One company. Seven models. One number they can't agree on.
The AI Perception Dossier is built on a claim you can measure. Ask seven frontier models about the same company and they come back with materially different verdicts. This is the paper that puts a number on the gap.
Andrew Pomazkov · VeritasLinks · 17 pages · Posted 19 July 2026 · SSRN
Abstract
Large language models are increasingly the first source consulted when an investor, analyst, or counterparty evaluates a company. Yet different models, asked identical questions about the same firm, routinely return materially different assessments—a phenomenon we term split-perception. Drawing on a dataset of AI-generated assessments covering more than 5,000 companies and several million individual model responses collected between 2025 and 2026, we introduce a structured methodology for scoring AI-mediated company perception on a 300–870 scale modeled on consumer credit scoring, decompose the score into visibility, positioning, and risk components, and quantify cross-model divergence. We find that divergence between models on the same company is large and structured: in the fully instrumented subset, the median gap between the highest- and lowest-scoring models is 100 points, with a 90th percentile of 134, and in the longitudinal case study the single-date split reached 115 points. While four of the five judge models were individually stable across repeated runs, the retrieval-grounded model's verdict on the same firm varied by more than 200 points. We argue that split-perception constitutes an unpriced information risk for M&A due diligence, KYB review, and credit assessment workflows that rely, formally or informally, on AI-assisted research, and we distinguish divergence that is a stable property of the model-evidence pair from volatility imported by live retrieval.
Keywords · large language models · company credibility · perception scoring · due diligence · KYB · information risk · generative engine optimization
JEL · D83 · G24 · G32 · G34 · M31 · O33
The gap is not noise. It is structured.
companies analyzed, across several million individual model responses collected through 2025 and 2026.
median gap between the highest and lowest scoring model on the same company, on the study's 300 to 870 scale.
how far a single retrieval grounded model moved its verdict on one firm across repeated runs, while the other models held steady.
An unpriced risk hiding inside everyone's research.
More and more diligence starts with an AI model. M&A review, KYB and counterparty checks, credit and risk assessment. If the model your counterparty happens to open lands a hundred points below the one you would have trusted, nobody sees the gap and nobody prices it. That gap is what VeritasLinks measures.
How the score is built.
Each company's perception score decomposes into three parts. Visibility, positioning, and risk. The method separates divergence that is a stable property of a model and its evidence from volatility imported by live retrieval, so a number that simply moves and a number that is structurally biased are never mistaken for each other.
The same 1000+ prompt, seven model method powers every dossier we generate.
Cite this paper
Pomazkov, Andrew. Split-Perception: Measuring Divergence in How Large Language Models Assess Company Credibility (July 19, 2026). Available at SSRN. https://doi.org/10.2139/ssrn.7145478
@misc{pomazkov2026splitperception,
title = {Split-Perception: Measuring Divergence in How Large Language Models Assess Company Credibility},
author = {Pomazkov, Andrew},
year = {2026},
month = {7},
note = {SSRN Working Paper},
doi = {10.2139/ssrn.7145478},
url = {https://doi.org/10.2139/ssrn.7145478}
}Version of record · DOI 10.2139/ssrn.7145478 · https://doi.org/10.2139/ssrn.7145478
See where the models place your company.
The research is the map. The dossier is your coordinates on it.