AI-Generated Images and SEO: What Google Requires and How to Label Them
Table of contents
AI-generated images and SEO are no longer separate topics: Google now asks you to declare where an image came from, and some platforms will reject a product listing that doesn't. Here is what is required today, how to label a file without exotic tools, and which mistakes cost the most.
AI-generated images and SEO: what Google actually means
Start with what does not happen: Google Images does not penalise an image for being machine-generated. Its published criteria are the same as always: the image should be useful for the query, live on a page with context, be of decent quality and be accessible (alt text, size, correct format). A well-made synthetic illustration competes on equal terms with a stock photo.
What did change is that origin became a mandatory field in two specific places. The first is product listings: Google Merchant Center requires every AI-generated image to carry its origin metadata and explicitly forbids removing it. The second is legal transparency: Article 50 of the EU AI Act has required machine-readable markings on generated or manipulated content since 2 August 2026.
Outside those two cases, declaring origin is an editorial decision. And because almost nobody does it, it makes you look like a serious source at zero technical cost.
Where provenance already shows up
This is no longer theoretical. Google reported in May 2026 that SynthID, its invisible watermark, is applied to more than 100 billion images and videos, and that verification is live in Lens, AI Mode and Circle to Search: a user can simply ask whether an image was generated and get an answer. C2PA Content Credentials verification already works in the Gemini app and is rolling out to Search and Chrome; Google also sits on the C2PA steering committee.
If you publish an unmarked synthetic image and someone checks it, the conversation stops being about your content and becomes about why you didn't disclose it.
The three IPTC values Google names
The standard Google points to is IPTC's, which defined a field called DigitalSourceType to describe how an image was made. For synthetic content, three values keep appearing in the official documentation:
- TrainedAlgorithmicMedia: an image created by a model trained on sampled content. This covers Midjourney, Firefly, Gemini, DALL-E and any diffusion generator. Full URI:
http://cv.iptc.org/newscodes/digitalsourcetype/trainedAlgorithmicMedia. - CompositeSynthetic: a composition mixing synthetic and real elements. A product photo with a generated sky, or an AI background behind a photographed object.
- AlgorithmicMedia: algorithmic output with no trained model behind it. Procedural rendering, data visualisation, generative formulas.
Why TrainedAlgorithmicMedia is the safe default
When in doubt, use it. Any diffusion or image-capable language model lands in that category, and it is the value Google names explicitly in its product feed requirements. Guessing CompositeSynthetic for a fully generated image is a minor error; declaring nothing is the error people notice.
How to label an AI image, step by step
Four steps:
- Keep the metadata the tool already writes. Firefly, Product Studio and Gemini record the source type on export. The most common mistake is not forgetting to label: it is wiping the field while "cleaning" the file with an aggressive optimiser.
- Write it yourself in XMP if it is missing. With exiftool it is a single line:
exiftool -XMP-iptcExt:DigitalSourceType=TrainedAlgorithmicMedia image.png
- Check the data survives export. Some WebP and AVIF converters drop EXIF and XMP. Re-read the final file with
exiftool -a -G1 image.webpand confirm the field is still there before uploading to the CDN. - Declare it on the page too when the visitor needs to know: a short caption under the image is enough. No drama, no warning banners.
What not to touch: credit, copyright and Creator
These are the fields Google Images uses to display image credit. Clearing them to "tidy up the file" removes the attribution that makes your image findable and shareable. Labelling origin adds information; it never replaces authorship. If you are also working on file weight, read the WebP vs AVIF guide first and re-check metadata afterwards, not the other way round.
On-page metadata: schema.org digitalSourceType
schema.org added digitalSourceType as a property of CreativeWork, expecting an IPTCDigitalSourceEnumeration value. In practice you declare it on an ImageObject alongside the rest of the image data:
{
"@context": "https://schema.org",
"@type": "ImageObject",
"contentUrl": "https://example.com/img/blue-sofa.webp",
"caption": "Two-seat sofa upholstered in blue",
"digitalSourceType": "https://cv.iptc.org/newscodes/digitalsourcetype/trainedAlgorithmicMedia"
}
Two warnings worth internalising. First, structured data does not replace file metadata. XMP travels with the image when someone downloads or reuses it; JSON-LD only informs crawlers about your page. They are different layers and they complement each other. Second, this is new ground: according to Google's index, fewer than 1,000 domains used the property as of mid-2026. Don't expect a ranking jump. What you get is consistency and, if your audience uses verification tools, an answer that is already there.
The product feed case (Merchant Center)
Here labelling is not a recommendation, it is a requirement. The current Merchant Center rules for AI-generated content say, in short:
- AI images are accepted in
image_link,additional_image_linkandlifestyle_image_link, provided they keep their IPTC metadata withTrainedAlgorithmicMedia. - Removing or altering those labels before uploading the feed is prohibited.
- AI-generated titles and descriptions go in
structured_titleandstructured_description, withdigital_source_type: trained_algorithmic_media.
The real risk is not paperwork. If you use AI where the product must be accurate, the representation stops matching what ships, and the listing fails on accuracy policy: first the item is disapproved, then the account gets suspended if it repeats.
When AI is a good bet and when it isn't
Good uses: backgrounds and staging, lifestyle imagery, colour variants when the physical item is identical, and campaign material that does not describe the item itself. Bad uses: the main photo of the product you ship, any image that promises a finish the customer will not receive, and an entire food or cosmetics catalogue where detail matters more than styling.
AI Act: obligations since 2 August 2026
Article 50 of the European AI regulation has applied since that date and asks two things of anyone publishing images: machine-readable markings that allow generated or manipulated content to be detected, and explicit disclosure for deepfakes. The Commission guidelines published on 20 July 2026 spell out how that works in practice.
There is a sensible carve-out: content with human review and declared editorial responsibility is treated less rigidly than fully automated output. For a site selling into Europe, the practical reading is simple: if you publish synthetic images regularly, you need a written policy on what gets labelled and how. It does not change your SEO, but it changes how you read to a customer who already knows AI exists.
Mistakes that actually cost rankings or customers
These are the ones that keep showing up when auditing a catalogue or a blog with generated imagery:
- Stripping metadata during compression. The optimiser of the day takes the file to 80 KB and removes the origin: exactly what the platform asked for.
- Using the same "stock" image as two hundred other sites. Same prompt, same result: a visual duplicate that adds no context of its own.
- Alt text stuffed with keywords. Alt text describes the image; keyword padding helps nobody and gets in the way of accessibility.
- Garbled text baked into a hero image. Generated typography usually comes out deformed and it is the first thing a visitor sees.
- AI on the main photo of a physical product. The single biggest source of feed disapprovals in shopping campaigns.
- Declaring nothing and getting caught. The cost is not a penalty: it is the credibility of the next thing you publish.
For your next batch of images, a four-point checklist: origin declared in the file, metadata verified after export, JSON-LD only where the page justifies it, and an honest caption when the image is not a photograph. If you are also working on image visibility, read up on how to appear in Google Images and the fundamentals of image optimisation for SEO. Labelling alone won't lift you; it just stops you from losing ground.