The AI Photo Is Allowed. The Missing Tag Is Not.

A photographer costs money and a light tent costs a Saturday. So the generated product image arrived fast in small shops, and it arrived without any of the paperwork that a stock photo used to bring with it. The question sellers ask is whether it is allowed. That turns out to be the easy half.

The harder half is that the rules which do exist are not about permission. They are about labelling, and they attach to the file rather than to the listing page. A file loses its labels easily. That is the whole story of this article.

Are AI product photos allowed at all?

Broadly yes, on the major sales channels, provided the image represents the actual product accurately. No large platform bans generated imagery outright, several of them now sell you a generator themselves, and the general purpose models keep getting better at holding your actual item steady, which is the part of the ChatGPT Images 2.5 release that matters to a catalogue. The restrictions that bite are the ordinary ones about what an image may show.

Google's image requirements for a product feed are a good proxy for what most channels expect, because they are written down in detail. The image has to display the entire product accurately and show the correct variant, matching colour, pattern and material. The product must fill no less than 75 percent and no more than 90 percent of the frame. Watermarks, borders, promotional adjectives, calls to action, pricing and placeholder graphics are all refused. The minimum is 500 by 500 pixels, with 1500 by 1500 recommended.

Read that list again with a generated image in mind. Nothing there mentions how the picture was made. Every clause is about whether it misleads. A generated shot that shows your actual product in your actual colourway passes. A generated shot with a slightly nicer weave than the fabric you ship fails, and it fails for the same reason an over edited photograph fails. The same gap opens up between a mockup and a real garment, which is where generated designs come apart at the print file.

What exactly does Google require you to tag?

Two things, and the second one surprises people. Images made with generative AI must carry IPTC metadata declaring it, and text made with generative AI has to be submitted through separate feed attributes.

On the image side, the Merchant Center policy on AI generated content names three DigitalSourceType values. TrainedAlgorithmicMedia covers an image created with a model derived from sampled content, which is what a normal image generator produces. CompositeSynthetic covers an image that mixes synthetic elements with real ones. AlgorithmicMedia covers output produced purely by an algorithm with no sampled training data behind it, which is rare in product photography. The requirement applies across image_link, additional_image_link and lifestyle_image_link, so gallery and lifestyle shots are in scope alongside the main image.

On the text side, an AI written title goes in structured_title, with a digital_source_type sub attribute set to trained_algorithmic_media and the wording in a content sub attribute. Descriptions follow the same pattern through structured_description. If you have ever generated 300 product descriptions in an afternoon and pushed them straight into the title and description fields, that is the gap. It is not enforced with the visibility of an image rejection, which is exactly why it goes unnoticed.

Why the tag disappears without anybody deciding to remove it

Here is the part that costs shops their compliance without a single bad intention. Metadata is not part of the picture. It travels alongside the pixels in the file container, and any step that re-encodes the file can drop it.

The IPTC, which defines these fields, published a reminder that Google asks publishers and merchants to preserve the tags, and it named inventory and content tools as frequent offenders. Its note on metadata stripping is blunt about a habit that predates AI entirely: web pipelines have been discarding EXIF and IPTC data for years to shave kilobytes, and nobody complained because nothing depended on it. Now something does.

Walk your own path once. The generator writes the tag. Your download may or may not preserve it. Your image editor probably drops it on export unless told otherwise. Your store platform generates half a dozen resized variants for thumbnails and responsive layouts, and those derivatives are new files. Your feed then points at whichever derivative the platform decided to serve. Four opportunities to lose a field you never see.

Which of your images even count as generated?

More than the obvious ones, and this is where the specification is more useful than the summaries. The distinction is not between real and fake. It is between what the file was at its inception and what was done to it afterwards.

The third row is the one small shops land on most often, because it is the cheapest useful trick in the box: shoot the item once on a white sweep, then place it on a marble counter, a picnic blanket or a shelf that suits the season. That is a composite, and it has its own label.

Does the EU AI Act make you label your product photos?

Not directly, and the confusion here is worth clearing up because it has been sold as a compliance emergency to people who do not have one. The marking duty in Article 50 falls on the provider of the AI system, meaning the company whose generator you use. It is their job to mark output in a machine readable format.

The European Commission's answers on the Article 50 transparency obligations set the application date at 2 August 2026, with marking and detection duties for systems already on the market extended to 2 December 2026, and no retroactive labelling of content published before the start date. The duty that lands on you as a deployer is narrower. It concerns deep fakes, defined by resemblance to real people or objects, a high degree of similarity and the capacity to appear authentic, and it is discharged by telling the viewer. The Commission also notes an exemption where the system performs an assistive function for standard editing.

The honest reading for a shop is this. A generated image of your own product, sold as a picture of your own product, is not a deep fake and is not the target of that provision. The Commission's material does not walk through advertising and product listings as a worked example, so anyone telling you Article 50 dictates a specific listing label is filling a gap with confidence rather than text. Your enforceable obligations on a listing come from the platform's feed policy and from consumer protection law about misleading commercial practices, both of which existed before any of this. We laid out the wider structure in the AI Act read as a decision tree and a timeline, and transparency is one branch of it.

Where do AI product photos actually get sellers in trouble?

Returns, not regulators. The mechanism is dull and expensive: a picture that flatters produces an order, the item arrives, the buyer sees a different thing, and the cost lands on you three times over.

You pay the outbound shipping, you pay the return, and you carry an item that is now used stock. Worse, the buyer files the claim as not as described rather than changed my mind, which shifts the postage liability and, on several platforms, counts against the account health metrics that decide your visibility. A generated image is unusually good at producing that outcome, because generators are trained to make things look appealing and have no idea what your fabric actually does under a window.

The three details that cause it are consistent enough to check for deliberately. Colour, because generators drift toward saturated versions of whatever you asked for. Texture, because a weave or a grain gets rendered as an idealised pattern. Scale, because nothing in a generated scene knows how big your product is unless the scene contains a real reference. If the picture is your only size cue, buyers will guess wrong, which is one more reason a real object in frame beats a beautiful empty one. Our piece on who decides on the refund when one order in five comes back covers what to do once it has already happened.

Do you have to tell the buyer the photo was generated?

On the listing itself, generally no, and the people insisting otherwise are describing a rule that does not exist in most places. What the current framework asks for is machine readable marking in the file and honest representation on the page. A visible sentence saying this image was generated is a choice, not an obligation, outside the narrower cases the transparency rules single out.

Social platforms are the exception worth knowing, because they add their own layer on top and they apply it automatically. Where a platform detects the provenance signal in a file it may attach its own label to your post without asking, which means the metadata you carefully preserved for the feed can produce a badge on an image you meant to look like a photograph. That is not a penalty. It does change how the post reads, and shops are better off deciding in advance which images they are happy to see labelled. We went through what a shop actually has to declare in the piece on the AI label and what a shop must declare.

There is a commercial argument buried in that decision as well. In several categories buyers now read a generated hero shot as a signal that the seller is far from the product, which is fatal for anything sold on craft or provenance. In other categories nobody notices or cares. Knowing which one you are in is a merchandising judgement rather than a compliance one, and no policy page will make it for you.

The provenance layer arriving underneath all of this

IPTC fields are one way to say where a file came from. The other is Content Credentials, the C2PA standard, which wraps the same claim in a signed manifest that travels with the asset and can be verified rather than merely read.

The C2PA guidance for AI generated assets goes considerably further than a single tag. It records which model produced the asset, links to that model's own credential, can carry the prompt as an input ingredient, and supports assertions about whether the result may be used for training. For a seller the practical meaning is modest today and larger later: as platforms begin reading signed credentials rather than trusting an editable field, the shops whose pipeline preserves provenance will get the benefit of the doubt and the shops whose pipeline flattens everything will not. We went through what these credentials do and do not prove in a separate piece on what Content Credentials actually prove.

None of this requires action this quarter. It does argue against building an image pipeline that strips everything, which is the default in a lot of store software. If you are choosing or rebuilding the storefront itself, the ability to control how images are stored and served is worth more than it looks on a feature list, and it is one of the arguments for a shop whose code and image pipeline you own outright rather than renting.

What a careful seller does in an hour

Start by finding out where you stand rather than changing anything. Take five product images that you know were generated, fetch them from the live URLs in your feed, and look at whether the source tag survived. Most shops discover the answer is no, and the discovery is worth more than any policy summary.

If the tag is gone, the fix is usually one setting in whichever tool re-encodes your files, not a new tool. Where the setting does not exist, writing the field back is a small scripted step at export time, and it only has to run on the images that need it.

Then look at your text. If you generated titles or descriptions in bulk and shipped them straight into the standard fields, moving them to the structured attributes is a feed mapping change rather than a rewrite. It costs an afternoon once and it removes a category of risk that grows quietly as the proportion of generated copy in your catalogue rises.

Last, write down which images in your catalogue are generated and which are photographed. Not for a regulator. For yourself, in six months, when a buyer disputes a colour and you need to know whether the picture they are holding you to came from a camera or from a prompt. Shops that keep the legal fields most product listings are missing tend to keep this one too, and it is the same ten minutes.

One more habit worth forming while you are in there. Keep the original camera file for anything you later composite, in a folder that is not your store's media library. When a buyer questions a colour, the unedited frame settles it in a minute, and when a platform asks how an image was produced you have an answer rather than a reconstruction.

The summary a busy seller can act on: generated images are fine, the labelling is real but small, the file is where the labelling lives, and your resize pipeline is the thing most likely to break it. Everything else in this subject is noise until one of those four is wrong.