feedsmerchant-centerai-images

AI-generated product images in Merchant Center: preserve provenance and label regulated assets

Preserve provenance metadata in AI-created Merchant Center images, label regulated creatives, and verify every delivered asset file.

The short version

Google’s Merchant Center specification says product images created with generative AI must retain embedded provenance metadata. It also says certain AI-generated or edited advertising assets need disclosures or labels under rules applying in the European Union, India and New York.

Do not let image optimizers, CDNs or export jobs strip the metadata. Inventory AI-created assets, preserve their IPTC digital-source tags, use Google’s AI-label setting where it applies, and verify the actual files delivered from every product-image URL.

The requirement applies to the image file

For both image_link and additional_image_link, Google says generative-AI images must contain metadata indicating how they were created. The specification names these IPTC DigitalSourceType values:

  • TrainedAlgorithmicMedia for images made with a model derived from sampled content;
  • CompositeSynthetic for composites containing synthetic elements;
  • AlgorithmicMedia for images produced by an algorithm that is not based on sampled training data.

Google explicitly says not to remove embedded tags from AI-created images, including images produced with Product Studio.

This is different from the structured feed fields used for AI-written text. structured_title and structured_description can carry digital_source_type for generated copy. Image provenance lives inside the image binary, not in a replacement feed attribute.

Find every place metadata can disappear

An original file can contain the right tag while the image served to Google does not. Check the complete delivery chain:

  1. the original asset from the studio or generation tool;
  2. the digital-asset manager;
  3. ecommerce-platform ingestion;
  4. resizing, cropping and format conversion;
  5. CDN optimization and caching;
  6. the final URLs submitted in image_link and additional_image_link.

Common image-processing libraries remove metadata to reduce file size. A new WebP, JPEG or PNG rendition may therefore lose provenance even when the source file is correct.

Build a register with the product ID, image role, source file, generation or editing method, expected DigitalSourceType, submitted URL and last verification date. Include every additional image, not only the main product image.

Preserve provenance through the pipeline

Use this deployment sequence:

  1. Classify each AI-created or AI-edited asset before it enters the catalogue.
  2. Embed the appropriate IPTC digital-source value in the master file.
  3. Configure every derivative-image job to copy supported provenance metadata.
  4. Purge or version CDN URLs after replacing a file so Google can retrieve the corrected asset.
  5. Download the public rendition from the exact submitted URL.
  6. Inspect that downloaded binary for the expected metadata.
  7. Save the verification result beside the product and asset record.

Do not use a successful feed upload as proof. A feed validator usually sees an image URL; it may not inspect the remote binary’s IPTC block. The acceptance test is the file Google can fetch.

Treat ad labels as a separate control

The same Google specification says ads with certain AI-generated or edited assets need disclosures or labels under rules applying in the European Union, India and New York. Google provides an AI-label setting across products including Merchant Center and Google Ads, and merchants can also add labels directly to creatives.

That control does not replace embedded image provenance. Run two checks:

  • File provenance: does the delivered image retain the correct digital-source metadata?
  • Creative disclosure: does the ad or asset carry the label required for the market and use case?

Google warns that using its AI-label setting does not guarantee compliance with local law. Assign legal or policy review for the markets where the asset will run rather than treating a platform toggle as final approval.

Verify a representative catalogue sample

Start with all known AI-created assets, then sample the rest of the catalogue to catch undocumented use. For each selected product:

  • fetch the main and additional images from their public feed URLs;
  • confirm the response is the intended product image, not a placeholder;
  • inspect the binary for DigitalSourceType and its value;
  • compare the value with the asset register;
  • check the market-specific creative label where the image is used in ads;
  • repeat after any resize, format-conversion or CDN configuration change.

Use the free product feed scanner for feed fields and URL-level problems, but keep binary metadata verification as a separate test. A clean feed cannot prove that an optimizer preserved IPTC metadata.

Make the control repeatable

Add provenance checks to the image-pipeline release process rather than running a one-off cleanup. Fail the release when a known AI-created asset reaches its public URL without the expected metadata. Re-run the sample after changing image software, CDN settings, file formats or asset-management vendors.

The finished control should prove four things: the asset was classified, the master carried the right provenance, every public derivative preserved it, and the applicable creative label was reviewed for the markets where the asset runs.

Sources