Google Merchant Center AI image metadata: the IPTC tag explained (2026)

Google Merchant Center requires images created using generative AI to carry IPTC DigitalSourceType metadata identifying their AI origin, and its dedicated guidance names TrainedAlgorithmicMedia. Keep that information in the actual image file Google fetches, including after resizing, compression or delivery through another service.

Which Google product images need AI-source metadata?

Google’s AI-generated content guidance says all images created using generative AI must contain the IPTC DigitalSourceType TrainedAlgorithmicMedia tag. The page identifies three relevant product-data attributes: image_link, additional_image_link and lifestyle_image_link.

This is not only a main-photo issue or a model-photo issue. A generated lifestyle scene containing a product but no people still needs its origin considered. Moving a generated image into a supporting position does not make the source information irrelevant.

Start by documenting how each image was made. Keep the original capture, the generated or edited version, and the destination export separately. This gives you a basis for selecting accurate metadata rather than guessing from the filename or visual style.

If you distribute the same images elsewhere, check each destination separately. The marketplace image and AI rules guide distinguishes source metadata, people disclosure and platform-specific image rules.

What is IPTC DigitalSourceType?

DigitalSourceType records the kind of process used to create an image. It is structured metadata, not a visible badge added over the photo.

The IPTC definition of trainedAlgorithmicMedia describes media created with an AI model trained on captured content. Its human-readable name is Created using Generative AI. The standard identifier is:

http://cv.iptc.org/newscodes/digitalsourcetype/trainedAlgorithmicMedia

Use the identifier as specified, rather than inventing a friendly value such as AI photo. In ExifTool, the writable field is named XMP-iptcExt:DigitalSourceType. The shortened ExifTool group name and the underlying IPTC namespace describe the same field; they are not two separate tags to add.

The tag says something about the image’s source. It does not say that the product is accurate, that an image has been approved for Shopping, or that its history has a verified digital signature.

Is edited-with-AI metadata the same as Google’s named value?

No. IPTC’s vocabulary is broader than the particular instruction on Google’s dedicated help page.

IPTC also defines compositeWithTrainedAlgorithmicMedia for media augmented, corrected or enhanced with generative AI, including inpainting and outpainting. That is a real vocabulary entry, but the Google page reviewed for this article does not explicitly name it as a substitute for its stated generative-image value.

Preserve accurate existing source information. If an image arrives with a different classification, establish how it was produced before replacing the field. When a mixed original-and-generated workflow creates uncertainty, consult the current Google instruction rather than treating a vocabulary label as an approval certificate.

This distinction also applies to editor controls. A tool can offer an accurate IPTC classification without promising that every platform treats every classification identically.

Does Amazon’s synthetic-performer keyword cover this requirement?

No. Amazon’s announcement about generated people asks for contains-synthetic-performer in XMP dc:subject for photorealistic AI-generated people. Google’s instruction addresses the image’s AI source.

Consider a product on a generated kitchen counter with nobody in the scene. Amazon’s people-focused condition does not require its keyword solely because that background was generated. Google’s generative-image rule still matters.

A realistic generated model wearing a garment can need both fields for the destinations concerned. Keep an existing Amazon keyword when adding Google’s source type, and keep the source type when adding the Amazon keyword. Do not use one as a shorthand for the other.

How do I add or inspect the source type in my browser?

The AI image metadata tagger reads and writes the relevant fields in JPEG and PNG without uploading your images.

For a qualifying image created using generative AI, select Made with generative AI. The separate checkbox, Shows a photorealistic AI-generated person, controls Amazon’s keyword; select it only where that description applies.

Choose your files and compare the Tags now and After tagging columns. Download the result, then choose that downloaded file again to read what it now contains. The displayed future state is useful, but reopening the download checks the file you actually saved.

Keep a backup before writing. The browser tool replaces existing XMP, so unrelated subject keywords, captions or rights information inside that packet are not merged into the new one. It is a focused AI-tagging utility, not a general metadata-preservation editor.

If your file already carries an Amazon people tag, make the checkbox reflect that applicable disclosure before downloading. The tool recalculates the output from the options you select; it does not promise to preserve every prior XMP value automatically.

If Tags now shows a source category that is not available among the radio options, keep the original and use a workflow that retains that field. Do not select the nearest-looking option merely to obtain a fresh download.

The Edited with generative AI option writes IPTC’s separate edited-image value. Its presence in the interface should not be read as a claim that Google expressly lists that value as interchangeable with TrainedAlgorithmicMedia.

How do I write the field using ExifTool?

Use a copy of a confirmed generative-AI image and inspect its existing metadata first:

exiftool -G1 -a -s -XMP-iptcExt:DigitalSourceType -XMP-dc:Subject product-final.jpg

For the appropriate created-with-generative-AI source, write:

exiftool "-XMP-iptcExt:DigitalSourceType=http://cv.iptc.org/newscodes/digitalsourcetype/trainedAlgorithmicMedia" product-final.jpg

Then run the read command again. The official ExifTool XMP table documents DigitalSourceType in its XMP-iptcExt group. Naming that field explicitly avoids changing dc:subject as part of the command.

This assignment replaces the existing source-type value, so establish that the change is correct before running it. It is not a command to apply blindly to every photograph in a catalog.

According to ExifTool’s application documentation, normal writing retains an _original backup. Keep it until the edited file has been checked. Do not add overwrite-original or delete-all-metadata options to these examples. Work on one representative final export before making the same change across an appropriate batch.

Why do AI tags disappear after compression or conversion?

A new export can contain newly encoded pixels without the earlier metadata packet. The visual result can appear unchanged while its source information is missing.

That makes the order of work important: finish visual edits, create the destination format and dimensions, add or retain the required metadata, and inspect the final output. If another transformation happens afterward, inspect that newer file too.

CatalogCut’s target-size compressor re-encodes its JPEG or WebP outputs and drops original metadata, including AI tags. Its page states that limitation. When you need CatalogCut’s browser tagger afterward, choose JPEG because the tagger does not write WebP.

The image optimizer has different behavior: it preserves recognized AI-disclosure tags in JPEG and PNG while removing private metadata. Its WebP output removes metadata. Do not assume that support for WebP as an image format means every workflow preserves AI disclosure inside it.

Should I use JPEG, PNG or WebP for a tagged final?

Google’s image_link specification lists all three among its supported formats. The practical choice also depends on the tools between your editor and the URL submitted to Google.

For CatalogCut’s browser tagging workflow, JPEG and PNG are the writable choices. WebP can be inspected for existing AI tags, but the tool cannot create a newly tagged WebP download. Convert before tagging if you need that workflow, then verify the converted file.

Do not rename a WebP extension to JPG. A filename change does not convert the file or create a metadata field. Google also says the extension in the image URL should correspond to the image format.

Choose a format that your complete editing, hosting and publication path can handle correctly. An image’s nominal format is only one part of preserving its source information.

How do I check the file Google will actually fetch?

Checking your local export is the first step. For a feed image, also inspect the image served at the submitted URL after your store or image delivery service has processed it.

A practical review sequence is:

1. Save the finished local image and read its source metadata.
2. Upload it through the normal store or hosting workflow.
3. Retrieve the image from the precise product-data URL.
4. Read that retrieved file’s DigitalSourceType and any other required disclosure.
5. Compare its format, dimensions and visible product details with the intended export.

If the served version lacks the tag, investigate the upload transformation or delivery settings. Do not keep retagging a local file while the serving path continues to replace it.

Google recommends stable image URLs when the image stays the same and provides separate guidance for notifying it about changed images. Follow that guidance when replacing an existing asset; do not invent a guaranteed recrawl time or approval outcome from a successful metadata check.

Are Stage downloads ready for this workflow?

CatalogCut Stage PNG and JPEG downloads come pre-tagged with the generative-AI source type, and on-model downloads also carry Amazon’s synthetic-performer keyword. This includes sized downloads and campaign ZIP images.

How Stage works explains the separate product-fidelity check: each image is compared with the original before delivery, with credits refunded for failed checks. Sellers still need to compare details themselves before publication.

Keep the tags when making subsequent exports, and verify the hosted version if your store transforms the download. Pre-tagging the original download does not control another service’s later processing.

Does the metadata guarantee Google approval?

No. Google’s image requirements still address the actual product, image content and technical suitability. Source metadata cannot correct an inaccurate garment, missing accessories, promotional overlays or the wrong variant.

This metadata is also not a signed C2PA credential or proof of an unbroken edit history. The guidance discussed here concerns source disclosure; do not make a broader claim about legal compliance or authenticity from an editable XMP field.

What to do next

Inspect one final JPEG or PNG in the AI image metadata tagger, preserve the appropriate source information, then check the file served at your product-image URL.