The Product Data Legibility Gap: Why LLMs Cannot Recommend What They Cannot Read

Large language models are now active participants in consumer purchase decisions. When a shopper asks ChatGPT which foundation suits combination skin, the model does not browse a retailer's website in real time. It draws on a training diet of product descriptions, retailer pages, and structured attributes ingested during pre-training and retrieval. If those attributes are inconsistent across sources — or absent entirely — the model preferences the source it has highest confidence in, regardless of whether that source reflects the brand's canonical positioning. This paper documents what we call the Product Data Legibility Gap: the structural mismatch between a brand's internal product information and the fragmented, contradictory version of that product data that LLMs actually read. Using a composite dataset drawn from AIVO Meridian diagnostic runs across multiple US cosmetics brand portfolios, we identify 12 PIM attributes that explain approximately 80% of cross-retailer LLM citation variance, demonstrate the mechanism by which attribute fragmentation suppresses AI purchase recommendations, and describe the remediation architecture required to close the gap. The findings indicate that product data legibility — not content volume, not brand visibility, and not prompt engineering — is a primary variable governing AI decision-stage outcomes in this dataset, and one whose structural logic applies to any consumer product brand distributing through a multi-retailer footprint.