Amazon describes an approach for evaluating A+ Content quality at scale on product detail pages. A+ Content is a rich-media widget through which brands use visual assets to communicate product and brand value. According to the source, selling partners using high-quality A+ Content increase sales by 8–20 percent compared with partners that do not use it.

Evaluating the page as customers see it

Instead of judging images and text blocks separately, the system renders an entire product detail page as one snapshot. A vision-language model then assesses the combined images, text, and layout as they appear to a customer. The unit of evaluation is therefore the composed page, preserving relationships among elements that could be lost if each asset were reviewed in isolation.

The model follows a fixed rubric covering four dimensions: Visual Storytelling, Feature Presentation, A+ Layout, and Conversion Optimization. For each dimension, it returns a score from 0 to 10, a confidence value, and a natural-language rationale. Both the vision-language model and rubric prompt remain fixed. Their structured outputs become features for a CatBoost classifier trained to predict expert quality labels.

Adding OCR-based readability signals

A separate OCR model hosted on SageMaker AI Inference extracts text blocks and their positions. It calculates Text Density, Character Density, Spacing, and Overall Readability. Those four signals are combined with the four vision-language scores, producing an eight-feature input for CatBoost. The design brings assessments of visual composition together with measurements derived from the placement and concentration of text.

The training labels came from independent ratings by one ecommerce marketing expert and four professional annotators. They reviewed 5,478 product detail pages spanning 10 product categories. Majority vote produced the consensus labels, and 49.6 percent of the pages were classified as high quality. Each annotator’s Cohen’s kappa against the consensus label ranged from 0.845 to 0.915.

Performance on the shared holdout set

All reported comparisons used the same 20 percent holdout set. An averaged-score baseline reached 0.80 accuracy and 0.84 F1. CatBoost using only the vision-language features reached 0.89 accuracy and 0.90 F1, while OCR-only CatBoost reached 0.82 accuracy and 0.83 F1. The fused OCR–VLM–CatBoost model produced the strongest reported result, with 0.94 accuracy and 0.95 F1.

SHAP analysis also showed how strongly the features contributed across the holdout set. The highest mean absolute SHAP value was OCR Character Density at 1.1. OCR Overall Readability and VLM Visual Storytelling each reached 1.0, followed by A+ Layout at 0.8, Feature Presentation at 0.6, and Conversion Optimization at 0.4. These values describe the classifier’s feature attributions on the evaluated set and help connect its predictions to the signals it used.

Turning a low-quality result into actions

For a page classified as low quality, the system ranks all eight features by SHAP value. It sends the most negative features, the product category, ecommerce best practices, and the vision-language model’s rationales to a language model. The resulting response consists of prioritized, specific actions. This explanation layer changes the output from a bare quality classification into guidance about which weaknesses should receive attention first.

After selling partners revised their content using these explanation-based recommendations, the average post-revision quality score increased by 1.4 points on a 10-point scale in a statistically significant sample. The reported result concerns the quality score after revisions and directly links the recommendation workflow with a measurable change in that score.

Operations and updating

The described SageMaker AI implementation maps page rendering, vision-language scoring, OCR extraction, feature aggregation, CatBoost training, SHAP explainability, and production monitoring to managed capabilities. Monitoring covers data quality, model quality, and feature-attribution drift. SageMaker Pipelines automates retraining when drift thresholds are breached.

Because the vision-language model and rubric prompt remain fixed, an update can be made by retraining only the classifier. The source characterizes this as a short, CPU-only job. That separation preserves the same structured evaluation dimensions while allowing the expert-label predictor to be updated without retraining the vision-language component.