AI images on marketplaces and in ads: what Amazon, Google, Meta and others require

Many stores show their product images not only in their own online store but also in Google Shopping, on marketplaces such as Amazon or Kaufland, in ads on Meta, TikTok or Pinterest, and in newsletters. An AI label that your store displays on the image does not arrive there. This guide sets out what eleven platforms currently require for AI-generated or AI-edited images, what they label themselves and what they forbid in the image, and what you can do for each channel. All platform information comes from the platforms’ own help and policy pages, retrieved on 1 October 2026. Platform rules change often; check them there before every campaign. For the underlying rule, see Article 50 explained.

Why does the label from your store not travel with the image?

A label that your theme displays over a product image is part of the store page, not part of the image file. A feed, a marketplace upload or an ad asset takes the file or its address, not the page. That is why the label is missing wherever the image appears outside your store. Easy AI Act Image Labels works exactly this way: the app displays the label as an overlay in the online store, changes no image file, and sets no label in feeds, on marketplaces, in ads or in emails.

The machine-readable marker in the file, which some platforms evaluate, does not reliably arrive either. On the copies Shopify delivers, only 6 of 38 marked images still carried their marker on 30 September 2026. Which image address a Shopify sales channel passes to a platform has not been measured.

Does the labelling obligation also apply in ads and on marketplaces?

Yes, if the image is a deep fake. Article 50(4) of the AI Act (Regulation (EU) 2024/1689) is tied not to a channel but to the content: a deployer who uses an AI system to generate or manipulate an image that resembles existing persons, objects, places, entities or events and would falsely appear to a person to be authentic (Article 3(60)) must disclose that it has been artificially generated or manipulated. Under Article 50(5), the information must be provided in a clear and distinguishable manner at the latest at the time of first exposure, and under the Commission guidelines of 20 July 2026 this applies to every person who sees the image (paragraph 143). The Digital Omnibus (Regulation (EU) 2026/1744) has not changed this.

The guidelines name advertising explicitly:

• An AI-generated image of a product in an advertisement that makes the product appear more appealing, of higher quality or different from real life is a deep fake (examples to paragraphs 113 to 116).
• A real product shown in an advertisement against an AI-generated background is not, as long as the ad is not likely to mislead about the product (same examples).
• Even a company outside the EU that uses AI to generate a deep fake for an advertisement displayed in the Union is a deployer (paragraph 14). What that means for stores based outside the EU: Does the EU AI Act apply to my store outside the EU?

On distribution through platforms, the guidelines say:

• You remain responsible. Deployers in distribution chains should take proportionate measures so that their label reaches the audience clearly, for example through contractual conditions with distributing partners or the settings of the interface (paragraph 12).
• The platform is not a deployer where it only disseminates content created by third parties. It is strongly encouraged to preserve labels and markers, but is not responsible for them under the AI Act (paragraphs 16 and 98).
• You may use the labelling tools of very large platforms. Where a very large online platform provides a tool that lets you disclose deep fakes clearly and distinguishably as AI, you can rely on it within that platform; your own responsibility is unaffected (paragraph 126).
• A note in menu layers is not enough. Information hidden under layers of menu options or in terms of use, which can easily be overlooked, is not considered clear and distinguishable (paragraph 142).

Two rules need to be kept apart from this. Article 35(1)(k) of the Digital Services Act lists, as a possible measure for very large online platforms, making deep fakes distinguishable through prominent markings on their interfaces; that is addressed to the platform and creates no image obligation for your own store. And the machine-readable marking under Article 50(2) of the AI Act is owed by the provider of the generator, not by you; for systems placed on the market before 2 August 2026, by 2 December 2026 (Article 111(4), inserted by Regulation (EU) 2026/1744). Your obligation is the visible disclosure under Article 50(4), and you cannot rely on the marker for it (paragraph 117).

What do the platforms require at a glance?

The table separates three things: whether a platform requires disclosure, whether it applies a label itself, and whether a burned-in label in the image is allowed. “No rule found” means that the publicly retrievable primary sources said nothing about it on 1 October 2026.

What applies on the marketplaces in detail?

Amazon

• Obligation for AI-generated people: If a product image shows photorealistic people who were entirely generated by AI, you must tag it with metadata before uploading: the keyword contains-synthetic-performer in the XMP field dc:subject. According to Amazon, it will then add a disclosure, where applicable, that the image includes AI-generated people. You should not tag images with real people, even if altered with AI, images without people, or people who are not photorealistic (product image guide).
• A+ Content: When uploading in A+ Content Manager, there are the checkboxes “AI-generated” for AI images without people and “AI-generated people”; Amazon then applies the metadata itself.
• Main image: It must show the product realistically and in its real scale, quantity and colour, on a pure white background. Amazon allows lifestyle main images only for a limited number of product types, and then with no text or extra logos. The retrieved version contains no explicit ban on text or watermarks in the main image on white; Amazon does, however, reserve the right to modify submitted images. Whether a burned-in AI label would stay there is open.
• Limit: Amazon’s rule covers only AI-generated people. An AI product image without people that makes the product look nicer or different is still a deep fake under the guidelines, and it also runs against Amazon’s requirement to show the product realistically. No disclosure route for such images was found outside A+ Content.
• Ads: The terms for Amazon’s advertising AI tools forbid using them to create or use deep fakes. No disclosure rule of its own for Amazon ads was found.
• Via Shopify: Whether the keyword reaches Amazon through a Shopify connection has not been measured.

eBay

• Ban in the image: eBay’s picture policy does not allow photos with added text, artwork or marketing material, added borders, or watermarks of any type. A burned-in AI label would fall under this wording (own reading). Photos that don’t accurately represent the item are not allowed either.
• AI rule: No rule of its own on disclosing AI images was found. eBay itself offers an AI feature that places item photos in front of an AI-generated background; the page does not mention labelling.

Etsy

Etsy’s policy and help pages, including the Creativity Standards, the seller policy and the help on listing photos, were blocked for automated retrieval on 1 October 2026. This guide therefore makes no statement about Etsy. Check directly with Etsy what applies there to AI designs and listing photos before you list.

OTTO Market

• In the publicly accessible documents, the Onboarding Guide of 16 February 2026 and the API documentation, no rule on AI images was found on 1 October 2026. The partner helpdesk requires a login.
• For the main image, the Onboarding Guide requires a cut-out, centred item on a white background, only the main product and a continuous white border. It contains no explicit ban on text or watermarks; the requirement leaves hardly any room for a label (own reading). Ask the partner helpdesk how OTTO deals with AI images.

Zalando Partner Program

• No visible label, no images that need one: Zalando’s image guidelines for partners state: “We don’t allow visible on-image labels (e.g. text overlays or watermarks). Don’t provide content that requires visible labelling under applicable laws”, explicitly including deep fakes. An AI image that you would have to disclose therefore does not belong in the Zalando catalogue.
• Invisible marking: Under the heading “Invisible marking (required by December 2026)”, Zalando requires identifying data to be embedded in the metadata or as an invisible watermark, and the metadata to be kept so that it is not stripped during upload.
• Product accuracy: Zalando reserves the right to reject AI images that do not represent the product accurately or break believability.

Kaufland

• Kaufland labels itself: According to a Kaufland Global Marketplace blog post of 23 July 2026, the technology behind the Kaufland marketplaces reads the metadata of AI-generated images on upload and automatically adds a visible AI label. This requires the metadata to remain unchanged; Kaufland names screenshots, format changes and sending via messaging services as points where it gets lost. The graphic “AI label variants” in the post shows a square field with the text “AI” in a light and a dark version.
• Ban in the image: The product data guideline allows only neutral product images and excludes text, logos, symbols and seals; for Sponsored Brand Ads the same applies to overlaid elements. A burned-in label would fall under this wording (own reading).
• Limits of the automation: It finds only what is in the file. Not every generator writes a marker: on the Easy AI Act Image Labels test bench on 1 September 2026, 13 of 47 models in a generator catalogue wrote none. As the check criterion, the post names the value trainedAlgorithmicMedia; whether Kaufland also detects compositeWithTrainedAlgorithmicMedia, the value for photos edited with AI, is not stated. And a label reading only “AI” leaves open whether an image was generated or modified; in the user tests for the Code of Practice, the variants with a word performed markedly better (see The EU icons for AI images).

What applies in ads and on social networks?

Google Ads, Shopping and Merchant Center

• Label setting: In Google Ads, in Merchant Center and in other Google advertising products, you can designate an asset as “AI created or edited”. Worldwide, a disclosure then appears in the “How this ad was made” section of My Ad Center, which users open via the ad’s three-dot menu. For campaigns that target the European Union, India and New York, ads with assets designated this way also include visible overlays, according to Google. In Merchant Center you set this per image under “Manage images” and “Manage AI label”.
• Your own label in the image: Since the policy update of 9 July 2026, text or visual labels for AI content in image and video creatives do not violate Google’s policies prohibiting text overlays and watermarks. The Merchant Center help says the same for ads and listings; if Google detects an AI label, automatic image enhancements such as overlay cleanup do not run. Google advises not to place the label in the corners or at the edges, because it can be cropped, and to turn off image enhancements.
• What Google does itself: Images from Google’s own AI tools carry SynthID watermarks and C2PA data. Google may also label assets itself, for example where it is legally required to or receives signals from other platforms; such labels cannot be overwritten.
• Metadata in the feed: Separately, Merchant Center requires AI product images to carry metadata such as IPTC DigitalSourceType and asks you not to remove it. What of this reaches Shopify’s copies: AI images in your Google Merchant Center feed.
• Open point: The image link help page still lists overlays and watermarks among the reasons for disapproval without mentioning the AI exception, and the requirements for YouTube Shopping ads also exclude overlaid elements without a note. The exception appears in the policy update and on the label setting pages.
• Google itself points out that the label setting does not guarantee compliance with regulations and that a label can affect where content appears.

YouTube

• Obligation: YouTube requires creators to disclose realistic content that was meaningfully altered or generated with AI, for example a realistic scene that never happened; this is done in YouTube Studio. For thumbnails created or improved with AI, YouTube explicitly requires no disclosure.
• What YouTube does itself: YouTube automatically applies a label to content made with YouTube’s own AI tools, to content with C2PA metadata and to content its internal systems detect as AI; since May 2026, gradually, also where a creator has not specified anything but the systems detect significant photorealistic AI use. For long-form videos the label sits below the player, for Shorts as an overlay on the video.
• Ads on YouTube run through Google Ads; the label setting above applies there.

Meta: Facebook, Instagram and ads

• Obligation: Meta requires self-labelling only for organic posts with photorealistic video or realistic-sounding audio; no such obligation was found for paid ads (details: Labelling AI videos). On images, Meta writes: “Meta does not require you to label images that have been created or modified with AI”; detected AI images still receive a label.
• What Meta does itself: The label is called “AI info”. Meta applies it when it detects industry signals such as C2PA or IPTC data, or when someone discloses the AI origin. Since an update of 12 September 2024, for content that was only edited with AI it sits in the post’s menu; for AI-generated content it remains on the post.
• Ads: Ads created or significantly edited with Meta’s own generative tools receive “AI info”; if this results in a photorealistic AI-generated human, the label appears next to “Sponsored”. With an update of 1 June 2026, Meta says it will also begin detecting ads made with third-party AI tools through industry signals; the label then sits under “About this ad” in the three-dot menu. An undated Meta help page still says that ads made with third-party tools do not always carry “AI info”.
• Ban in the image: No rule forbidding text or a burned-in AI label in the ad image was found in the retrievable primary sources; the help page on text in ad images could not be retrieved as text.
• For ads about social issues, elections or politics, a separate disclosure obligation applies to photorealistic AI images.

TikTok, TikTok Shop and TikTok ads

• Obligation: Under the Community Guidelines effective 24 September 2026, you must label content that uses AI to show realistic-looking people or scenes, explicitly including AI images that look like real photographs. You can use TikTok’s AIGC label or your own clear caption, sticker or watermark. Small edits such as colour correction or cropping need no disclosure.
• TikTok Shop: The EU version of the Shop policy on AI content of 28 August 2026, which also applies in Germany and Austria, requires disclosure in the content itself (on-screen text, watermark, sticker, note in the description) or via the “AI-generated content” toggle; a label reading “Disclosed by creator as AI-generated” then appears in the bottom-left corner. For product images in the catalogue, TikTok Shop in Germany recommends avoiding watermarks and too much text (best practice, as of 10 September 2026).
• What TikTok does itself: If content contains specific technical metadata, TikTok automatically applies an “AI-generated” label that you cannot remove. For this, TikTok reads C2PA Content Credentials, and in November 2025 it announced that it would also add invisible watermarks to content made with TikTok’s own AI tools and to content uploaded with Content Credentials.
• Ads: You must label fully AI-generated images, videos or audio in ads, or ones significantly modified by AI, with the toggle “This ad contains AI-generated content” in Ads Manager or your own clear disclaimer, watermark or sticker; otherwise TikTok rejects or restricts the ad. TikTok counts removing or modifying backgrounds among insignificant AI edits (ads policy, last updated April 2026). Spark Ads follow the Community Guidelines.

Pinterest

• Obligation: No explicit labelling obligation was found. Content owners can indicate that a Pin was generated or modified with AI.
• What Pinterest does itself: Pinterest applies the label “AI modified” when it detects the AI origin via IPTC metadata or its own classifiers, or when the content owner indicates it. The label appears at the bottom left in close-up after someone clicks on the Pin; for ads, only in the menu with the ad details.
• Ads: The advertising guidelines exclude overly busy images packed with text overlay, but do not name a ban on a single label. From 12 November 2026 they prohibit ads for products misrepresented through synthetic media or manipulated content (preview of the new version).

What can you do for each channel?

1. Classify the image first, then the channel. Whether an image must be disclosed depends on whether it is deceptively realistic, not on the platform. Which image types fall under this is covered in the guide Which Images in Your Store Need an AI Label?. Some platforms go further than the AI Act: TikTok requires a disclosure in ads for every fully AI-generated image. Beauty images have their own guide: Labelling AI images in a cosmetics and beauty store.
2. Use the platform’s toggle and check where the label appears. Google Ads and Merchant Center, TikTok, YouTube, Pinterest, Meta for posts and Amazon for A+ Content offer a setting on upload or in the account. What matters is whether viewers see the note on the image without clicking, like the overlay Google shows in EU campaigns. If it sits only behind a menu, as for ads on Pinterest or under “About this ad” on Meta, that is not enough under the guidelines (paragraph 142). In that case your own label belongs in the image, where the platform allows it; the Wettbewerbszentrale also recommends this where platform labels are not clear enough.
3. Where allowed, upload a labelled version of the image. For Google ads, TikTok and Meta, a separate derivative with the word variant burned in, AI GENERATED or AI MODIFIED, works well, because there is no explanatory text next to it. The original file stays without a burned-in label, so that you can derive the right version for each channel. The free AI Label Generator creates such a version in the browser; the copy keeps the IPTC data of the original, but not a C2PA manifest. Variants and placement: The EU icons for AI images.
4. Where text in the image is forbidden, solve it differently. eBay and Kaufland forbid added text in product images, Zalando does not allow visible labels, and on Amazon a lifestyle main image is allowed only without text. What remains there is the platform’s own route where one exists (Kaufland reads the metadata, Amazon the keyword for AI-generated people), a real product photo instead of a deep fake, or asking seller support how you should disclose a deep fake there. According to its guidelines, Zalando does not accept images that need a label at all.
5. Keep the marker in the original file. Kaufland, Meta, TikTok, Pinterest and YouTube evaluate provenance data, Amazon a keyword of its own for AI-generated people; Google Merchant Center and Zalando require metadata. Download AI images with the tool’s download button in the native format, with no screenshot and no converter in between. Whether a file carries evidence is shown by the AI image checker, without the file leaving your device. If it finds none, that says nothing about the origin. And a marker does not replace a visible label (paragraph 117).

What remains open?

• Which image address and format the Shopify sales channels for Google, Meta, TikTok, Pinterest or the marketplace connections pass on, and whether the marker arrives there.
• How OTTO deals with AI images: the rules in the partner helpdesk are not public.
• What Etsy requires for AI images: the pages could not be retrieved.
• Whether Amazon keeps a burned-in AI label in a main image on white, and whether Meta’s rules on text in ad images cover such a label.
• Whether Google’s AI exception also applies to YouTube Shopping ads, whose requirements page excludes overlaid elements without a note.
• Whether a platform label meets the requirements of Article 50(5) in an individual case is assessed by the guidelines only in general terms (paragraphs 126 and 142); no decision by an authority or court on this is known (as of 1 October 2026).

Status of all information: platform pages and legal texts retrieved on 1 October 2026, test bench measurements of 1 September 2026 and 30 September 2026. General information, not legal advice in an individual case.