Is Your AI Due Diligence Stable or Stochastic?

A single AI report is often treated as a definitive fact, yet it is usually just one random draw from a stochastic process. If you ask the same model to evaluate a vendor twice, you may get two different answers. This variability is not a glitch. It is a feature of how large language models generate text.

The problem is that most visibility tools report a single pass as if it were the whole truth. They count citations or provide a score based on one generation. This method fails to distinguish between signal and noise. VeritasLinks measures the spread of answers by running many prompts across several models with repeats [F4533]. This approach reveals the stability of the perception, not just its presence.

The stakes for buyers and sellers are rising. Gartner predicts that by 2027, 95% of sellers’ research workflows will begin with AI [F4659]. Yet, 51% of buyers say they are more likely to encounter misleading information from GenAI compared to 49% for sales reps [F4658]. A single-pass report can mask the true consensus. It can make a marginal perception look like a strong recommendation, or hide significant disagreement between models.

VeritasLinks addresses this by analyzing a company across ChatGPT, Claude, Grok, Perplexity, and DeepSeek [C2]. It does not just average the scores. It measures where the models disagree. This Split-Perception view shows whether a company is broadly accepted or if the judgment is fragile and model-dependent. The gap between the highest and lowest model score for the same company is itself a risk signal.

The chart shows that while a typical company has a cross-model gap of 100 points on the VeritasScore scale, one in ten companies sees a gap of 134 points, indicating significant perception drift.

Different models do not see the same company. They project distinct perceptions that shape real decisions like purchases and due diligence [F4512]. When you ask ChatGPT and Claude about a vendor, you get two different answers. This is not a bug. It is the nature of stochastic reasoning.

The gap between these models is a risk signal. Research by VeritasLinks founder Andrew Pomazkov, published on SSRN, defines this as Split-Perception [C6]. It measures where AI models disagree about the same company [C5]. A company might be highly visible but barely preferred by the model that matters most for your specific deal.

This disagreement has tangible costs. AI washing can lead buyers to overvalue a target [F4704]. If you rely on a single model pass, you might miss the divergence. You assume consensus where there is only coincidence. The stability of your AI research determines whether you see a real opportunity or a random snapshot.

Consider Bluefish AI, analyzed by VeritasLinks in July 2026. Five models gave five different opinions. The spread was 121 points on the 300 to 870 scale. DeepSeek scored it 530, while ChatGPT scored it 409. That gap changes the narrative. It changes the recommendation. A 121-point difference is not noise. It is a structural shift in how the market views the company. DeepSeek sees a stronger candidate than ChatGPT. If your due diligence relies on ChatGPT, you might undervalue the target. If you rely on DeepSeek, you might overpay.

The chart shows how the same company can be perceived differently by DeepSeek, Claude, Perplexity, Grok, and ChatGPT, with a 121-point spread between the highest and lowest scores.

VeritasLinks measures how large language models perceive, describe and judge a specific company [C1]. It is not a GEO visibility tool that counts citations. It measures the judgment layer underneath. This means it determines whether models consider a company credible, whether they recommend it, and where models disagree. Citation is not recommendation. A company can be highly visible yet barely preferred by the stochastic engines that now influence purchasing and partnership decisions.

The platform analyzes a company across several AI models, including ChatGPT, Claude, Grok, Perplexity and DeepSeek, and shows a separate score for each model [C2]. Each analysis runs many prompts across these models with repeats and measures the spread of answers, instead of reporting a single model answer as a fact [C3]. This approach separates signal from noise. If ChatGPT gives a positive verdict but Claude is skeptical, the gap itself is a risk signal. The analysis includes the emotional tone each model carries about the company, known as the Perception Affect Field [C14].

Scores are independent and cannot be bought, ensuring no paid placement bias [F4317]. This independence is critical for M&A analysts and KYB reviewers who rely on AI outputs for due diligence. AI-driven hiring or employee evaluation tools may create potential discrimination concerns if they produce biased outcomes [F4705]. Similarly, AI-generated outputs are generally not copyrightable or patentable, and inclusion of trade secrets in AI data sets jeopardizes protection [F4706]. VeritasLinks captures the stability of these judgments over time through continuous re-runs on subscription [C10].

The VeritasScore is a perception score on a 300 to 870 scale built from several components [C4]. The chart shows where companies land on this scale, with a median score of 437 among 139 fully scored companies, and only one in eight scoring 500 or more.