Treyci Data: AI Engines Disagree on Brand Recommendations

On September 3, 2026, AI search intelligence platform Treyci published benchmark measurement research revealing that commercial conversational AI engines disagree sharply regarding which enterprise brands to recommend for identical commercial buying inquiries. In a rigorous multi-engine study analyzing B2B software categories, one leading generative engine recommended tracked brands across 81% of generated responses, while a competing major engine recommended the identical vendor entities in only 43% of answers—representing an almost two-to-one visibility disparity across platforms that corporate buyers routinely treat as interchangeable.

The research highlights a critical vulnerability in how marketing organizations currently evaluate Generative Engine Optimization (GEO). Many digital marketing teams assess their presence within artificial intelligence answers through anecdotal manual checks: a team member enters a single product query into ChatGPT, takes a screenshot of the response, and presents the result as definitive proof of market leadership or brand absence. Treyci’s longitudinal telemetry demonstrates that because foundation models operate through non-deterministic token sampling, the identical engine submitted the exact same query in separate user sessions frequently delivers divergent vendor selections. A single query execution captures statistical variance rather than an authoritative baseline.

Equally consequential for corporate content strategies, Treyci discovered that when conversational models answer commercial evaluation queries, the overwhelming majority of cited sources originate from independent third-party platforms rather than official corporate domains. Review aggregators such as G2, Trustpilot, and Capterra, alongside independent trade publications and comparative industry analyses, accounted for over 70% of total attributed citations. This empirical reality inverts traditional search engine marketing: optimizing on-site landing pages provides negligible generative visibility if an organization fails to cultivate active, positive consensus across the external review networks and independent media outlets that models crawl for validation.

Probabilistic Model Divergence and Third-Party Citation Mechanics

The marked divergence between generative engines when evaluating enterprise brands stems from the complex interplay of model training corpora, retrieval mechanics, and system architecture. Understanding why ChatGPT, Perplexity, Google Gemini, and Claude arrive at vastly different vendor recommendations requires examining the underlying technology driving each platform.

First, the underlying pre-training datasets establish divergent baseline entity relationships. Large foundation models are trained on distinct web scrapes, academic archives, and licensed publication feeds. If an enterprise possesses deep historical documentation across GitHub, technical developer forums, and open-source documentation, a model with higher pre-training density in software engineering archives will assign higher baseline probability to that brand. Conversely, a model trained on broader consumer web data may favor better-known consumer software suites.

Second, the real-time retrieval-augmented generation (RAG) architectures deployed across these engines utilize fundamentally different search backends and ranking heuristics. For example, Perplexity Pro routes queries through dedicated multi-source search scrapers optimized for real-time journalistic and research retrieval, weighting recent review updates and trade press heavily. OpenAI’s search-enabled models leverage specialized web indexing pipelines integrated with Bing infrastructure, while Google Gemini queries Google’s real-time Knowledge Graph and local business indexes. Because each search backend applies unique passage relevance scoring, the context windows of these models are populated with completely different source snippets for the identical user question.

Third, temperature settings and top-p sampling introduce inherent non-deterministic behavior. When a user asks an AI engine to recommend three CRM solutions, the model calculates a probability distribution across dozens of qualified software entities. Unless the temperature parameter is pinned to zero, the model randomly samples from high-probability candidate tokens. In Session One, the model might select vendors A, B, and C; in Session Two, small probability shifts might yield vendors A, B, and D. Consequently, evaluating AI visibility requires aggregate probabilistic testing across dozens of independent sessions rather than relying on point-in-time spot checks.

Finally, Treyci’s findings confirm that foundation models maintain severe algorithmic skepticism toward self-published corporate claims. When synthesizing answers to buying questions, models are instruction-tuned to prioritize neutral, multi-source corroboration. Citations from independent aggregators like G2, Capterra, and TrustRadius provide third-party validation that insulates the model from repeating promotional hyperbole. An organization that concentrates its SEO investments solely on its proprietary domain while neglecting independent review platforms effectively renders itself invisible to the retrieval mechanisms powering conversational AI.

AI Visibility Measurement Best Practices

The table below contrasts rigorous, statistically valid AI visibility measurement methodologies against flawed, anecdotal approaches:

Real-World Utility & Policy Implementation

Marketing organizations seeking to capture market share within generative answer engines must abandon subjective manual testing in favor of automated, multi-engine measurement infrastructure. Measuring baseline recommendation share allows teams to allocate content and reputation resources where algorithmic deficiencies are most acute.

The 4-Step AI Visibility Measurement Playbook

1. Establish a Statistically Valid Multi-Engine Benchmark: Construct a standardized prompt library containing at least fifty commercial intent variations reflecting your prospective buyers’ decision journey. Execute these prompt matrices programmatically across ChatGPT, Perplexity, Gemini, and Claude across multiple independent sessions. Calculate your baseline recommendation share, competitor inclusion rates, and sentiment scores across each engine.
2. Diagnose and Prioritize Engine-Specific Gaps: Analyze the variance across engines. If your brand achieves 80% recommendation share in ChatGPT but languishes at 40% in Perplexity, inspect Perplexity’s citation sources. Identify which trade reviews, comparison portals, or industry analyses Perplexity references for that query category, and prioritize outreach to those specific external platforms to close the visibility gap.
3. Reallocate Content Budgets Toward Third-Party Consensus: Shift content marketing investments from self-published corporate blog posts toward external authority hubs. Launch coordinated review acquisition initiatives on verified software review platforms (G2, TrustRadius, Capterra). Ensure executive thought leadership and case studies are distributed across independent media outlets that search-integrated AI scrapers crawl for factual validation.
4. Integrate AI Attribution into CRM Lead Capture: Upgrade digital marketing analytics to identify buyers originating from generative search engines. Implement referral tracking for AI domain referrers (e.g., chatgpt.com, perplexity.ai) and update sales intake forms with open-ended attribution fields (“How did you discover our solution?”). Track pipeline velocity and customer acquisition cost for AI-sourced opportunities.