One Prompt, Two Different Search Worlds
We used to optimize for a single search engine index. Now, we're optimizing for different modes of the same engine. When Semrush teamed up with Kevin Indig to analyze how ChatGPT handles searches in their original reasoning visibility study, they uncovered a massive divide. When you run the same 100 prompts through ChatGPT's instant (minimal) and thinking (high) reasoning modes, only 25.6% of the cited domains overlap.
Let that sink in. Nearly three-quarters of the websites that win visibility when the AI thinks deeply are completely invisible when it responds instantly. We don't just fight for page one anymore. We're fighting for two different states of AI contemplation.
Back in my agency days, we tracked rankings like they were absolute truths. Today? The truth depends on how long the model has to think before it answers. Kevin Indig put it best: "The brand that wins under minimal reasoning is not the brand that wins under high reasoning." He's spot on. If you only optimize for the quick answer, you might lose the deep research journey entirely.
The Search Methodology Explained
This isn't another small-scale AI experiment. Semrush and Indig ran a rigorous test. They created 100 distinct prompts mapping onto 20 complete buyer journeys. These journeys spanned four major verticals: B2B SaaS, Finance, Consumer Tech, and Health & Lifestyle.
They ran the entire list twice. The first run used Instant mode, ChatGPT’s default response path. The second run forced the engine into Thinking mode, triggering its deep reasoning framework. By analyzing the citations across both runs, the team mapped exactly how the AI’s source selection alters when it has time to work through a problem. It gives us a clear look at how search visibility splits down the middle. What we're seeing is a fragmentation of organic search that traditional SEO metrics aren't built to track. We finally have real data on how search is turning into a customer acquisition channel, but measuring it requires throwing out the old playbook.
Inside the AI Reasoning Loop
What happens when ChatGPT enters Thinking mode? It doesn't just sit there and ponder. It runs queries. Behind the scenes, the AI goes to work, breaking down your prompt into sub-components.
The study tracked this and found a massive lift. In Instant mode, ChatGPT made just 245 internal sub-queries across the tests. When switched to Thinking mode, that number surged to 1,130. That's a 4.6x increase.
This depth shows up in the output. The citation rate jumped from 50% in Instant mode to 68% in Thinking mode. That's an 18-percentage-point increase. The average number of citations per response climbed from 2.6 to 4.5. It does more research, links to more sources, but keeps the answer length steady. It's higher-density reference work. In a world where publishing more thin content actually hurts your visibility, the AI's preference for dense, cited details is a warning sign. The era of high-volume, low-effort pages is over.
The Rise of Authority Sources
For a long time, Reddit dominated AI answers. The instant response mode still loves forums. But when ChatGPT is allowed to think, UGC takes a hit.
Reddit’s share of citations dropped from 15% under minimal reasoning to just 7% under high reasoning. Other user-generated content and product review sites saw a similar decline, falling from 14.3% down to 6%.
So, where does that attention go? The AI pivots to official and institutional sources. Government and academic domains surged from a tiny 1.9% share to 8.8%. Official documentation and support websites jumped from 12.4% to 17.5%.
This tells us something important. When the AI has time to check its work, it rejects quick opinions. It swaps out forum threads for reference material. If your SEO strategy depends on people talking about you on Reddit, your visibility has a weak shelf-life. You need official, structured documentation to survive the deep-thinking passes. This aligns perfectly with what we see in topical authority and knowledge graph strategies—the search engines are looking for verified entity profiles, not just casual mentions.
Citations Across the Funnel
The disparity becomes even clearer when you look at different stages of the buyer journey. Early in the funnel, during the initial "Problem" stage, the AI citation gap is huge. In fact, there is a total 35-percentage-point difference in citation rates, with the high-reasoning mode citing far more heavily.
As buyers move toward evaluation, the research gets intense. During the "Comparison" stage, the sub-query activity peaks. The AI runs an average of 24 sub-queries in Thinking mode, compared to a meager 5.5 sub-queries in Instant mode. This translates to a massive jump in links: the AI provides an average of 9.8 citations under deep reasoning, versus 5.8 in instant replies.
Winning this journey is hard. The study tracked brand persistence across the entire funnel. Out of the 20 buyer journeys tested, a brand managed to maintain visibility across the full funnel only 4 times—and in those cases, it happened under high reasoning. If you aren't visible in the deep research phase, you won't survive the transition to the buying decision.
Industry Winners and Losers
Different industries feel this split differently. Finance saw the biggest citation lift. Its citation rate jumped by 28 percentage points under high reasoning. Health & Lifestyle followed closely with a 24-point increase. These are areas where truth matters, and ChatGPT knows it.
B2B SaaS experienced a 16-point lift. For software buyers, this means product sheets and verified specs become currency.
Then there is Consumer Tech. It had only a 4-point citation rate increase. Yet, it triggered the highest number of sub-queries. The AI does a lot of work comparing consumer technical specs, but it rarely goes to new domains to find the answers. It sticks to a tight group of reviews. If you are in Consumer Tech, the gate is narrow. You must be on the primary sites the engine trusts.
A Playbook for Dual Visibility
You can't optimize for both modes with the same old playbook. You need a split approach. I always tell my mentees to stop chasing the same keywords and start structuring their knowledge to feed the AI reasoning loops.
First, split your measurement. Run tracking diagnostics on your brand across both modes. Know where you stand when the AI answers quick, and where you stand when it thinks.
Second, double down on official documentation. The massive climb in official docs and support resources (from 12.4% to 17.5%) means your knowledge base is your best defense. Keep it clean. Stop writing vague blog posts. Write clear, technical manuals, API guides, and spec sheets.
Third, create primary research assets. Academic and government sources saw their share lift to 8.8%. While you can't easily get a government link, you can publish original data studies, whitepapers, and structured reports that authorities cite.
Fourth, map your content to intent. Early-stage problem searches need deep, cited explainers. Mid-stage comparison queries require clean feature tables that the AI can easily parse in its sub-queries. The tools are changing, but the core rule of search remains: feed the crawler what it needs.