Google Shopping Requires AI Image Metadata. Here’s What Sellers Must Preserve

An AI product image can look clean, accurate and ready for Google Shopping while still missing a requirement that shoppers cannot see. Google Merchant Center says images created with generative AI must contain machine-readable metadata that identifies their digital source.

This is not a ban on AI imagery. Google explicitly allows AI-generated assets in the main image link, additional image links and lifestyle image links. The hidden risk is losing the embedded provenance information somewhere between generation, editing, export, upload and feed submission.

For sellers using AI backgrounds, virtual try-on, colour variants or lifestyle scenes, metadata now needs to be part of the image-production checklist, not an afterthought.

The rule is about the file, not just the pixels

Google requires generative AI images to carry an IPTC DigitalSourceType value. IPTC is a widely used photo-metadata standard, and Digital Source Type records how an image came into existence.

Google's guidance names several values that should be preserved. TrainedAlgorithmicMedia identifies media created using a generative AI model trained on captured content. CompositeSynthetic covers a composite that includes synthetic elements. AlgorithmicMedia is for media produced by an algorithm that is not based on sampled training data, such as a mathematically generated pattern.

Most seller workflows will involve the first two. A fully generated product scene usually fits TrainedAlgorithmicMedia. A real product photograph placed into an AI-generated background is a mixed image, for which IPTC recommends CompositeSynthetic. The important point is to use the value that truthfully describes the file rather than treating every AI-related image as identical.

Why normal export habits can create a blind spot

Metadata is stored inside the image file, separate from the visible pixels. IPTC explains that Digital Source Type is carried in the XMP portion of the file. A seller can therefore open two visually identical JPEGs and have one contain the correct source information while the other does not.

That makes the full delivery chain relevant. An AI tool may embed the correct metadata in its download, but the image can then pass through a photo editor, compressor, marketplace app, content-management system or image CDN before Google fetches it. You should not assume the final hosted version still contains the same information simply because the original download did.

Google specifically warns merchants not to remove embedded DigitalSourceType metadata from images made with tools such as Product Studio. The practical interpretation is simple: preserve the original file, and verify the finished file after the last transformation in your workflow.

A five-step metadata check for product images

First, keep the original AI output as a master. Do not overwrite it when resizing, compressing or preparing channel-specific versions. If a later export loses metadata, the master gives you a known starting point.

Second, record what kind of image you created. Was the entire frame generated, or was a genuine product cutout combined with a synthetic background? This determines whether the final asset is fully generative or a composite containing synthetic elements.

Third, inspect the exported production file. IPTC provides a Get Photo Metadata tool for reading embedded metadata from an uploaded file or a web image. Look for Digital Source Type and confirm that the value matches the image's origin. Technical teams can also inspect the XMP field named Iptc4xmpExt:DigitalSourceType with metadata software such as ExifTool.

Fourth, upload the image through your normal store or feed workflow, then inspect the exact public image URL that Merchant Center will crawl. Checking only the local file is not enough if another service reprocesses the image during upload. The hosted file is the version that matters to the feed.

Fifth, check the product itself again. Metadata does not make an inaccurate image compliant. Google still expects each variant image to show the correct colour, finish and material, and it tells sellers of personalised products to show the customisation rather than a blank substitute. For jewellery, verify the engraving, stone count, chain, clasp and metal finish before submission.

Visible AI labels and embedded metadata are different controls

Merchant Center also lets sellers mark an image or video as created or edited with AI through its Manage AI label control. That setting can create disclosures in Google's ad experiences. Google also says its own AI tools add machine-readable signals such as SynthID and C2PA information.

These controls sit alongside the IPTC image requirement. A visible or interface-level AI label tells a shopper or advertising system that AI was involved. Embedded DigitalSourceType metadata travels inside the asset and describes its source in a standard machine-readable field. Sellers should not assume that adding a visible label repairs a file whose required embedded metadata is missing.

There is another separate rule for AI-written product data. Google says AI-generated titles should use the structured title attribute with a digital source type, while AI-generated descriptions should use the structured description attribute. Image metadata does not replace those feed fields, and those text fields do not replace the image tag.

What to do with Product Studio and external AI tools

If you generate an image in Google Product Studio, keep the metadata Google embeds. Be cautious when downloading it, editing it elsewhere and exporting a new copy. Recheck the final version before it returns to Merchant Center.

For images created in another AI service, inspect the first download rather than assuming the service writes the IPTC field. If the value is missing, use a metadata-capable tool or update the workflow that produces the file. Adding provenance once at the final production stage is better than manually repairing random catalogue images after diagnostics appear.

For high-volume catalogues, make the check automatic. A pre-publish script can scan every image for DigitalSourceType, flag missing values and compare the expected source type with the file. That is especially useful when employees or agencies use several image generators and export tools.

Build provenance into catalogue QA

The useful lesson is broader than one metadata field. AI product-image workflows now have two quality layers. The visible layer asks whether the product is accurate, attractive and compliant with image rules. The invisible layer asks whether the file carries the provenance information expected by the platform.

A strong process checks both before publication: preserve the master, classify the edit, inspect the production export, inspect the hosted URL and then confirm the actual SKU is still correct. That is a small addition to the workflow, but it can prevent a technically polished AI image from becoming a catalogue problem.

Sources: Google Merchant Center Help, “AI-generated content” and “Image link [image_link],” reviewed 30 September 2026; Google Merchant Center Help, “Use AI content label settings and disclosures,” reviewed 30 September 2026; IPTC Photo Metadata User Guide, November 2025 revision; IPTC Digital Source Type controlled vocabulary, reviewed 30 September 2026.