Split-Perception: When AI Models Disagree, Risk Increases

You have likely asked two different AI models for a vendor recommendation and received conflicting answers. This is not random noise. It is a structural feature of how large language models process information. VeritasLinks calls this Split-Perception: the phenomenon where different models judge the same company differently, creating a measurable gap that signals risk [F32].

Most visibility tools report a single pass as if it were a fact. They miss the spread. VeritasLinks measures this disagreement by running many prompts across several models with repeats [C3]. This method separates the signal from the noise of random variation. The mechanism is grounded in academic research. Andrew Pomazkov’s Split-Perception study, published on SSRN, validates that the divergence between models is a distinct risk signal [C6].

The gap between models matters for enterprise risk teams. A company might be highly visible but barely preferred if the models are split. The disagreement itself is the data point. When models agree, confidence is high. When they split, the uncertainty is real and quantifiable. This split-perception metric notes that the gap between different models judging the same company is a risk signal [F3556].

VeritasLinks analyzes a company across several AI models, including ChatGPT, Claude, Grok, Perplexity and DeepSeek, and shows a separate score for each model [C2]. The analysis includes the emotional tone each model carries about the company, known as the Perception Affect Field [C14]. This allows risk teams to see not just the score, but the sentiment behind the judgment.

The chart [cross_model_gap] shows how far apart AI models score the same company, with a typical median gap of 100 points and a 90th percentile gap of 134 points on the VeritasScore scale.

You ask an AI model for a vendor recommendation and receive a confident answer. The model mentions your company repeatedly. It uses your name. It sounds authoritative. Yet when you dig deeper, you realize the model is reciting a press release rather than making a genuine choice. This is the core trap of single-pass AI visibility tools. They count citations as if they were votes. They are not.

Citation is not recommendation. A company can be highly visible and barely preferred by the models that judge it [F31]. Visibility is often a function of data volume and recency. Recommendation is a function of judgment and fit. When a model cites a brand, it acknowledges existence. When it recommends a brand, it endorses utility. The gap between these two states is where risk hides for M&A analysts and risk reviewers.

VeritasLinks measures this gap. By running multiple prompts across several models with repeats, it separates the signal of true preference from the noise of mere mention [C3]. The result is a clearer picture of how a company is actually perceived by the stochastic engines that increasingly influence enterprise decisions. If a model mentions you but does not prefer you, your AI visibility score may be inflated.

The data supports this separation. An analysis of AI-visibility vendors in July 2026 reveals a stark divergence between brand visibility and buyer preference. Semrush led visibility at 135 points but had a preference score of only 3. Profound followed with 130 visibility and 5 preference. AthenaHQ sat at 129 visibility and 3 preference. Ahrefs Brand Radar showed 127 visibility but just 2 preference. Nightwatch and Bluefish hovered near 125 and 121 visibility respectively, with preference scores of 6 and 5.

This chart shows that high visibility does not guarantee high buyer preference.

VeritasLinks avoids the single-pass trap by running repeated prompts across five distinct models: ChatGPT, Claude, Grok, Perplexity, and DeepSeek [C2]. This multi-model approach captures the stochastic nature of large language models, measuring the spread of answers rather than treating one random draw as a fixed fact [C3]. The result is a separate score for each model, revealing where perceptions diverge [C2]. This divergence, or Split-Perception, is itself a risk signal for due diligence [C5].

The core output is the VeritasScore, a perception metric on a 300 to 870 scale [C4]. This score aggregates several components, including the emotional tone each model carries toward the company [C14]. It also includes an AI-simulated focus group where synthetic buyer personas discuss the firm and surface objections [C7]. An AI Investor Review adds another layer, showing how a credit officer or investor would read the company through AI [C8].

VeritasLinks also tracks recommendations derived from the relationships between these metrics [C9]. Because the analysis is independent, the score cannot be bought through paid placement [C13]. This independence matters for M&A analysts and KYB reviewers who need to trust the data source. The system monitors where AI models disagree about the same company, treating this disagreement as a risk signal [C5].

The chart illustrates this distribution. Among 139 companies with fully scored analyses in July 2026, the median VeritasScore is 437 [score_population]. Only about one in eight companies scores 500 or higher, and none have reached 700 [score_population]. The highest score measured so far is 564 [score_population]. This data shows that while visibility is common, high perceived credibility is rare.

The chart shows that among 139 companies with fully scored analyses in July 2026, the median VeritasScore is 437, with only about one in eight companies scoring 500 or higher and none reaching 700 [score_population].