As of 8 November 2026, the way marketplaces judge a product page has changed shape. An AWS machine learning write-up published on 5 November describes a content-quality system for Amazon A+ Content that reached 0.94 accuracy and 0.95 F1 on a 20 percent holdout of 5,478 expert-labelled pages, against a 0.80 accuracy and 0.84 F1 baseline from simply averaging model scores. The system renders each full page as one snapshot, scores it on a four-dimension rubric, adds four OCR readability signals, and lets a small classifier decide. Sellers raised their quality score by 1.4 points out of 10 after revisions.
A page is scored as a whole, not asset by asset
The argument behind scoring a whole page is that checking assets one at a time misses the failure. A blender page can have a sharp hero image, grammatical bullets and a well-lit lifestyle photo and still fail, because the bullets advertise a 64-ounce pitcher while the photo shows a single-serve model, or because the lifestyle shot sits above the hero image. Those problems only appear when the page is read as a unit, which is why the system renders the entire experience into a single snapshot instead of grading components separately.
For sellers, that has a practical consequence: a page assembled from individually good parts can still score badly. Teams that optimise images, bullets and A+ modules in separate workstreams, each hitting its own checklist, routinely ship pages that contradict themselves. The score is not measuring effort, it is measuring coherence.
There is also a product concept problem hiding inside the same rubric. A listing is usually written once, at launch, and then left alone while the product concept moves on: a reformulated ingredient, a new pack size, a flavor extension. The page keeps describing the original product concept, so feature presentation decays silently long before anyone books a design review.
What the rubric actually rewards
| Dimension | What it checks | Common failure | Fix that moves the score |
|---|---|---|---|
| Visual storytelling | Images build a sequence | Lifestyle shot above the hero | Reorder to hero, use, detail |
| Feature presentation | Claims match the product | Bullets describe another variant | Rewrite bullets against the SKU |
| A+ layout | Modules read in order | Comparison table buried at the end | Move comparison up front |
| Conversion optimisation | Page answers buying questions | No sizing or compatibility data | Add structured specs |
The rubric is useful precisely because none of its four dimensions is about persuasion. Visual storytelling checks whether the images form a sequence, not whether they look premium. Feature presentation checks whether claims match the product, not whether the copy is clever. A+ layout checks reading order. Conversion optimisation checks whether the page answers the questions a buyer would ask before paying. A seller who treats the score as a copywriting contest will keep losing to a seller who treats it as a consistency audit.
Consistency between bullets and images
The most common score killer is a mismatch between what the text promises and what the image shows. It happens because listings are updated piecemeal: a supplier changes the pack size, the bullets get edited, the photography does not. The system cannot know which is right, so it penalises the page. A quarterly check that reads the page as a shopper would, image by image against claim by claim, catches most of these before a revision cycle does.
Readability is measured, not assumed
Four OCR signals are computed from the rendered page: text density, character density, spacing and overall readability. This matters for mobile, where dense image text becomes unreadable and the page loses the buyer in the first scroll. Sellers who pack five benefits into one infographic often improve their design score in internal reviews while lowering the readability signal. The two have to be balanced, and the balance is measurable rather than a matter of taste.
How to use the score without overfitting
A quality score is a diagnostic, not a target. Optimising for the number alone produces pages that satisfy the rubric and still convert poorly, because the rubric does not know your category's objections. The useful pattern is to let the score point at the module that is failing, then fix the underlying product information, then check whether conversion moved. Vendor figures claim high-quality A+ content lifts sales 8 to 20 percent and can lift conversion by up to 15 percent, but those results have not been independently verified, so treat them as a direction rather than a forecast.
A four-step revision loop
The reason to fix facts before design is that most low scores come from information problems wearing a design costume. A comparison table that omits the variant a shopper wants looks like a layout issue and is actually a data issue. Once the data is right, layout changes become cheap; before that, they are decoration. This is also why the machine-scored approach is worth copying internally even if you never see the official score: it forces the page to be evaluated as one artefact.
FAQ
What does the score actually measure?
It measures whether a rendered page is internally consistent and readable across four dimensions, plus four OCR readability signals. It does not measure persuasion or brand taste.
Can a seller see this score directly?
Not necessarily. The published system is described as an internal content-quality model; sellers can approximate it by auditing their own pages against the same rubric.
Why do individually good assets still score badly?
Because the page is judged as a whole. Bullets describing one variant while images show another is a coherence failure that no per-asset checklist catches.
Should sellers optimise for the score?
Use it as a diagnostic. Fix the lowest-scoring module, correct the underlying product facts, then verify against conversion and returns rather than against the score alone.
