What Merchant Center requires
The Merchant Center help page on AI-generated content states the image rule briefly: every image created with generative AI must contain metadata saying so — the IPTC DigitalSourceType tag with the value TrainedAlgorithmicMedia. Google explicitly says not to remove such tags from files produced by AI tools. The same page lists where AI images are allowed: the main image (image_link), additional images (additional_image_link) and the in-context or lifestyle image (lifestyle_image_link).
So there is no separate feed field here, unlike AI descriptions. For text Google requires the structured_description attribute — see our piece on AI product descriptions. For an image, the label lives inside the file the feed links to.
The second half of the requirements is the same for AI and ordinary photos. According to the image_link attribute page, the main image must clearly show the exact item being sold. Images get disapproved for:
- text over the photo: "Buy" calls to action, prices, "Free shipping", words like "Best" or "New";
- watermarks, barcodes, QR codes, added shop logos;
- a border, or a single-colour square instead of the product;
- items not sold together with the product;
- placeholders, mockups and illustrations that do not depict the physical product.
For products with variants Google recommends a separate image for each colour or material, and it must accurately reflect the colour, pattern and finish. The product should fill 75% to 90% of the frame. The minimum size is 500×500 pixels, and from 31 January 2027 that becomes mandatory for all products; 1500×1500 or more is recommended.
Which images you may generate
The rules boil down to a few questions worth asking of every image before it goes into the feed.
- Was the image drawn from a text description, with no photo of the real item?The model draws a plausible object, not yours. Dimensions, seams, hardware and shade will be invented. Yes → not for the main image: it is an illustration that does not depict the physical product
- Was a colour or material changed by generation, for a variant you have not photographed?Google requires a variant image to reflect its colour, pattern and finish accurately. A product recoloured by AI may differ from what the buyer receives. Yes → photograph each variant separately
- Did the generator add text, a logo, a border or another object to the frame?Models often add lettering and "decor". For Merchant Center these are the same violations as in ordinary photos. Yes → regenerate or remove the extras
- A real photo of the item, with AI replacing the background or placing it in a room?The product is unchanged; only its surroundings are new. Yes → fine as main, additional or lifestyle image, with the label in the file
- Does the file the feed links to carry the IPTC DigitalSourceType label?Check the file your site serves, not the one from the generator. No → restore the tag, otherwise the image breaks the rules
One gap in Google's own help. For a photo where AI changed only part of the frame, the IPTC vocabulary has a separate value, compositeWithTrainedAlgorithmicMedia — "edited using generative AI", including inpainting and outpainting. Google does not mention it: the help names TrainedAlgorithmicMedia, CompositeSynthetic and AlgorithmicMedia. The requirement itself is broad — a label on every image created using generative AI. So do not leave a photo with a replaced background unlabelled; which of the two values to choose, the help does not say directly.
Where the label disappears
The IPTC tag is written into the file's metadata, in the XMP block. The generator or editor set it, but between it and the feed the image is usually compressed, resized or converted to WebP. On 15 October we checked which common tools keep the label. We took a JPEG with trainedAlgorithmicMedia in its XMP and ran it through each tool with default settings.
Test on 15 October 2026 with one 1536×1024 JPEG carrying IPTC DigitalSourceType = trainedAlgorithmicMedia in XMP. The label was detected by searching the output file for the string trainedAlgorithmicMedia.
Pillow is used by many home-grown scripts and feed export modules, and cwebp sits inside many tools that convert images to WebP. Both drop the label by default. For WordPress we only read the code: the Imagick-based image editor keeps the iptc and xmp profiles when it creates resized copies (that is what the core code says, in the strip_meta function). We did not test it on a live site, and servers without Imagick process images with a different editor.
The practical conclusion: a label set by the generator does not mean a label in the feed. An optimisation plugin, a CDN that compresses images, or a module that makes square thumbnails for Merchant Center can remove it, and diagnostics will not tell you straight away.
How to check your feed
- Take a few
image_linkURLs of AI images from the feed. The exact addresses Google sees, not the files in your media library. - Check the label in the file at that URL. The simplest way is a terminal:
curl -s <image URL> | grep -a -c trainedAlgorithmicMedia. A result of 1 or more means the tag is there, 0 means it is gone. This only searches for a string and does not validate the metadata, but it is enough to catch a stripped label. - If the label is missing, find what strips it. Compare the file right after generation with what the site serves. Usually one setting is to blame: metadata retention in the optimisation plugin or a parameter in the conversion script.
- If your generator does not set the label, add it yourself: the XMP field
Iptc4xmpExt:DigitalSourceTypewith the valuehttp://cv.iptc.org/newscodes/digitalsourcetype/trainedAlgorithmicMedia. Metadata editors and the exiftool utility can do this. - Look at every AI image with your own eyes. Text, logos, stray objects, the wrong colour — the model adds these itself, and no automated check will see them.
Where AI genuinely saves a shop time is covered in AI for an online shop, and Google's requirements for AI text on the site in our piece on the October 2026 changes.
What generation costs
Money is the smallest part of the cost. In OpenAI's calculator for the GPT Image 2.5 models, one 1024×1024 image costs from $0.006 at the lowest quality to $0.21 at the highest. That covers image output tokens only; the prompt text and any input photo, if you edit a real shot, are billed separately.
Image output tokens at $30 per million, token counts from OpenAI's calculator for GPT Image 2.5. NBU rate on 15 October 2026: UAH 44.9801 per dollar. Excludes prompt text and input images.
Even at maximum quality a thousand images cost about UAH 9,500, and about UAH 2,400 at high quality. So the comparison with a photo shoot should not be the price of a file but the price of a usable image:
price of a usable AI image = (price of one generation × number of attempts) + minutes of review × the cost of a minute of the person reviewing.
Generation wins when the product is already photographed and you only need a background, a room or a set of catalogue shots in one style. It does not replace a shoot when there are no real photos at all: then the AI draws something other than your product, and under Google's rules that is not fit for the main image, however cheap the file.
Our position
Generate the surroundings, not the product. A real photo of every variant plus an AI background or room is a safe and cheap way to get consistent catalogue images. And build the label check into your process the way you check image size: the file at the feed URL, not the file in the generator.
This does not apply if you do not send products to Merchant Center and do not plan to. Then the IPTC requirement does not formally concern you. But the rule "the photo must show the exact item" remains common sense: a buyer who receives something other than what was pictured sends the parcel back.
Checked on 15 October 2026 against the Google Merchant Center help pages (AI-generated content, the image_link attribute), the IPTC Digital Source Type vocabulary, OpenAI's image generation pricing calculator and the WordPress core code (class-wp-image-editor-imagick.php); exchange rate from the NBU. The label retention test is our own, on one file, with the tool versions listed in the table.