- The exposure is not "did a model make this image". It is whether the picture creates an impression of the product that the product does not deliver.
- The UK advertising regulator states plainly that a disclosure of AI use does not rescue an image that is misleading in the first place, and gives a cosmetics example.
- The US standard is older and blunter: an ad is deceptive if it is likely to mislead a reasonable consumer on something material, and you must hold proof before it runs.
- Of the five jobs a catalogue image does, generated imagery is safe on one, conditional on two, and a poor idea on two.
- Google's shopping feed rejects any image that is not the actual product, including illustrations and generic graphics, which settles the hero shot before advertising law even gets involved.
- The EU marking rules for synthetic images started applying on 2 August 2026, with a grace period to December 2026 for systems already on the market.
- The single question that resolves most cases: will the customer physically receive the thing in the picture.
A generated image is not automatically a lie, and treating it as one costs small shops money they do not need to spend. A generated image is also not automatically fine, and treating it that way is how a merchant ends up arguing about a refund with a photograph they cannot defend. The same line applies to moving pictures, where exact product fidelity is still out of reach. The line between those two runs somewhere specific, and it was drawn by advertising regulators long before anyone could type a product into an image model.
The useful thing about that timing is that the rules are settled. Nothing here is speculation about pending legislation. Every rule below is already enforced, most of it for decades, and it maps onto generated images without much interpretation. What follows is the mapping, job by job.
Does a generated catalogue image break advertising rules?
Not by existing. It breaks them when the image creates a materially misleading impression, and regulators on both sides of the Atlantic apply that test to the picture regardless of how the picture was made.
The clearest statement of this comes from the UK's Advertising Standards Authority, whose guidance on disclosing AI use in advertising makes two points that a merchant should internalise. The first is that disclosure is not required in all cases; the question they pose is whether the audience would be misled if the AI involvement were not mentioned. The second is the one that catches people out: a disclosure is very unlikely to fix the harm caused by a message that is fundamentally misleading. Their worked example is a cosmetics product, where showing an AI generated result that does not match real world results and then captioning it as AI does not repair the deception.
The American standard arrives at the same place by a different road. The Federal Trade Commission's advertising guide for small businesses defines a deceptive ad as one that is likely to mislead a consumer acting reasonably, on something material to their decision to buy. Ads are judged from the point of view of a typical person seeing them, taking words and pictures together rather than parsing sentences in isolation. And the requirement most small sellers have never heard: the law expects you to have proof before the ad runs, not to assemble it if someone complains.
Both regimes treat fine print the same way. The FTC guide's example is a weight loss ad promising results, with "diet and exercise required" in small type underneath. The small type does not cure it. If your instinct is to publish a generated image and add a caption saying it is illustrative, check first whether the caption is doing the work the image should have done.
Which image jobs can a generated picture honestly do?
A product listing does not use one image, it uses several, and they are not doing the same job, a split that matters more now that shopping assistants assemble a basket straight from listing data. Treating them as one category is why this argument goes in circles. Split them and most of the difficulty disappears.
| Image job | What it promises the buyer | Can a generated image do it honestly? | What decides it |
|---|---|---|---|
| Hero shot | This is the item you will receive | No | It is a direct representation of the goods. Shopping feed policy also rejects non actual product images outright |
| Variant image | This is what the blue one looks like | Only if generated from a photo of the real variant | A recoloured render of a product you never photographed in that colour is a claim about stock you cannot substantiate |
| Scale cue | This is how big it is next to a familiar object | Risky | Scale is the easiest thing to get subtly wrong, and it is material to a purchase. A ruler beside the real item beats any render |
| Detail crop | This is the stitching, the grain, the finish | No | It is evidence of quality. Generating it is generating the evidence |
| Lifestyle scene | Here is a context this could sit in | Yes, with care | Nobody expects to receive the kitchen. The product inside the scene still has to be the real product |
The pattern in that table is worth stating explicitly because it is the whole article compressed. Generated imagery is safest exactly where it is furthest from the transaction. A background is not part of what the buyer receives. A detail crop of the weave is close to being the reason they clicked buy.
What did the regulator actually rule on images?
Rulings are more instructive than principles, and the ASA has published a set that translates directly. Its guidance on avoiding misleading imagery in ads collects several, and none of them originally involved AI at all, which is the point.
Two categories recur. The first is images that show more than the price buys. A food ad picturing a complete meal when the price covers one component was ruled misleading, as was a bathroom listing showing an entire fixture when only a part was included. The second is images that exaggerate what the product does. A teeth whitening ad implying instant results and a tanning product shown through a filter that improved skin tone beyond the product's actual effect were both ruled against. The games industry example is the sharpest: ads must show representative gameplay, and a disclaimer saying the footage is not representative does not save an ad that is not representative.
Map those onto a generated catalogue and the risky cases name themselves. An image that shows accessories not in the box. An image where the colour is richer than the dye. A render where the fabric drapes better than the fabric does. Every one of those is the old ruling wearing new clothes, and the regulator does not need a new rule to reach it.
What does the shopping feed require, separately from the law?
This is where the hero shot stops being a judgement call. Google's specification for the image link attribute in Merchant Center requires the image to accurately display the entire product with minimal or no staging, and lists what will get a product disapproved. The prohibited list includes a placeholder or any image that does not show your product, a generic image, graphic or illustration that is not the actual product image, and a logo or icon in place of a product photo. Promotional overlays, watermarks and borders are out too.
Variants get their own instruction: submit a unique image showing the distinguishing details of each variant, showing only one variant per image. Sizes that look identical may reuse an image, colours may not. Anyone planning to generate a colour range from one photograph should read that line twice.
The size floor moves as well. Merchant Center will require at least 500 by 500 pixels for all products from 31 January 2027, while recommending 1500 by 1500 or above for best performance across listing formats. Useful to know now if you are about to redo a catalogue, because doing it twice is the expensive outcome, and the same pass is the moment to sort which catalogue images need real alt text and which need an empty one.
What changed on 2 August 2026?
The European transparency rules for synthetic content started to apply. The Commission's summary of the AI Act transparency rules puts the date at 2 August 2026 and requires providers of generative systems to mark synthetic image, video, audio and text in a machine readable way. The obligation sits with the provider of the tool rather than with you, but it changes the environment you publish into, because the marks travel with the file.
Two carve outs matter to a shop. Marking does not apply where the system performs an assistive function for standard editing or does not substantially alter the input or its meaning, which covers most retouching. And systems already on the market before that date have until December 2026. Penalties run to 15 million euros or 3 percent of worldwide annual turnover, which is aimed at providers rather than at a merchant with forty products, but it explains why the tools are being changed underneath you.
Our earlier piece on what generated product photography holds up on and where it visibly breaks goes through the technical side of this, including the origin metadata Merchant Center expects on generated files. The two pieces are complementary: that one is about whether the image survives inspection, this one is about whether it survives a complaint.
The question that resolves most cases
Will the customer physically receive the thing in the picture? If yes, the picture has to be of the thing. If no, you have room, and a low denoise pass in a ComfyUI graph you can rerun per product keeps that room consistent across a catalogue.
That single test handles the majority of decisions a small shop faces, and it is worth applying before reaching for any of the rules above. A soap maker photographing bars needs photographs of bars. The same soap maker generating a warm bathroom shelf for the bar to sit on is not making a claim about the shelf, because nobody buying soap believes the shelf is included.
Where it gets genuinely harder is the middle ground the test does not reach: made to order goods, digital products, and services. If the customer receives a wooden table built after they order, no photograph of their table can exist. The honest solution is a photograph of a previous table with the difference stated, not a render, because the render invites a comparison the finished object may lose. For a purely digital product the physical test does not apply at all and the FTC standard takes over: does the image create an impression of what the product does that the product does not deliver.
Does labelling an image as AI made protect you?
It protects you from an accusation of concealment. It does not protect you from an accusation of misrepresentation, and those are different complaints with different consequences.
This is the ASA's point restated, and it is the most commonly misunderstood thing in this area. A merchant who labels a misleading image has been transparent about the method and still misleading about the product. A merchant who does not label a fair, accurate lifestyle image has done nothing wrong under either regime, because the audience was not misled about anything material.
So labelling is a good habit for reasons of trust rather than a compliance shield. Where it earns its keep is internal: if your files carry their origin, you can answer a question about any image in your catalogue in seconds instead of guessing a year later.
What a returns dispute actually looks like
Nothing above has teeth until money is attached, and the surface where it gets attached is the refund. A buyer receives an item, compares it to the listing image, and says it is not what was shown. At that point the marketplace, the card scheme or the small claims process asks a very simple question, and it is not about AI. It is whether the listing image represented the goods.
A photograph of the item answers that immediately. A generated hero image cannot answer it at all, because there is no version of the argument where "the picture was an impression of the product" helps the seller. This is why the hero shot row in the table is not marked risky but simply no. The cost of getting it wrong is not a regulatory fine that will probably never come. It is a chargeback that certainly will.
If you are setting up a catalogue and want the flexibility to swap imagery per product without a redesign each time, keeping the storefront code in your own hands makes that a five minute job rather than a support ticket. Our AI ecommerce store builder generates the product pages into a repository you own, which is the difference between changing an image policy and requesting one. For the broader question of which AI tools are worth a slot in a small operation at all, the scored list in our piece on the AI tools worth paying for in a small business covers image generation among the rest.
Who is responsible when the tool made the picture?
You are. The ASA states it without hedging: delegating production to an algorithm does not delegate accountability, and the primary responsibility for compliance stays with the advertiser. The same holds under the American framework, where the party running the ad carries the substantiation duty regardless of who or what produced the creative.
This matters more than it sounds, because it removes a defence people expect to have. A merchant who buys a generation tool, types a description of their own product, and publishes what comes back has not outsourced any part of the judgement. If the output shows a feature the product lacks, the merchant made that claim. The tool's terms of service are between the merchant and the vendor and have no bearing on the buyer or the regulator.
The practical consequence is a workflow one, and it is small. Someone who knows the product has to look at every generated image before it goes near a listing, with one specific question in mind rather than a general sense of whether it looks good. The question is not "is this a nice picture". It is "is anything in this frame a claim I could not defend".
A practical policy for a small catalogue
Written out, the rules a one person shop can actually follow come to five lines.
- Photograph everything the buyer receives. Hero, variants and detail crops come from a camera. A phone on a windowsill is sufficient; the requirement is truth, not production value.
- Generate only what the buyer does not receive. Backgrounds, scenes and context are fair game, provided the product inside them is the real photographed product.
- Never generate evidence. If an image is the proof behind a claim about durability, finish, capacity or result, it has to be real. This is the ASA's cosmetics case and the FTC's substantiation rule in the same sentence.
- Match the variant to the stock. One unique image per colour, taken from that colour. This is feed policy, not opinion, and it is enforced by disapproval rather than by complaint.
- Keep the origin on the file. Not because a rule forces you to, but because in a dispute the fastest thing you can produce is a clear record of where an image came from.
None of this asks a merchant to avoid the tools. It asks for the distinction the regulators have already made and the platforms have already coded: an image that sets a mood is decoration, an image that stands in for the goods is a promise. Decoration can be made. Promises have to be kept.