If your ecommerce pipeline generates apparel images with AI and then runs them through an optimizer, there's a decent chance you're silently violating Google Merchant Center's image policy — and hurting your feed without knowing it.
The metadata requirement almost nobody checks
As of July 2026, Google Merchant Center requires AI-generated product images to retain IPTC DigitalSourceType metadata, and explicitly tells merchants not to strip embedded tags. The relevant values are TrainedAlgorithmicMedia (generated from a trained model), CompositeSynthetic (composites with synthetic elements), and AlgorithmicMedia (purely algorithmic).
Here's the problem: most image-optimization steps — sharp, imagemin, Pillow's save(), most CDN transforms — strip metadata by default. Your generation tool writes the tag; your build step deletes it.
A quick verification script:
Turning “does it look right” into a testable contract
The second engineering problem with AI apparel images: QA is subjective. “Looks good” isn't a test. We ended up encoding per-SKU review rules as a schema, so the checkable parts are automated and humans only review what machines can't judge:
Color drift is automatable (Delta-E against the source swatch). Button counts are borderline (a small vision model does okay). Print alignment across body curves and seams is still a human call — the failure modes are too varied, and a misaligned plaid is exactly the kind of error that looks fine at thumbnail size and generates returns at delivery.
The two-size review rule
Every image gets reviewed twice: once at full size (texture, seams, hardware), once at mobile-thumbnail size (what buyers actually see). An image that only communicates when zoomed in is not a listing image.
This is the engineering half. The seller-facing half — a garment-type difficulty matrix and gallery-slot design — is documented in the full
guide:https://loomadesign.ai/en/blog/ai-virtual-try-on-clothing-ecommerce
Top comments (0)