A Semrush study found only 25.6% of the sources ChatGPT cites in high-reasoning mode overlap with its fast mode. Getting cited in the quick answer and the considered answer are now two different jobs.
Getting cited in the fast answer and getting cited in the considered answer are now two different jobs.
Turn on ChatGPT's Thinking mode and you are, for all practical purposes, looking at a different search engine. A Semrush analysis with Kevin Indig found that only about a quarter of the sources cited in high-reasoning mode overlap with the ones cited in fast, instant mode.
The study ran the same prompts through minimal reasoning and high reasoning and compared what came back. Only 25.6% of cited domains overlapped. Nearly three in four sources changed. High reasoning also ran close to five times as many web searches, cited a source in more of its answers (up from 50% to 68%) and pulled in more sources per answer, rising from 2.6 to 4.5.
The mix shifted too. Reddit's share of citations dropped from 15% to 7%. User-generated content and review sites fell from 14.3% to 6%. Government and academic sources jumped from 1.9% to 8.8%. In plain terms, the harder ChatGPT thinks, the more it reaches for authoritative primary sources and the less it leans on forums and reviews.
Why it matters
Every business chasing visibility in AI answers has been optimising for one surface. There are at least two. If your brand only shows up when the model is skimming, you vanish the moment a user asks a harder question and the model switches to reasoning. For considered, high-value purchases, which is exactly where the money sits, the reasoning surface is the one that matters.
Share of cited sources that overlap between ChatGPT's minimal and high reasoning modes
What to do about it
The answer box is not one box, and the brands that work out which one their buyers use will own the citation that counts.