As of 6 November 2026, Amazon Marketplace is grading A+ Content, the rich-media block on product detail pages, with a frozen vision-language model combined with OCR signals and a small CatBoost classifier. According to an AWS Machine Learning blog post published on 5 November, the fused model reached 0.94 accuracy and 0.95 F1 on a 20 percent holdout of 5,478 labelled pages, against 0.80 accuracy and 0.84 F1 for a simple average of the model's own scores. Amazon also reports that sellers using high-quality A+ Content see sales 8 to 20 percent higher than those without. For brands running large catalogues, this is the first public look at how a marketplace turns page quality into a measurable score.
Why the page is judged as a whole
The design choice matters more than the accuracy number. Checking assets one at a time misses the failure mode that actually hurts conversion: a page can have a sharp hero image, grammatical bullets and a good lifestyle shot, and still be wrong. The bullets may describe a 64-ounce pitcher while the photo shows a single-serve model, or the lifestyle image may sit above the hero.
Amazon's answer is to render the entire content experience as one snapshot and score it against a four-part rubric: visual storytelling, feature presentation, A+ layout and conversion optimisation, each scored 0 to 10 with a confidence value and a written rationale. The scores and confidences become features; the rationales go back to sellers as guidance.
Why the model is frozen
Fine-tuning a vision-language model on a few thousand labels risks overfitting and needs GPUs. Amazon kept the model and prompt fixed, used its structured outputs as inputs, and trained a small classifier on top. Only the classifier is retrained when labels or criteria change, which turns a GPU problem into a short CPU job.
OCR still carries weight
The pipeline does not rely on vision alone. OCR signals, including text density, character density, spacing and overall readability, are added as features. That makes sense for a marketplace: much of what makes a page hard to read is typographic, and a vision model scoring a rendered image can miss how dense the copy actually is.
| Signal | What it captures | Why it matters |
|---|---|---|
| VLM rubric scores | Story, features, layout, conversion | Holistic judgement of the page |
| OCR readability | Text density and spacing | Catches unreadable layouts |
| Classifier output | Calibrated quality label | Turns judgement into a score |
SHAP values explain which features drove each score, and an LLM converts those explanations into recommendations for the seller. In the reported test, sellers raised their quality score by 1.4 points out of 10 after acting on the recommendations. That is a modest but real improvement, and it suggests the feedback loop is usable rather than decorative.
What this means for catalogue teams
First, the unit of work changes. Instead of briefing a photographer, a copywriter and a designer separately, teams should review the rendered page as a customer sees it, because that is what gets scored. Most obvious inconsistencies, such as a variant mismatch between image and bullet, only become visible at page level. In a catalogue of thousands of SKUs, that means adding a review step that looks at the finished page rather than at each asset in isolation.
Second, the rubric gives a shared vocabulary. Visual storytelling, feature presentation, layout and conversion are broad, but they are stable enough to use in internal reviews, which is more than most content guidelines achieve. It also makes category growth comparable: two teams can argue about a page without arguing about what quality means.
Third, the sales claim should be read carefully. The 8 to 20 percent lift is reported by Amazon and covers sellers with high-quality content versus those without; it is not a promise that improving a score by one point yields a proportional sales gain. Category growth, price and traffic all sit between page quality and revenue.
Where AI-generated listings fit
The same week, tools that generate a full listing from a single product photo in minutes went on sale to small merchants. If listing production gets that cheap, page quality stops being a production problem and becomes a review problem: the constraint moves from writing copy to catching errors before they reach the snapshot.
That is why page-level scoring is likely to spread. When anyone can generate a plausible page in minutes, the scarce skill becomes judging whether the page is internally consistent, and that judgement is exactly what a rubric-based model can be trained to reproduce. Marketplaces have an obvious incentive to automate it, because they carry the cost of a bad page through returns and complaints. Sellers, in turn, get a score they can act on before the customer sees the page.
FAQ
What exactly is being scored?
A+ Content, the rich-media block on Amazon product detail pages, rendered as one full-page snapshot rather than judged asset by asset.
How accurate is the system?
AWS reports 0.94 accuracy and 0.95 F1 on a 20 percent holdout of 5,478 labelled pages, against 0.80 and 0.84 for a simple average of model scores. The labels came from one marketing expert and four annotators.
Does a better score mean more sales?
Amazon reports that sellers with high-quality A+ Content see sales 8 to 20 percent higher, but the comparison baseline is not described and the effect varies by category.
What should a catalogue team do first?
Review the rendered page as a whole and check that the hero image, bullets and price describe the same variant. Variant mismatches are the most common page-level failure.
