AI image watermarking for eCommerce: what survives the upload

TL;DR. AI image watermarking for eCommerce splits into two layers that behave nothing alike. SynthID sits in the pixels. It survives cropping, compression and re-encoding. C2PA Content Credentials and IPTC tags sit in the file metadata, which is the part most store image pipelines throw away on upload. Amazon now requires an XMP keyword on synthetic-performer images, and EU AI Act Article 50 has applied since August 2, 2026. The rules ask for the fragile layer.

CNBC reported on July 23, 2026 that Amazon told third-party sellers to tag any listing image or A+ content asset showing a photorealistic AI-generated person. The tag is a keyword. Sellers open the file in an IPTC-compatible metadata editor and write contains-synthetic-performer into the XMP field dc:subject before upload. Amazon then shows a disclosure indicator to shoppers.

New York General Business Law section 396-b forced the change. That statute has been in force since June 9, 2026. It sets a $1,000 civil penalty for a first undisclosed synthetic performer in a commercial advertisement and $5,000 for each one after that. The Amazon rule skips images of real people edited with AI tools. It also skips non-photorealistic characters and images with no people in them.

That rule exposes the shape of the wider problem. Every provenance signal on a product image lives in one of two places. It sits in the pixels or it sits in the metadata. eCommerce pipelines handle the two in opposite ways. Pixels survive because pixels are the product. Metadata goes because metadata is weight. Almost every rule now on the books asks for the metadata layer.

AI image watermarking for eCommerce starts in the pixels

SynthID comes from Google DeepMind. The SynthID watermark goes into the image at the moment a model generates it. A shopper sees nothing. The signal is a change to the image data itself and holds up through cropping, filters, frame rate changes and lossy compression.

At I/O on May 19, 2026 Google said SynthID covers more than 100 billion images and videos plus 60,000 years of audio. SynthID verification inside the Gemini app has run 50 million times worldwide. Verification landed in Search that day and in Chrome over the weeks after. A shopper on desktop can right-click a product image and ask whether AI made it. Google also said OpenAI, Kakao and ElevenLabs are building SynthID into their own generation products, so images out of ChatGPT carry it. NVIDIA has applied it to video from its Cosmos models since January 2026.

Two things follow for a seller. Buyers now see this signal on Search and Chrome, the two surfaces where product discovery happens. No export setting changes that. And you cannot write SynthID yourself or check it at catalog scale. The SynthID Detector portal opened to journalists and researchers first and runs a waitlist. Google’s AI Content Detection API launched on the Gemini Enterprise Agent Platform with a small group of trusted partners. You sit downstream of a signal you neither create nor audit, which makes it a poor foundation for a process.

C2PA lives in the metadata, and re-encoding drops it

C2PA Content Credentials work the opposite way. C2PA metadata for product images travels in a signed manifest attached to the file. The manifest records which tool produced the image, what changed, and whether a model played a part. Version 2.4 is the current specification family. The Conformance Program and the official Trust List now carry the trust layer, after the older Interim Trust List froze on January 1, 2026 and stopped taking new entries.

Adoption on the capture side is real. Pixel 10 became the first phone to sign camera captures with Content Credentials in its native camera app. Google said in May 2026 that it extends this to video on Pixel 8, 9 and 10. Meta sits on the C2PA steering committee and plans to label camera-captured media with Content Credentials on Instagram.

Credentials are detailed, human-legible and auditable in a way a pixel watermark never manages. Anyone can also remove them, including by accident. A C2PA manifest sits in a format-specific container, the APP11 segment in a JPEG. Any process that re-encodes the file drops it. That is a normal image pipeline doing its job.

Your own CDN is the compliance problem

Here eCommerce splits from publishing, and most guidance on this topic stops being useful.

Shopify’s image CDN compresses on upload and strips non-essential EXIF and IPTC data to cut file size. Sellers documented the result in the Shopify developer forum in May 2026. An AI-generated image carrying the TrainedAlgorithmicMedia tag on their own disk lost it once cdn.shopify.com served it. WordPress and WooCommerce store the file you uploaded and keep metadata intact. Etsy strips most of it. Different platform, different answer. None of them tell you which one you run on.

Google Merchant Center asks for the exact tag that gets stripped. Google’s policy states that every image created with generative AI must carry the IPTC DigitalSourceType property, and instructs merchants to keep it in place. Three values apply:

TrainedAlgorithmicMedia for an image a model trained on sampled content produced.
CompositeSynthetic for a composite that includes synthetic elements.
AlgorithmicMedia for purely procedural output.

The requirement covers the image_link, additional_image_link and lifestyle_image_link attributes.

Most AI product photography falls under CompositeSynthetic, because the standard workflow puts a real product photo inside a generated scene. Tools also set that value least often by default. So the common failure runs like this. A store never writes the tag, uploads to a platform that would have stripped it anyway, and submits a feed to Merchant Center that fails a policy nobody enforces yet.

The fix is a checkpoint rather than a setting. An image can carry an honest classification at the moment someone generates it. Six months later the same job means reconstructing which of 4,000 SKUs used a synthetic background. That holds whether you generate in Vision or anywhere else. It is a workflow decision.

What the law asks a seller to do

Article 50 of the EU AI Act has applied since August 2, 2026. That date survived the Digital Omnibus, which the European Parliament approved on June 16, 2026 and the Council adopted on June 29, 2026. The Omnibus pushed standalone high-risk obligations under Annex III to December 2, 2027 and product-embedded systems under Annex I to August 2, 2028. Article 50 stayed where it was. Penalties for breaching it reach 15 million EUR or 3 percent of global annual turnover.

Read who Article 50 addresses. Article 50(2) obliges providers of generative AI systems to mark their outputs in a machine-readable format. Systems already on the market get a transition to December 2, 2026. Article 50(4) obliges deployers to disclose deepfake content in a clear and distinguishable way. Article 3(60) defines a deepfake as content resembling existing persons, objects, places or events that would falsely appear authentic.

For a catalog that draws a line. A real product photograph placed in a generated scene sits far from the deepfake definition, and the machine-readable marking duty sits with the tool vendor. A generated human face that reads as a real person sits inside it. Marketplace policy and ordinary consumer-protection law reach wider than Article 50 does on both counts. Treat the split as a starting point for a conversation with counsel rather than a ruling. Our 2026 regulations overview works through the jurisdiction map in more detail.

California is the one to watch next, because it attacks the stripping problem head on. AB 853 moved the California AI Transparency Act’s operative date to August 2, 2026 and added a duty from January 1, 2027. Large online platforms must detect and display provenance data on the content they distribute. They must also keep standards-compliant provenance data and digital signatures in place where that is technically feasible. Penalties run at $5,000 per violation and each day counts as its own violation. Marketplaces above the threshold will have to stop doing what image CDNs do today. That is the first regulatory pressure aimed at the pipeline instead of the seller.

What to do now

Classify at the point of generation. Whoever produces the image records what it is in the same step, using one of the three IPTC values. A shared field in your asset workflow covers it. Retroactive classification of a live catalog is the expensive version of this task and nobody ever finishes it.

Test one file end to end, and take it from the live page. Download a listing image from the product page URL rather than from your drive, then inspect it. exiftool on the command line or the Inspect tool in Adobe’s Content Authenticity app answers within seconds whether the tag survived. Assume it did not if you have never run this check.

Own the layer you control. You cannot write SynthID and your platform may drop C2PA, and you can still write IPTC tags before upload and keep an internal record of which SKUs use AI-assisted imagery. That record is what a marketplace appeal or a regulator letter asks for. The Amazon synthetic performer tag goes in before upload too, which our Amazon guide covers alongside the rest of the marketplace’s image rules.