AWS described a system for assessing A+ Content quality on Amazon Marketplace using vision-language models

A+ Content is a module with images, text, and video on a product page in Amazon Marketplace, used by sellers to present their brand. According to AWS, sellers with high-quality A+ Content sell 8-20% more than those without it, but because AWS does not impose strict content specifications (image aspect ratios, word count), quality varies strongly between brands — ranging from companies with their own design teams to sellers with limited resources.

AWS described an approach that evaluates the entire product page as a single snapshot, rather than assessing individual elements in isolation — a typical computer vision model will detect a blurry image, but won't notice, for example, a discrepancy between text describing a container's capacity and a photo of a different model. A vision-language model (of the CLIP, Flamingo, LLaVA type) scores the page snapshot according to a rubric with four dimensions — visual storytelling, feature presentation, A+ layout, and conversion optimization — and returns a score of 0-10, a confidence level, and a text justification for each dimension. These outputs serve as input for a CatBoost classifier, trained on expert ratings; according to AWS, the system can be retrained with a quick CPU-based job, without the need to train the large model itself. According to the source, diffusion models then generate improved visual elements based on the resulting quality signals.

The source text is incomplete — the description of the process for creating expert training data and other technical details are missing. See the source article for details.