By Ali Ikhwan 
In the rapidly evolving landscape of Generative Engine Optimization (GEO) and artificial intelligence, marketers have long treated AI chatbots as singular entities. However, a groundbreaking joint study by SEO platform Semrush and industry analyst Kevin Indig has shattered this assumption.
The research reveals that ChatGPT’s "high-reasoning" mode—which utilizes advanced, multi-step thinking processes to solve complex queries—behaves like an entirely different search engine compared to its standard, "minimal-reasoning" counterpart.
By analyzing how these two distinct cognitive modes crawl, search, and cite web sources, the study exposes a dramatic divergence in brand visibility. For enterprises and search marketers, the implications are clear: a brand that dominates standard AI search results may completely vanish when a user switches to a high-reasoning model to make a final buying decision.
Main Facts: The Bifurcation of AI Search
The core finding of the Semrush and Kevin Indig analysis is that ChatGPT’s high-reasoning mode does not merely refine standard answers; it executes a completely different retrieval strategy. The study yielded several critical insights:
- Minimal Citation Overlap: Only 25.6% of the domains cited in minimal-reasoning mode overlapped with those cited in high-reasoning mode for the exact same prompts.
- Aggressive Information Gathering: High-reasoning mode ran nearly five times as many web searches (1,130 searches across the test set) compared to minimal-reasoning mode (245 searches).
- Decline of User-Generated Content (UGC): Reddit and review-heavy platforms suffered massive drops in visibility when high-reasoning was activated, as the AI favored authoritative, primary documentation.
- Expanded Citation Pools: The frequency and volume of citations increased significantly under high-reasoning conditions, with the AI citing more unique sources per query.
- Funnel Persistence: Brands that successfully secured citations in the early research phase of a high-reasoning interaction were far more likely to retain their visibility throughout the later stages of the buyer journey.
Chronology: The Evolution of AI Search and the Study’s Genesis
To understand why these findings are so disruptive, it is necessary to trace the technological shift that led to this point.
The Rise of Retrieval-Augmented Generation (RAG)
In the early days of generative AI, models like GPT-3.5 relied solely on pre-trained datasets. This led to "hallucinations" and a lack of real-time information. To solve this, OpenAI and its competitors integrated web-browsing capabilities, utilizing Retrieval-Augmented Generation (RAG). In this setup (now referred to as "minimal reasoning"), the AI takes a user’s prompt, performs a quick, single-step web search, pulls a few top pages, and synthesizes an answer.
The Shift to Agentic Reasoning
In late 2024, OpenAI introduced its "o1" and subsequent reasoning models. These models do not just retrieve and summarize; they "think" before they respond. They break complex problems down into smaller logical steps, self-correct, and run parallel web searches to verify facts.
The Semrush-Indig Investigation
Recognizing that this shift would fundamentally alter digital marketing, Semrush partnered with Kevin Indig to stress-test this new paradigm.
The researchers designed a highly controlled experiment to observe how these algorithmic changes affect brand visibility. They drafted a test suite of 100 prompts representing 20 distinct customer buyer journeys across four key verticals: B2B SaaS, finance, consumer technology, and health and lifestyle.
Each prompt was executed twice under identical conditions: once in ChatGPT’s minimal-reasoning mode and once in its high-reasoning mode. The resulting data, published in early 2025, provides the first empirical look at how deep-thinking AI engines select the brands they recommend to users.
Supporting Data: A Deep Dive into the Numbers
The study’s data points paint a vivid picture of two distinct search algorithms operating under the hood of a single interface.
1. The Search and Citation Explosion
When ChatGPT is given "time to think," its appetite for information grows exponentially. The high-reasoning model conducted 1,130 web searches across the 100-prompt test set, compared to just 245 searches executed by the minimal-reasoning model.
Web Searches Executed across Test Set (100 Prompts)
┌────────────────────────────────────────────────────────┐
│ Minimal Reasoning: 245 searches │
├────────────────────────────────────────────────────────┤
│ High Reasoning: 1,130 searches (4.6x increase) │
└────────────────────────────────────────────────────────┘This aggressive search behavior directly translated into more diverse and frequent citations:
- Citation Rate: The overall citation rate rose from 50% in minimal reasoning to 68% in high reasoning.
- Citations Per Response: When an answer did include citations, the average number of sources cited jumped from 2.6 (minimal) to 4.5 (high).
2. The Marginalization of Reddit and UGC
One of the most striking shifts occurred in the types of sources the AI trusted. Under standard search algorithms (including Google’s recent updates), user-generated content (UGC) from platforms like Reddit and Quora has enjoyed unprecedented visibility. Minimal-reasoning ChatGPT followed this trend.
However, when high-reasoning mode was engaged, the AI actively deprioritized these platforms:
- Reddit’s Citation Share: Fell from 15% in minimal reasoning to just 7% in high reasoning.
- UGC & Review Sites: Overall visibility for forums, community hubs, and review platforms dropped from 14.3% to 6%.
This suggests that when the AI performs deeper logical validation, it devalues subjective public forums in favor of objective, structured data and authoritative corporate or journalistic sources.
3. Comparison Prompts as Search Engines
The divergence between the two modes was most pronounced during "comparison" prompts (e.g., "Compare the security features of Salesforce vs. HubSpot").
At this critical mid-funnel stage, the high-reasoning mode ran an average of 24 sub-queries per prompt, compared to a meager 5.5 sub-queries in minimal-reasoning mode. To synthesize these complex comparisons, the high-reasoning engine cited an average of 9.8 sources per response, compared to 5.8 for minimal reasoning.
Average Sub-Queries Executed for Comparison Prompts
┌──────────────────────────────────────┐
│ Minimal: 5.5 sub-queries │
├──────────────────────────────────────┤
│ High: 24.0 sub-queries │
└──────────────────────────────────────┘4. Vertical-Specific Variances: Finance Leads the Charge
The impact of high-reasoning search was not uniform across all industries. Highly regulated, data-dense sectors experienced the most significant shifts in visibility.
- Finance: This sector saw the most dramatic jump in citation rates and source shifting. Because financial queries require high factual accuracy and compliance, the reasoning engine executed exhaustive parallel searches, discarding shallow consumer reviews in favor of regulatory filings, institutional reports, and official corporate documentation.
- B2B SaaS & Consumer Tech: These verticals also saw substantial citation lifts, particularly during technical feature comparisons and integrations, where the AI bypassed marketing copy to read developer documentation and API guides.
- Health & Lifestyle: While citation rates increased, the shift was more muted compared to finance, though the AI showed a clear preference for peer-reviewed medical databases and established health institutions over wellness blogs.
Industry Perspectives: Decoding the AI’s "Thought Process"
Search industry experts and artificial intelligence engineers point to the underlying mechanics of large language models (LLMs) to explain these dramatic shifts.
In standard RAG (minimal reasoning), the system acts as a "lazy searcher." It takes the user’s query, inputs it into a search index, grabs the top three to five snippets, and writes a response. Because of this, standard SEO tactics—such as optimizing for high-ranking keywords, schema markup, and securing mentions on popular third-party review sites or Reddit threads—are highly effective.
In contrast, high-reasoning models utilize an "Agentic RAG" workflow. When presented with a complex query, the model does not just run one search. It breaks the query down into a tree of sub-questions. For instance, if asked to compare two enterprise software platforms, the AI might independently search for:
- The pricing pages of both companies.
- The user manuals or technical documentation for specific features.
- Security whitepapers and compliance certificates.
- Independent cybersecurity audits.
Because the AI is running dozens of micro-searches behind the scenes, it bypasses the surface-level web pages that traditional SEO targets. It seeks out the ground-truth data. Consequently, if a brand has optimized its blog posts but neglected its technical documentation, help center, or API logs, the high-reasoning AI will likely drop that brand from its final response.
Strategic Implications: How Marketers Must Adapt
The Semrush and Kevin Indig study serves as a wake-up call for digital marketers, SEO specialists, and brand managers. Relying solely on traditional search engine rankings or basic AI visibility metrics is no longer sufficient.
To maintain visibility across all layers of AI search, companies must adopt a dual-track optimization strategy.
1. Optimize for the "Deep Crawl"
Because high-reasoning engines perform exhaustive background searches, brands must ensure that their most valuable technical and product data is fully indexable and structured.
- Expose Help Centers and Documentation: Do not hide user manuals, API documentation, or customer support forums behind logins or unindexable formats. These are prime targets for reasoning engines executing technical sub-queries.
- Maintain Structured, Fact-Rich Formats: Use clear tables, bullet points, and explicit specifications. Reasoning models favor structured data that can be easily parsed and compared against competitors.
2. Focus on "Brand Persistence" and the Early Funnel
The study noted that high-reasoning engines tend to carry brands discovered in early research phases through to the final buying recommendations.
- Target Informational Queries: Ensure your brand is heavily associated with foundational industry concepts. If the AI identifies your brand as an authority during its initial broad searches, you are highly likely to remain in its "consideration set" as it drills down into complex comparison queries.
3. Diversify Beyond UGC and Forum Seeding
While "Reddit SEO" and forum seeding have become highly popular tactics to capture traffic from Google’s search layouts, this study proves those tactics have a shelf life in an AI-dominated world.
- Build Primary Authority: As high-reasoning engines deprioritize user-generated forums (dropping Reddit’s citation share by half), brands must reinvest in authoritative, third-party journalistic coverage, industry whitepapers, and rigorous case studies. The AI wants verified, professional sources, not anonymous forum comments.
4. Monitor the "Fan-Out" Queries
Marketers must begin tracking not just what keywords their brands rank for, but what sub-queries (or "fan-out" queries) AI engines generate when analyzing their product category. Tools that monitor generative search behavior will become essential for identifying the specific technical and comparative questions ChatGPT asks behind the scenes.
Conclusion
The era of treating AI search as a single, predictable algorithm is over. The Semrush and Kevin Indig study proves that as AI models become more intelligent, their methods of discovering and recommending brands diverge sharply from traditional search engines.
By running five times as many searches and ignoring the superficial forum chatter that currently dominates search engine results pages, ChatGPT’s high-reasoning mode demands a higher standard of digital presence. For brands to survive this transition, they must shift their focus from superficial keyword optimization to deep, authoritative, and structured information architecture.
Basiran 
Asep Darmawan 