Amazon is trying to grade A+ Content the way a shopper actually sees it: one page, not a stack of separate files. Picture a lifestyle photo of a single glass parked under copy that promises a 64-ounce pitcher. The photo can be sharp. The sentence can be grammatical. Scroll the listing, and it is arguing with itself.
That mismatch is the whole problem. A+ Content (the brand modules of images, video, carousels, hotspots, and story on an Amazon detail page) is not pinned to a required image ratio or a word-count cap. Large brands with in-house design teams use it. So do smaller brands that do not. Quality swings, and it does not live in any single asset. It shows up in how the pieces sit together.
Grading those pieces in isolation is like reviewing a band by listening to each instrument in a different room. You can confirm that everyone showed up. You still have no idea if they were playing the same song. The useful system is a model that grades the whole page, then tells a seller what to fix first.
A loose spec, a strict customer
On Amazon’s seller site, A+ Content is a Brand Registry program: larger images, video, interactive modules, Q&A, shoppable content, and a separate Brand Story. Publishing it requires a Professional selling account and, for branded products, a Brand Representative or Reseller role. Submitted content is reviewed before it appears on brand-owned listings, and status lives in A+ Content Manager. That public workflow is a gate. By itself, it does not say whether the finished page is any good.
In a technical write-up on how Amazon Marketplace evaluates content quality with Amazon SageMaker AI, Amazon’s applied scientists describe a different layer: an internal quality signal for the page as the customer sees it. Render the full A+ experience as one snapshot. Ask a vision-language model to score that snapshot. Add a second read on whether the text is actually readable. Then let a small classifier decide high or low, and tell the seller what to change.
A vision-language model, in plain terms, is an image reader wired to a language model. You hand it a picture and a rubric. It writes an assessment without being retrained for that exact job. Amazon does not name the production model. What the authors specify is the rubric: Visual Storytelling, Feature Presentation, A+ Layout, and Conversion Optimization. Each comes back as a score from 0 to 10, a confidence value, and a short rationale. Those names are the team’s, not a public Amazon style guide.
They matter because they force a judgment of composition. Hand the model that pitcher page and the failure is not blur or grammar. It is a claim and a photo that are not telling the same story, a lifestyle image stranded above the hero, or a layout that buries the feature a shopper came to compare.
Why a second opinion from the letters on the page
Vision models have a bias that will sound familiar to anyone who has watched a draft get longer in the name of being thorough. The authors say they can prefer dense, detailed copy even when a shopper wants something they can scan on a phone.
So the team adds optical character recognition (software that reads the text off the rendered page) and turns that read into four scores. Text density is how much of the page area is covered by words. Character density is how many characters sit in a given area. Spacing looks at whether the whitespace is consistent and enough. Overall readability is then built from those three. The same paragraph can score differently if you cram it or give it room. Layout is part of the meaning.
Overall readability is built from the other three, so it is not a separate fact about the page. The authors still feed all four scores into the classifier alongside the four rubric scores. Eight numbers, with some overlap built in. Worth knowing before anyone treats the list as eight clean measurements.
The classifier is CatBoost, a form of gradient-boosted decision trees. Less glamorous than a foundation model, and the right tool for a small table of features. Trees can learn that the signals are not equally important, and that they interact. A beautiful story on an unreadable layout is not “medium.” It is a specific kind of failure.
The cheap model learned the weights
The expensive work was human labeling and the vision pass. The labels came from people, not from the model grading its own homework. The authors report that one ecommerce marketing expert and four professional annotators independently marked 5, 478 product detail pages, across 10 categories, as high or low quality. Furniture, Shoes, Grocery, Beauty, and Apparel are named among those categories. The training target was the majority vote. Classes were nearly balanced: 49.6 percent high quality.
Agreement with that consensus was strong. Cohen’s kappa (agreement beyond what you would get by chance, where 0 is chance and 1 is perfect) ran from 0.845 to 0.915 for each annotator against the consensus. Values above 0.8 are generally treated as strong. The authors also say content standards differ by category, language, and marketplace, and that these annotations cannot stand in for every context.
They held out 20 percent, about 1, 100 pages. They do not say whether that split was balanced by category. On the pages they do report, the pattern is hard to shrug off:
- Averaging the four vision scores and applying a fixed threshold: 0.80 accuracy, and 0.84 F1 on the high-quality class.
- Readability scores alone, in CatBoost: 0.82 accuracy, 0.83 F1.
- Vision scores alone, in CatBoost: 0.89 accuracy, 0.90 F1.
- Vision scores and readability together: 0.94 accuracy, 0.95 F1.
F1 is a combined read of precision and recall. On a nearly even split, it is not a vanity number hiding behind a lopsided dataset. The authors report the fused model as 0.05 higher on both accuracy and F1 than vision scores alone, and 0.14 higher in accuracy and 0.11 higher in F1 than the equal-weight average. Letting the trees learn unequal weights did most of the first jump, from 0.80 to 0.89, before readability was added. Fusion then beat both inputs. The text read was not a duplicate of the vision model. It was the missing argument.
A smaller check asked whether this only works with one heavy model. On a single-category set of 499 content projects (the authors’ unit, and a much narrower set than the 5, 478 pages), a lighter vision model moved from a baseline of 0.74 accuracy and 0.72 F1 to 0.90 on both once the signals were fused. A more capable model moved from 0.79 accuracy and 0.84 F1 to 0.94 and 0.95. Accuracy gains of 0.16 and 0.15. The backbone changed. The lesson did not. Amazon still does not name either model, so this is evidence about the pattern, not a shopping list of vendors.
Fix the layout before you reshoot the photo
A score that cannot explain itself is a dead end for a selling partner. The team uses SHAP, a way of showing how far each input pushed a specific decision toward high or low quality, on a shared scale the raw 0-to-10 scores cannot provide.
On the holdout, average influence on the yes-or-no call ranked like this: character density 1.1, overall readability 1.0, visual storytelling 1.0, A+ layout 0.8, feature presentation 0.6, and conversion optimization 0.4. Read that slowly. The dimension with the most CFO-friendly name contributed the least, on average, to the quality call. Conversion copy can still matter to revenue. In this model, readability and storytelling moved the decision more than the score explicitly labeled for conversion.
Local explanations then set the work order. As an illustration of the logic, not as the audit of a live listing: if layout pulls the decision down by 1.2 and visual storytelling by only 0.3, restructuring comes before a new photo shoot. An LLM turns the most negative signals into a prioritized list for the seller.
The authors treat that list as the point, not as a transparency annex. Pages revised from those recommendations rose 1.4 points on a 10-point scale. They report the gain as statistically significant and do not publish the sample size, the test, or a p-value. They also do not show whether the sellers who revised were already the ones most likely to improve. Directionally useful. Not a clean causal proof.
There is a further loop: a diffusion model that can take the same signals and generate a stronger lifestyle image, a rebalanced layout, or a cleaner product shot. That capability is described. A separate quality metric for the generated assets is not. Capability and a published win rate are different sentences. Keep them apart.
The revenue sentence is still incomplete
Two sales claims travel with this work, and they are not the same claim.
The same write-up says selling partners with high-quality A+ Content see an 8 to 20 percent sales lift versus pages without it. That is a statement about the content program, not a result of this classifier. Amazon’s public A+ marketing page does not use that range. It sells a softer bet: brand story, discovery, trust, and sales, without a number you can audit. A brand manager’s testimonial on that page is marketing, not an evaluation of rubric scores or SHAP.
The experiment tied to the classifier is narrower, and more interesting for anyone thinking about AI for sales. The authors describe a worldwide test that changed the visibility of low-quality A+ Content based on the model’s decisions, and report a statistically meaningful improvement in Ordered Product Sales, the value of products customers actually ordered, not clicks. They do not report the size of that lift, the markets, the duration, the sample, or exactly how visibility was changed. “Statistically meaningful” is not a number you can take to a board. It is a direction, with the magnitude left in the drawer.
Sellers already have a different experiment tool. In Seller Central discussions, Amazon has described Manage Your Experiments as covering A+ Content, titles, and main images, with image tests framed as free, creatable in less than five minutes, and reported weekly, aimed at conversion. Program details move, so treat that as Amazon’s description, not a permanent spec. It helps a brand pick a variant. It is not the same thing as a marketplace model deciding that weak modules should be seen less. One is a seller’s A/B test. The other is a platform quality lever. If you sell on Amazon, know that both exist, and do not assume you can see the second one’s math.
What to copy, and what to leave on the slide
The architecture choice is the part worth stealing for AI automation outside retail.
Fine-tuning the vision model on a few thousand labels would have meant GPU spend and a real risk of overfitting: teaching the big model the quirks of this label set rather than a stable idea of quality. The team left the vision model and the rubric frozen. When standards drift, from new module types, shifting taste, or sellers writing differently, they describe monitoring and retraining the small classifier once drift thresholds are breached. Those thresholds are not published. An update is a CPU retrain of CatBoost, not a foundation-model fine-tune. The same features are stored for live scoring and for later training, so the model they measure is the model they serve.
Most AI for business decks still assume the prestigious move is to customize the largest model in the building. Sometimes the better move is to refuse. A frozen judge plus a cheap, retrainable head stays swappable when a better vision model shows up. The 499-project check is the hint: fusion helped on both a lighter and a heavier backbone.
There is a cousin of this system already in sellers’ hands, and it should not be mashed together with the 0.94 figure. In A+ Content Manager, modules marked AI Ready can draft text and images from the product listing and from key insights on that product and on top performers in the category. That is generative drafting. Amazon’s seller page reports no quality or sales metric for those drafts. A writer that cannot be graded will fill the catalog. A judge that cannot explain itself will not be trusted by the people who have to reshoot the photo. The interesting stack is both, kept honest about which half has evidence.
The same split matters if you are staffing AI agents to manage listings, ads, or proposals. An agent that rewrites without a standard, and cannot say whether to fix the structure before the hero image, will scale mediocre pages faster than a creative team can review them.
The limits are not footnotes. The authors say these annotations cannot stand in for every category, language, and marketplace. The seller flow most people know is a US-oriented one. A global catalog, they write, will need cheaper inference and a broader label set. Until then, a single accuracy number is a result on a reported holdout, not a passport.
Questions worth asking before you copy the playbook
-
Why not grade each image, caption, and video on its own?
Because the failures that matter (a claim that contradicts the photo, a lifestyle shot sitting where the hero should be) only appear when the page is judged as one snapshot. Isolated checks can catch blur and bad grammar. They miss the argument the page is making.
-
How much did combining the signals actually help?
On the reported holdout, the fused model reached 0.94 accuracy and 0.95 F1, against 0.80 and 0.84 for an equal-weight average of the vision scores, and against 0.89 and 0.90 when the trees saw vision scores alone. Readability alone was weaker than vision alone. Together they beat both.
-
Did better scores turn into more sales?
The authors report a statistically meaningful lift in Ordered Product Sales after a worldwide test that changed visibility of low-quality A+ Content. They do not publish the size of that lift. The separate 8 to 20 percent range is their claim about high-quality A+ Content versus pages without it, not a measured output of this model.
-
Is this the AI sellers already use to draft A+ modules?
No. AI Ready modules in A+ Content Manager generate copy and images from the product listing and from category insights. The system described here judges a rendered page, explains the weak signals, and can optionally generate replacement visuals. Adjacent tools, not one verified stack, and the public generator has no published accuracy figure to borrow.
-
Do you have to fine-tune a giant model to get a useful quality score?
Not in the design Amazon reports. The vision model stayed frozen. A small tree model, retrained on CPU, learned the weights and the interactions. That is cheaper to update when taste and modules drift, and the fusion gain showed up on more than one vision backbone in the 499-project check.
-
If I am not an Amazon seller, what do I copy?
Freeze the expensive judge (the vision model and the rubric) and retrain a small model on labels from your own experts when standards drift. Do not treat a generator as a grader. The half worth copying is the one that can rank the fixes.
What travels outside Amazon is the design: leave the expensive vision model frozen, let a cheap classifier learn the weights from expert labels, and make the explanation specific enough that a team knows whether to restructure the page or reshoot the photo.
Outside the Marketplace write-up
If you want a source that is not Amazon describing its own stack, this is the one that did not fit the body.