study used a specific model snapshot (early July 2026). The specific percentages will shift as models are updated. The structural pattern, that extended reasoning favours primary sources and penalises aggregated user-generated content, is more likely to persist as a directional finding than the exact numbers.
For teams in Singapore, Malaysia, and Australia, the practical constraint is that ChatGPT Pro or Plus subscription costs vary by market and the Thinking mode may roll out at different rates across regions. We track that in our broader visibility optimisation service work, where client access to the right testing infrastructure is part of the engagement scoping.
