Amazon’s A+ Content quality AI is a system AWS describes for judging a rendered product page as a whole, combining visual and readability signals to classify its quality. That page-level view lets the evaluator consider how images, text and layout work together. AWS outlines the system in its technical account of the evaluator.
A+ Content is assessed as a complete page: its images, text and layout contribute to one overall quality classification.
What the evaluator assesses
Amazon A+ Content is the text, images, video and layout presented as part of a product detail page. AWS says its system evaluates a rendered snapshot of that full page rather than scoring each asset on its own.
A vision-language model (VLM), which analyzes visual and textual information together, scores four dimensions: Visual Storytelling, Feature Presentation, A+ Layout and Conversion Optimization. Each receives a score from 0 to 10, along with a confidence value and written rationale. The categories cover both the presentation of the product and how the page is organized.
How visual and readability signals work together
Optical character recognition (OCR) extracts text and its placement. AWS uses it to calculate four layout-based readability signals: Text Density, Character Density, Spacing and Overall Readability. As a result, identical wording can produce different readability signals when its layout changes.
CatBoost, a classification model, combines those four OCR signals with the VLM’s four rubric scores. It predicts a high- or low-quality label using labels derived from human review. In AWS’s main dataset, 5,478 product detail pages across 10 categories were rated by an e-commerce marketing expert and four professional annotators; majority vote set the consensus label. AWS says 49.6% of those consensus labels were high quality.
SHAP identifies which input features pushed a prediction up or down. AWS says a large language model then uses negative feature contributions, category context, e-commerce practices and the VLM’s rationale to prioritize recommendations. The explanation links a page’s classification to factors the system considered, while the recommendation stage turns those factors into suggested revisions.
What AWS reports about model performance
AWS evaluated the configurations below on the same 20% holdout from its main dataset, about 1,100 pages. Accuracy is the share of pages classified correctly. The F1 score combines precision and recall, so it reflects how well the classifier identifies both quality classes.
| Configuration | Accuracy | F1 score |
| Averaged VLM scores with a fixed threshold | 0.80 | 0.84 |
| OCR-only CatBoost | 0.82 | 0.83 |
| VLM-only CatBoost | 0.89 | 0.90 |
| Fused OCR-VLM CatBoost | 0.94 | 0.95 |
The fused configuration had the highest reported accuracy and F1 score in this holdout comparison. Those figures describe classification against the dataset’s labels; they are not sales measurements.
AWS also reports an average 1.4-point increase on a 10-point quality scale after sellers revised content using recommendations. In a separate worldwide A/B experiment that adjusted the visibility of low-quality A+ Content, AWS reports a statistically meaningful improvement in Ordered Product Sales (OPS). The revision score, holdout metrics and OPS result describe different outcomes.
