Amazon Marketplace has developed a system for evaluating A+ Content as a complete product-page experience rather than checking each image or text block separately. The approach renders the full page as one snapshot, allowing a vision-language model to assess how its imagery, copy, and layout work together.
The model scores four dimensions: visual storytelling, feature presentation, A+ layout, and conversion optimization. Amazon keeps the vision-language model fixed and feeds its scores into a CatBoost classifier trained on expert labels. The training set contained 5,478 pages across 10 product categories, with consensus labels drawn from one ecommerce marketing expert and four professional annotators.
Optical character recognition adds four layout-based signals covering text density, character density, spacing, and overall readability. On a 20 percent holdout set, a simple averaged-score baseline reached 0.80 accuracy and a 0.84 F1 score. The combined OCR, vision-language, and CatBoost system reached 0.94 accuracy and a 0.95 F1 score.
The pipeline uses SHAP values to identify which features most affected each decision. An LLM then turns those findings and the model’s rationales into prioritized recommendations for sellers. Amazon reports that revised content gained an average of 1.4 points on a 10-point quality scale, while a production experiment produced a statistically meaningful improvement in Ordered Product Sales.