Most AI visibility methods deal with ChatGPT as a single system. The information reveals that it won’t be clever.
When ChatGPT operates in high-reasoning mode, it cites a special set of manufacturers, surfaces completely different supply varieties, and behaves otherwise than when it’s in minimal reasoning mode.
Kevin Indig calls this hole between what reveals in a single mannequin versus one other “reasoning elevate.” To analyze it, we partnered with Kevin and analyzed knowledge from the Semrush AI Visibility Toolkit.
Here is what we discovered:
Key takeaways
- ChatGPT with greater reasoning is actually a special search engine. Solely 25.6% of cited domains overlap between minimal and excessive reasoning for a similar prompts. Practically three in 4 cited sources are completely different.
- Quotation habits modifications dramatically with greater reasoning on. When evaluating low reasoning to excessive reasoning, the quotation price jumps from 50% to 68%, the sources per response almost double (2.6 → 4.5), and the high-reasoning mannequin fires 4.6x extra inner sub-queries.
- Supply varieties shift when reasoning activates. Reddit and different user-generated content material (UGC) websites lose roughly half their share of citations in Considering mode in comparison with On the spot mode, whereas authorities, educational, and official documentation websites acquire floor.
- Beneath excessive reasoning, the identical model typically stays within the dialog from a purchaser’s first query to their final. This occurred in Four of the 20 journeys we examined. Beneath minimal reasoning, full-funnel persistence was uncommon.
- Switching from minimal to excessive reasoning impacts some industries way over others. Quotation charges for Finance content material bounce by 28 share factors. Client Tech barely modifications.
- Prime-of-funnel content material has actual worth underneath excessive reasoning. Manufacturers cited in a consumer’s early analysis questions are likely to hold showing of their later, extra particular queries from the identical dialog — however solely with a high-reasoning mode.
- Switching from minimal to excessive reasoning impacts some industries way over others. Quotation charges for Finance content material bounce by 28 share factors. Client Tech barely modifications.
Methodology
We partnered with Kevin Indig from Growth Memo to research knowledge from the Semrush AI Visibility Toolkit.
We ran 100 prompts by means of GPT-5.2 twice: as soon as with minimal reasoning, and as soon as with excessive reasoning. So, we received 200 complete responses.
In ChatGPT’s interface, minimal reasoning corresponds to On the spot mode (the default fast-response expertise), and excessive reasoning corresponds to Considering mode (the deeper, multi-step analysis mode).
On the spot is the default expertise, whereas Considering mode is designed for extra advanced, multi-step duties.
The 100 prompts we analyzed cowl 20 purchaser journeys throughout 4 classes:
- B2B SaaS
- Finance
- Client Tech
- Well being and Way of life
Every shopping for journey breaks into 5 phases:
- Downside: Recognizing a necessity or ache level
- Exploration: Researching what choices exist
- Comparability: Evaluating options aspect by aspect
- Validation: Confirming the main alternative
- Choice: Committing to a selected model or product
For every response, we tracked:
- Quotation price: The share of responses that cite at the very least one exterior supply
- Common citations: The variety of sources per cited response
- Fan-out queries: The variety of sub-queries the mannequin runs to analysis a immediate earlier than answering
Let’s discover the findings.
1. Excessive reasoning cites sources and makes use of internet searches way more
Whenever you flip excessive reasoning on, ChatGPT depends extra closely on energetic analysis:
- Quotation price: This climbs from 50% in On the spot mode to 68% in Considering mode (+18 share factors)
- Common citations: The variety of citations per response almost doubles from On the spot mode to Considering mode (2.6 to 4.5)
- Fan-out queries: The variety of sub-queries run is 4.6x greater in considering mode than in On the spot mode
Excessive reasoning additionally pulled from 173 distinctive domains throughout the take a look at set vs. 127 for minimal reasoning. And 99 of these domains that present utilizing the high-reasoning mode by no means seem underneath minimal reasoning in any respect.
On the identical time, high-reasoning mode provides solely barely longer responses. Which means the rise in citations is not merely a byproduct of producing extra textual content. As an alternative, the mannequin is doing considerably extra analysis behind the scenes and packing extra proof into roughly the identical size of output.
This issues even for free-tier customers, as a result of ChatGPT routes advanced prompts (comparisons, evaluations, regulatory questions, and different multi-step selections) into high-reasoning mode robotically.
For manufacturers, the implication is direct: when your viewers asks a kind of advanced questions, you’re not competing for a single placement in a single response. You’re competing for visibility throughout each sub-search the mannequin runs alongside the best way to that reply.
2. Every reasoning mode cites completely different domains
For a similar immediate, solely 25.6% of cited domains are shared between minimal- and high-reasoning modes. Nearly three in 4 cited sources are completely different.
The general supply combine additionally shifts:
- Reddit appearances drop from 15% with low reasoning to 7% with excessive reasoning
- UGC and evaluate websites shrink from 14.3% with low reasoning to six% with excessive reasoning
- Authorities and educational sources quadruple from 1.9% with low reasoning to eight.8% with excessive reasoning
- Official documentation and assist pages develop from 12.4% with low reasoning to 17.5% with excessive reasoning
- Manufacturers seem nearly equally (62.4% with low reasoning v.s 60.6% with excessive reasoning)
“The model that wins underneath minimal reasoning just isn’t the model that wins underneath excessive reasoning. The combo of supply varieties is completely different. The phases the place citations seem are completely different. These are two completely different methods.”
— Kevin Indig, Development Advisor
Right here’s the sensible implication: If most of your AI citations at the moment come from Reddit threads, Quora, or UGC evaluate websites, you are profitable through On the spot mode however is likely to be dropping through Considering mode.
To stability efficiency in each modes, focus your content material funding on the supply varieties excessive reasoning truly pulls from.
Which means proudly owning extra official documentation and reference pages by yourself website, publishing authentic analysis that provides writers and teachers one thing to quote, and getting your model referenced in .gov, .edu, and trade-association sources by means of partnerships, professional contributions, and knowledge sharing.
3. The largest mode hole reveals up early within the purchaser journey
The quotation price hole between minimal and excessive reasoning isn’t fixed. It relies on the place the consumer sits within the purchaser journey, and what sort of query they’re asking at that time.
As an instance, a purchaser evaluating CRM software program may progress by means of the 5 phases utilizing these questions:
- Downside: “How do I do know if my gross sales crew wants a CRM?”
- Exploration: “What forms of CRM software program exist for B2B SaaS?”
- Comparability: “HubSpot vs. Salesforce vs. Pipedrive for a 50-person gross sales crew”
- Validation: “Is HubSpot definitely worth the worth for mid-market B2B firms?”
- Choice: “How do I get began with HubSpot Gross sales Hub?”
Throughout all 20 journeys, three patterns stood out:
- Early within the journey, the 2 modes barely overlap. On the Downside stage, the quotation price in excessive reasoning mode is 35 share factors greater than in minimal reasoning. By the Validation stage, the hole shrinks to five factors. Minimal-reasoning mode typically solutions early-funnel questions with out citing exterior sources, whereas high-reasoning mode is extra prone to analysis and cite them.
- The Comparability stage is the place high-reasoning mode does probably the most analysis. It fires 24 sub-queries per Comparability immediate, in comparison with 5.5 for minimal reasoning. Common citations per response peak right here too: 9.Eight with excessive reasoning vs. 5.Eight with minimal reasoning.
- On the Choice stage, excessive reasoning nonetheless pulls extra sources than minimal reasoning. Every high-reasoning response cites 4.7 sources on common, vs. 2.6 for minimal reasoning. Each modes cite the net closely right here; excessive reasoning simply goes deeper.
Throughout the 100 prompts we examined, minimal reasoning ran 245 internet searches in complete. Excessive reasoning ran 1,130 internet searches, nearly 5x extra. Most of that additional analysis occurs on the Comparability and Choice phases, when the consumer is selecting between particular merchandise.
Fan-out queries observe the identical form and are considerably greater underneath excessive reasoning at each stage. They spike at Comparability (24 sub-queries per response vs. 5.5 for minimal reasoning) and once more at Choice (15.Four vs. 2.6), that are the phases the place the mannequin is actively working by means of particular product choices.
When high-reasoning mode will get a immediate like “Salesforce vs. HubSpot vs. Pipedrive for a 50-person gross sales crew,” it does not simply seek for that particular immediate. It breaks the query into roughly Eight sub-queries (issues associated to pricing tiers, API integrations, safety compliance, and developer documentation) and runs a separate seek for every one.
The model that wins the reply is not essentially the one which ranks for the unique immediate. It is the one which has pages exhibiting up clearly throughout lots of these sub-searches.
What this implies is you shouldn’t dismiss top-of-funnel content material as simply model consciousness. Most customers ask a mixture of informal and sophisticated prompts, and the advanced ones set off high-reasoning mode robotically.
Deal with your early-funnel content material items as quotation sources. Identify your product, methodology, or framework explicitly, so the AI has one thing to attribute when it surfaces these pages.
4. Beneath high-reasoning mode, manufacturers persist throughout the journey
LLM classes are conversations fairly than single queries. So a key query is: Does a model cited firstly of a journey carry by means of to the top?
Beneath excessive reasoning, sure. Beneath minimal reasoning, no.
We measured model persistence by checking whether or not a model cited on the Downside stage survived to the Choice stage of the identical journey:
- Minimal reasoning: No journeys present this type of full-funnel persistence
- Excessive reasoning: Model continuity is maintained in 4 of the 20 journeys
Excessive reasoning additionally returns to the identical supply greater than as soon as inside a single reply. In 51 of 100 high-reasoning responses, the identical area seems a number of occasions in the identical response (vs. 26 of 100 for minimal).
This can be a completely different impact than journey persistence: anchoring is about depth (how closely the mannequin leans on one supply inside a single reply), whereas persistence is about continuity (whether or not the identical model retains showing throughout a multi-step dialog).
“Prime-of-funnel content material is not simply model consciousness for AI visibility. Beneath high-reasoning mode, it is a main indicator of the place the mannequin lands at determination time.”
— Kevin Indig, Development Advisor
To make sure model continuity, audit your AI visibility throughout full purchaser journeys and intent classes. Within the AI Visibility Toolkit, open the Questions report and discover the important thing matters your clients ask AI instruments, categorized by intent and funnel stage.
Then, analyze the precise questions individuals ask throughout every stage and matter.
Lastly, head to the Narrative Drivers report back to see how your model seems in key conversations throughout the funnel in comparison with your rivals.
In the event you present up for decision-stage prompts (Comparability, Validation, Choice) however not for early-stage ones (Downside, Exploration), that is a spot price closing.
With high-reasoning mode, manufacturers cited early in a journey typically proceed to be cited later, so investing in Downside-stage content material can compound your present Choice-stage visibility.
5. Reasoning elevate varies sharply by class
Not all classes we analyzed profit from elevated quotation charges equally when the high-reasoning mode activates. It varies by business:
- Finance: A 28 share level enhance in quotation price from low reasoning to excessive reasoning
- Well being and Way of life: A 24 share level enhance in quotation price from low reasoning to excessive reasoning
- B2B SaaS: A 16 share level enhance from low reasoning to excessive reasoning
- Client Tech: A Four share level enhance from low reasoning to excessive reasoning
Client Tech stands out.
Regardless that excessive reasoning runs extra sub-queries per Client Tech immediate (13.4) than some other class we examined, it finally ends up citing most of the identical manufacturers and sources as minimal reasoning.
In different phrases, the additional analysis barely modifications the Client Tech reply, which suggests ChatGPT already has sturdy inner information of frequent Client Tech matters from its coaching knowledge and doesn’t want recent analysis to land on the identical manufacturers.
For Finance and Well being manufacturers, optimizing for prime reasoning means producing the content material the mannequin actively pulls into its sub-searches.
In apply, which means publishing official product documentation, white papers backed by your personal knowledge, and structured content material (clear claims per part, named entities, express stats) the mannequin can pull cleanly right into a single sub-query response.
How one can modify your AI visibility technique for every reasoning mode
The findings counsel minimal-reasoning and high-reasoning habits shouldn’t be handled as a single visibility floor. They pull from completely different sources, favor completely different content material varieties, and may produce very completely different winners for a similar model.
The aim is to not decide one mode and optimize for it. It’s to ensure you’re seen in each.
Right here’s how:
- Cut up your monitoring by reasoning mode. Use a software like Prompt Tracking to group the prompts you already monitor into two buckets: advanced queries (multi-criteria analysis, side-by-side comparisons, regulatory or compliance questions) and easy queries (definitions, single-factor lookups, primary “what’s X” questions). Observe quotation price, point out price, and the highest cited domains for every bucket individually. The place the 2 buckets diverge most is the place reasoning elevate is reshaping who wins.
- Construct a two-track content material technique. For minimal-reasoning visibility, spend money on comparison-stage content material, Reddit, and review-site presence, and clear product-focused pages by yourself website. For top-reasoning visibility, spend money on early-funnel schooling, official product documentation, white papers, and authoritative reference materials that lives at a citable URL.
- Map and audit your precedence purchaser journeys by stage. For every precedence journey, write down the query a purchaser would ask at every of the 5 phases (Downside, Exploration, Comparability, Validation, Choice). Then run these questions by means of ChatGPT with Considering mode on and word the place your model seems and the place it drops out. Levels the place you’re lacking are your highest-leverage content material gaps.
Understanding these variations begins with measuring AI visibility on the immediate and journey stage.
The Semrush AI Visibility Toolkit reveals you which ones prompts and intent classes drive your model’s visibility in AI solutions, which sources affect these solutions, and the way your presence shifts throughout the client journey.
Even and not using a built-in reasoning-mode filter, that knowledge is what tells you the place reasoning elevate is most certainly to be in play and the place to spend money on closing the hole.
