Amazon's A+ Content quality check hits 0.94 accuracy on 5,478 pages

The system starts from a pragmatic choice: rather than fine-tuning a vision-language model on a few thousand labeled examples, Amazon keeps the model and its rubric prompt fixed and trains a CatBoost classifier on the resulting scores. That keeps updates to a short CPU-only job, which matters as annotations accumulate or criteria change.

A second layer addresses what the vision-language model misses. OCR extracts each block of text and its position, producing readability signals for text density, character density, spacing, and overall readability. Fusing those with the model's four rubric scores lifted accuracy to 0.94, and swapping the vision-language backbone still produced gains of 0.16 and 0.15, suggesting the complementary signal travels across models.

Beyond the score, SHAP assigns each feature a value showing how far it pushed a prediction, and an LLM turns the most negative features into prioritized actions. A diffusion-based component then generates ready-to-use image assets aimed at the dimensions SHAP flagged. Amazon notes that content standards differ across categories, languages, and marketplaces, and it plans to evaluate more efficient backbones for a global catalog.