Thinking, fast and cited: what ChatGPT's reasoning mode means for AI citations and GEO
· GEO Fundamentals
A Nobel Prize in psychology, a pub quiz and a Semrush study walk into a search results page — and between them, they explain why your brand can be cited by ChatGPT one minute and invisible the next.
Daniel Kahneman spent a career proving that human beings think in two speeds. System 1 is fast, instinctive and frequently wrong. System 2 is slow, effortful and usually right, if you can be bothered to summon it.
You might rightly be wondering what a Nobel-winning psychologist has to do with AI citations. Rather a lot, as it turns out — because the large language models now doing a fair chunk of the world's research have started exhibiting the same split personality, and the consequences for anyone practising generative engine optimization are significant.
Two minds, one machine
Generative engine optimization — GEO to its friends, though the acronym still causes the odd mix-up with geo-targeting — is the discipline of getting your brand mentioned, recommended or cited inside AI-generated answers, rather than merely ranked in a list of ten blue links. It is the successor to classic SEO in a world where a growing share of searches never produce a click at all, only a paragraph with your competitor's name in it (or, with a bit of luck and some careful content structuring, yours).
Until recently, most GEO advice treated an AI platform as a single, consistent entity: optimise your content, earn a citation, done. New research suggests that's no longer good enough — because ChatGPT itself now operates in two distinct modes, and it cites almost entirely different sources depending on which one is switched on.
The 25.6% problem
A Semrush analysis conducted with SEO consultant Kevin Indig, reported by Search Engine Land on 1 July 2026, ran 100 prompts across 20 realistic buyer journeys — four categories, five funnel stages each — through ChatGPT twice: once in quick-answer mode, once in high-reasoning ("Thinking") mode. The overlap in cited domains between the two runs was just 25.6%.
Put plainly: ask the fast-thinking version of ChatGPT a question and ask the slow-thinking version the same question, and roughly three-quarters of its sources change. A brand that shows up beautifully in one mode can be entirely absent from the other.
The gap isn't cosmetic. When high-reasoning mode is switched on:
- Sources per answer nearly double, from 2.6 to 4.5.
- Web searches per query rise from around 245 to roughly 1,130.
- Sub-queries fired for a single comparison question — evaluating CRM platforms, say — climb from 5.5 to as many as 24.
- Citation rate overall jumps from 50% to 68%.
- Reddit's share of citations more than halves, from 15% to 7%, as the model reallocates its trust toward government, academic and official documentation sources.
That last figure is the one worth sitting with. Quick-answer mode leans on forum chatter and community consensus — cheap to retrieve, broadly representative, good enough for a fast answer. High-reasoning mode has the time and the query budget to go looking for primary sources instead, and it visibly prefers them.
Why the model changes its mind
None of this is ChatGPT being fickle for the sake of it. Reasoning mode works by decomposing a question into many more sub-queries and running many more retrieval passes before it starts writing — the numbers above (245 searches becoming 1,130) are the mechanism, not a side effect. More searches surface a wider, deeper pool of candidate sources, and a model given more "thinking time" appears to weight authority and specificity more heavily than it does when working from a thin, fast retrieval pass.
It's a reasonable parallel to Kahneman's own finding: System 1 reaches for the answer that's cognitively cheap and available (Sydney; the top Reddit thread). System 2, given the chance, goes and checks.
What this means for your GEO strategy
The practical upshot for anyone doing generative engine optimization work is uncomfortable but useful: being cited once is worth precisely nothing on its own. A single citation-tracking snapshot, taken in whichever mode your tool happens to query by default, tells you about one half of ChatGPT's split personality and nothing about the other.
A few adjustments follow directly from the data:
- Track AI citations across both modes, not one. If your GEO or LLM-citation-tracking tool only samples quick-answer mode (the cheaper, faster, more common default), you are measuring a shrinking half of the picture as reasoning-mode usage grows.
- Build content that survives a deep-research pass, not just a skim. Reasoning mode rewards primary data, named sources, specific figures and documentation-style clarity — the same qualities that make content genuinely citable rather than merely search-engine-friendly.
- Don't over-index on UGC-style visibility. A strong Reddit presence or forum footprint helps in quick-answer mode, where community content is cited more freely. It buys you noticeably less in high-reasoning mode, where the model actively rotates away from it.
- Publish the standalone facts, not just the narrative. AI systems extract and reproduce specific sentences and data points far more readily than they summarise flowing prose — write the sentence you'd want quoted verbatim, and put it near the top.
- Treat citation frequency as a moving target. Share of voice inside AI answers should be measured per reasoning mode, per platform and over time — not as a single number to be checked once a quarter.
Frequently asked questions
What is generative engine optimization (GEO)? Generative engine optimization is the practice of structuring content and brand presence so that AI systems such as ChatGPT, Google AI Overviews, Perplexity and Claude cite, recommend or mention it in generated answers, rather than optimising purely for a ranked list of links.
Why does ChatGPT cite different sources in reasoning mode? High-reasoning mode runs substantially more web searches and sub-queries before answering, surfacing a wider pool of sources and shifting the model's preference toward primary, authoritative and documentation-style content over forum and community content.
How can brands track AI citations properly? By monitoring citation frequency and source overlap across both quick-answer and high-reasoning modes, across multiple AI platforms, since Search Engine Land's data shows the overlap between modes on the same platform can be as low as 25.6%.
The kicker
Kahneman would have had a field day with this one. It turns out your brand doesn't need one convincing story to be believed by a machine — it needs two, because the same model, asked to think just a little harder, goes looking for an entirely different set of receipts!
Sources: ChatGPT Thinking mode changes which brands get cited — Search Engine Land, 1 July 2026; Semrush analysis with Kevin Indig.
