Only 25% of cited sources overlap between ChatGPT's different reasoning modes [Study]

Most AI visibility strategies dainty ChatGPT arsenic a azygous system. The information shows that it mightiness not beryllium wise.

When ChatGPT operates successful high-reasoning mode, it cites a different group of brands, surfaces different root types, and behaves otherwise than erstwhile it’s successful minimal reasoning mode. 

Kevin Indig calls this spread betwixt what shows successful 1 exemplary versus different “reasoning lift.” To analyse it, we collaborated pinch Kevin and analyzed information from the Semrush AI Visibility Toolkit.

Here's what we found:

Key takeaways

• ChatGPT pinch higher reasoning is fundamentally a different hunt engine. Only 25.6% of cited domains overlap betwixt minimal and precocious reasoning for the aforesaid prompts. Nearly 3 successful 4 cited sources are different.

• Citation behaviour changes dramatically pinch higher reasoning on. When comparing debased reasoning to precocious reasoning, the citation complaint jumps from 50% to 68%, the sources per consequence astir double (2.6 → 4.5), and the high-reasoning exemplary fires 4.6x much soul sub-queries.

• Source types displacement erstwhile reasoning turns on. Reddit and different user-generated contented (UGC) sites suffer astir half their stock of citations successful Thinking mode compared to Instant mode, while government, academic, and charismatic archiving sites summation ground.

• Under precocious reasoning, the aforesaid marque often stays successful the speech from a buyer's first mobility to their last. This happened successful 4 of the 20 journeys we tested. Under minimal reasoning, full-funnel persistence was rare.

• Switching from minimal to precocious reasoning affects immoderate industries acold much than others. Citation rates for Finance contented jump by 28 percent points. Consumer Tech hardly changes.

• Top-of-funnel contented has existent worth nether precocious reasoning. Brands cited successful a user's early investigation questions thin to support appearing successful their later, much circumstantial queries from the aforesaid speech — but only pinch a high-reasoning mode. 

• Switching from minimal to precocious reasoning affects immoderate industries acold much than others. Citation rates for Finance contented jump by 28 percent points. Consumer Tech hardly changes.

Methodology

We collaborated pinch Kevin Indig from Growth Memo to analyse information from the Semrush AI Visibility Toolkit.

We ran 100 prompts done GPT-5.2 twice: erstwhile pinch minimal reasoning, and erstwhile pinch precocious reasoning. So, we sewage 200 full responses. 

In ChatGPT’s interface, minimal reasoning corresponds to Instant mode (the default fast-response experience), and precocious reasoning corresponds to Thinking mode (the deeper, multi-step investigation mode). 

Instant is the default experience, while Thinking mode is designed for much complex, multi-step tasks.

The 100 prompts we analyzed screen 20 purchaser journeys crossed 4 categories:

• B2B SaaS
• Finance
• Consumer Tech
• Health and Lifestyle

Each buying travel breaks into 5 stages:

• Problem: Recognizing a request aliases symptom point
• Exploration: Researching what options exist
• Comparison: Evaluating alternatives broadside by side
• Validation: Confirming the starring choice
• Selection: Committing to a circumstantial marque aliases product

For each response, we tracked:

• Citation rate: The stock of responses that mention astatine slightest 1 outer source
• Average citations: The number of sources per cited response
• Fan-out queries: The number of sub-queries the exemplary runs to investigation a punctual earlier answering

Let's research the findings.

1. High reasoning cites sources and uses web searches overmuch more

When you move precocious reasoning on, ChatGPT relies much heavy connected progressive research:

• Citation rate: This climbs from 50% successful Instant mode to 68% successful Thinking mode (+18 percent points)

• Average citations: The number of citations per consequence astir doubles from Instant mode to Thinking mode (2.6 to 4.5)
• Fan-out queries: The number of sub-queries tally is 4.6x higher successful reasoning mode than successful Instant mode

High reasoning besides pulled from 173 unsocial domains crossed the trial group vs. 127 for minimal reasoning. And 99 of those domains that show utilizing the high-reasoning mode ne'er look nether minimal reasoning astatine all.

At the aforesaid time, high-reasoning mode gives only somewhat longer responses. This intends that the summation successful citations isn't simply a byproduct of generating much text. Instead, the exemplary is doing substantially much investigation down the scenes and packing much grounds into astir the aforesaid magnitude of output.

This matters moreover for free-tier users, because ChatGPT routes analyzable prompts (comparisons, evaluations, regulatory questions, and different multi-step decisions) into high-reasoning mode automatically. 

For brands, the accusation is direct: erstwhile your assemblage asks 1 of those analyzable questions, you’re not competing for a azygous placement successful 1 response. You’re competing for visibility crossed each sub-search the exemplary runs on the measurement to that answer.

2. Each reasoning mode cites different domains 

For the aforesaid prompt, only 25.6% of cited domains are shared betwixt minimal- and high-reasoning modes. Almost 3 successful 4 cited sources are different.

The wide root operation besides shifts:

• Reddit appearances driblet from 15% pinch debased reasoning to 7% pinch precocious reasoning
• UGC and reappraisal sites shrink from 14.3% pinch debased reasoning to 6% pinch precocious reasoning
• Government and world sources quadruple from 1.9% pinch debased reasoning to 8.8% pinch precocious reasoning
• Official archiving and support pages turn from 12.4% pinch debased reasoning to 17.5% pinch precocious reasoning
• Brands appear almost arsenic (62.4% pinch debased reasoning v.s 60.6% pinch precocious reasoning)

"The marque that wins nether minimal reasoning is not the marque that wins nether precocious reasoning. The operation of root types is different. The stages wherever citations look are different. These are 2 different systems."

— Kevin Indig, Growth Advisor

Here’s the applicable implication: If astir of your AI citations presently travel from Reddit threads, Quora, aliases UGC reappraisal sites, you're winning via Instant mode but mightiness beryllium losing via Thinking mode. 

To equilibrium capacity successful some modes, attraction your contented finance connected the root types precocious reasoning really pulls from. 

That intends owning much charismatic archiving and reference pages connected your ain site, publishing original investigation that gives writers and academics thing to cite, and getting your marque referenced successful .gov, .edu, and trade-association resources done partnerships, master contributions, and information sharing.

3. The biggest mode spread shows up early successful the purchaser journey

The citation complaint spread betwixt minimal and precocious reasoning isn’t constant. It depends connected wherever the personification sits successful the purchaser journey, and what benignant of mobility they’re asking astatine that point.

To illustrate, a purchaser evaluating CRM package mightiness advancement done the 5 stages utilizing these questions:

• Problem: "How do I cognize if my income squad needs a CRM?"
• Exploration: "What types of CRM package beryllium for B2B SaaS?"
• Comparison: "HubSpot vs. Salesforce vs. Pipedrive for a 50-person income team"
• Validation: "Is HubSpot worthy the value for mid-market B2B companies?"
• Selection: "How do I get started pinch HubSpot Sales Hub?"

Across each 20 journeys, 3 patterns stood out:

• Early successful the journey, the 2 modes hardly overlap. At the Problem stage, the citation complaint successful precocious reasoning mode is 35 percent points higher than successful minimal reasoning. By the Validation stage, the spread shrinks to 5 points. Minimal-reasoning mode often answers early-funnel questions without citing outer sources, while high-reasoning mode is much apt to investigation and mention them.

• The Comparison shape is wherever high-reasoning mode does the astir research. It fires 24 sub-queries per Comparison prompt, compared to 5.5 for minimal reasoning. Average citations per consequence highest present too: 9.8 pinch precocious reasoning vs. 5.8 pinch minimal reasoning.

• At the Selection stage, precocious reasoning still pulls much sources than minimal reasoning. Each high-reasoning consequence cites 4.7 sources connected average, vs. 2.6 for minimal reasoning. Both modes mention the web heavy here; precocious reasoning conscionable goes deeper.

Across the 100 prompts we tested, minimal reasoning ran 245 web searches successful total. High reasoning ran 1,130 web searches, almost 5x more. Most of that other investigation happens astatine the Comparison and Selection stages, erstwhile the personification is choosing betwixt circumstantial products.

Fan-out queries travel the aforesaid style and are substantially higher nether precocious reasoning astatine each stage. They spike astatine Comparison (24 sub-queries per consequence vs. 5.5 for minimal reasoning) and again astatine Selection (15.4 vs. 2.6), which are the stages wherever the exemplary is actively moving done circumstantial merchandise options.

When high-reasoning mode gets a punctual for illustration "Salesforce vs. HubSpot vs. Pipedrive for a 50-person income team," it doesn't conscionable hunt for that circumstantial prompt. It breaks the mobility into astir 8 sub-queries (things related to pricing tiers, API integrations, information compliance, and developer documentation) and runs a abstracted hunt for each one. 

The marque that wins the reply isn't needfully the 1 that ranks for the original prompt. It's the 1 that has pages showing up intelligibly crossed galore of those sub-searches.

What this intends is you shouldn’t disregard top-of-funnel contented arsenic conscionable marque awareness. Most users inquire a operation of casual and analyzable prompts, and the analyzable ones trigger high-reasoning mode automatically. 

Treat your early-funnel contented pieces arsenic citation sources. Name your product, methodology, aliases model explicitly, truthful the AI has thing to property erstwhile it surfaces those pages.

4. Under high-reasoning mode, brands persist crossed the journey

LLM sessions are conversations alternatively than azygous queries. So a cardinal mobility is: Does a marque cited astatine the commencement of a travel transportation done to the end?

Under precocious reasoning, yes. Under minimal reasoning, no.

We measured marque persistence by checking whether a marque cited astatine the Problem shape survived to the Selection shape of the aforesaid journey:

• Minimal reasoning: No journeys show this benignant of full-funnel persistence
• High reasoning: Brand continuity is maintained successful 4 of the 20 journeys

High reasoning besides returns to the aforesaid root much than erstwhile wrong a azygous answer. In 51 of 100 high-reasoning responses, the aforesaid domain appears aggregate times successful the aforesaid consequence (vs. 26 of 100 for minimal). 

This is simply a different effect than travel persistence: anchoring is astir extent (how heavy the exemplary leans connected 1 root wrong a azygous answer), while persistence is astir continuity (whether the aforesaid marque keeps appearing crossed a multi-step conversation).

"Top-of-funnel contented isn't conscionable marque consciousness for AI visibility. Under high-reasoning mode, it's a starring parameter of wherever the exemplary lands astatine determination time."

— Kevin Indig, Growth Advisor

To guarantee marque continuity, audit your AI visibility crossed afloat purchaser journeys and intent categories. In the AI Visibility Toolkit, unfastened the Questions study and research the cardinal topics your customers inquire AI tools, categorized by intent and chimney stage.

Then, analyse the circumstantial questions group inquire crossed each shape and topic.

Finally, caput to the Narrative Drivers study to spot really your marque appears successful cardinal conversations crossed the chimney compared to your competitors.

If you show up for decision-stage prompts (Comparison, Validation, Selection) but not for early-stage ones (Problem, Exploration), that's a spread worthy closing. 

With high-reasoning mode, brands cited early successful a travel often proceed to beryllium cited later, truthful investing successful Problem-stage contented tin compound your existing Selection-stage visibility.

5. Reasoning assistance varies sharply by category

Not each categories we analyzed use from accrued citation rates arsenic erstwhile the high-reasoning mode turns on. It varies by industry:

• Finance: A 28 percent constituent summation successful citation complaint from debased reasoning to precocious reasoning
• Health and Lifestyle: A 24 percent constituent summation successful citation complaint from debased reasoning to precocious reasoning
• B2B SaaS: A 16 percent constituent summation from debased reasoning to precocious reasoning
• Consumer Tech: A 4 percent constituent summation from debased reasoning to precocious reasoning

Consumer Tech stands out. 

Even though precocious reasoning runs much sub-queries per Consumer Tech punctual (13.4) than immoderate different class we tested, it ends up citing galore of the aforesaid brands and sources arsenic minimal reasoning. 

In different words, the other investigation hardly changes the Consumer Tech answer, which suggests ChatGPT already has beardown soul knowledge of communal Consumer Tech topics from its training information and doesn’t request caller investigation to onshore connected the aforesaid brands.

For Finance and Health brands, optimizing for precocious reasoning intends producing the contented the exemplary actively pulls into its sub-searches.

In practice, that intends publishing charismatic merchandise documentation, achromatic papers backed by your ain data, and system contented (clear claims per section, named entities, definitive stats) the exemplary tin propulsion cleanly into a azygous sub-query response.

How to set your AI visibility strategy for each reasoning mode

The findings propose minimal-reasoning and high-reasoning behaviour shouldn’t beryllium treated arsenic a azygous visibility surface. They propulsion from different sources, favour different contented types, and tin nutrient very different winners for the aforesaid brand. 

The extremity is not to prime 1 mode and optimize for it. It’s to make judge you’re visible successful both.

Here’s how:

• Split your search by reasoning mode. Use a instrumentality for illustration Prompt Tracking to group the prompts you already show into 2 buckets: analyzable queries (multi-criteria evaluation, side-by-side comparisons, regulatory aliases compliance questions) and elemental queries (definitions, single-factor lookups, basal “what is X” questions). Track citation rate, mention rate, and the apical cited domains for each bucket separately. Where the 2 buckets diverge astir is wherever reasoning assistance is reshaping who wins.

• Build a two-track contented strategy. For minimal-reasoning visibility, put successful comparison-stage content, Reddit, and review-site presence, and clear product-focused pages connected your ain site. For high-reasoning visibility, put successful early-funnel education, charismatic merchandise documentation, achromatic papers, and charismatic reference worldly that lives astatine a citable URL.

• Map and audit your privilege purchaser journeys by stage. For each privilege journey, constitute down the mobility a purchaser would inquire astatine each of the 5 stages (Problem, Exploration, Comparison, Validation, Selection). Then tally those questions done ChatGPT pinch Thinking mode connected and statement wherever your marque appears and wherever it drops out. Stages wherever you’re missing are your highest-leverage contented gaps.

Understanding these differences starts pinch measuring AI visibility astatine the punctual and travel level. 

The Semrush AI Visibility Toolkit shows you which prompts and intent categories thrust your brand’s visibility successful AI answers, which sources power those answers, and really your beingness shifts crossed the purchaser journey.

 Even without a built-in reasoning-mode filter, that information is what tells you wherever reasoning assistance is astir apt to beryllium successful play and wherever to put successful closing the gap.