The EU AI Act became enforceable on 2 August 2026. Most ecommerce teams have spent the summer worrying about AI labelling requirements on their product pages. The regulation that actually binds them shipped in April 2024, inside the Google Merchant Center feed spec — not from Brussels.

The EU Act’s text-labelling duty almost certainly does not apply to your product descriptions. The rule that does apply has been live for over two years, and most stores are quietly non-compliant. This article is a companion to the technical breakdown of AI watermarking and Article 50, which covers mechanism and the three-layer regulatory structure in full. Read that first or after — you can follow this piece either way.

The Content-Type Matrix

This table is the primary reference. Every row maps a content type to its actual obligations.

How Watermarking Works

The full mechanism is covered in the parent article. Short version: the watermark lives in word choice, not hidden Unicode characters — the model biases token selection according to a secret key. As of August 2026, Anthropic and Google are the only providers with a text watermark in production for text, and neither has a public detector. No third party can read the mark, which is why every enforcement question in this article rests on self-declaration, not detection.

The Merchant Center Requirement

Google’s AI-generated content policy for Merchant Center requires that AI-generated product titles and descriptions be submitted with explicit source declarations.

What is required:

AI-generated titles must use structured_title with two sub-attributes: digital_source_type set to trained_algorithmic_media, and content containing the actual text. AI-generated descriptions must use structured_description with the same sub-attributes. This applies whether the copy came from Product Studio, ChatGPT, Claude, Shopify Magic, a custom model, or a third-party feed tool.

The digital_source_type vocabulary is defined by the IPTC. For model-generated text, the correct value is trainedAlgorithmicMedia. The IPTC’s metadata guidance for synthetic media provides the broader context.

The XML:

<g:structured_title>
  <g:digital_source_type>trained_algorithmic_media</g:digital_source_type>
  <g:content>Product title text here</g:content>
</g:structured_title>

<g:structured_description>
  <g:digital_source_type>trained_algorithmic_media</g:digital_source_type>
  <g:content>Product description text here</g:content>
</g:structured_description>

Three complications:

1. The documentation inconsistency. The AI-generated content policy page treats digital_source_type as required for generative AI content. The title attribute reference calls it optional, defaulting to default. These are in tension. Follow the policy page — it is the operative document.

2. The precedence gotcha. If your feed submits both title and structured_title, Google uses title. A feed that emits both is, in effect, submitting the undeclared version and discarding the declared one. Verify what your feed tool actually emits. Many platforms add structured_title correctly but leave the legacy title field populated, which silently wins.

3. It is self-declaration. Google has no watermark check here. It cannot read the mark in your copy, because detection requires the provider’s private key. Compliance is an honesty and process question, not a detection question. This means the risk of non-compliance is a feed-disapproval risk on audit, not an automated detection event. State this plainly in your internal documentation so nobody treats the absence of an error as confirmation of compliance.

Where the boundary sits. “AI-generated” means the model produced the text. Using AI to brainstorm, refine structure, or check grammar on copy you wrote is categorically different. Amazon’s KDP AI content guidelines draw exactly this line and are a useful reference when writing internal policy for content teams.

For images (just briefly): Google requires that AI-generated images retain IPTC DigitalSourceType metadata, including from Product Studio output. Most image optimisation pipelines — CDNs, resizers, compression tools — strip metadata by default. This is a real and very common silent failure. Verify yours.

What the EU AI Act Actually Asks of a Store

Three questions, in the order teams ask them.

Do product descriptions need a visible AI label?

Almost certainly not. The EU AI Act’s text-labelling duty for deployers — Article 50(4) — applies only when three conditions are simultaneously met: the content is published, it informs the public, and it concerns a matter of public interest. The Commission FAQ and the Guidelines on Transparency list the qualifying topics: politics, public administration, justice, fundamental rights, public security, public health, environment, consumer safety, and economic, financial, scientific or cultural developments open to public debate.

Advertising copy, product descriptions, social posts, and newsletters generally fall outside this test. Gleiss Lutz on AI transparency in advertising and Davis+Gilbert for advertisers both reach the same conclusion. The duty is also removed entirely where there is genuine human review — someone with editorial authority who substantively examined and approved the text. That is the normal state of affairs in marketing.

The misinformation circulating in the compliance-vendor space: several vendors are claiming that AI-written product descriptions require a visible on-page disclosure under Article 50(2). This conflates the provider’s machine-readable marking duty — which sits on Anthropic and Google, not you — with the deployer’s visible labelling duty, which has a public-interest threshold you almost certainly do not meet. They are different obligations on different parties. Greenberg Traurig and Lewis Silkin provide correct readings of the four-obligation structure.

Where does a store actually have exposure?

Two places, and neither is the product page.

Your chatbot. Article 50(1) requires that users be told they are interacting with an AI, clearly, at the point of first interaction. No public-interest threshold. No human-review exemption. This applies to product finders, support bots, sizing assistants, and WhatsApp flows. It is visible from the outside — a regulator, competitor, or procurement team can check it in thirty seconds. Lausen’s analysis confirms there is no grey zone here. Fix this first; it takes an afternoon.

Editorial-shaped content on public-interest topics. Pure product advertising is outside the test. Marketing designed to look editorial — native advertising on health, finance, or safety subjects — may fall inside it. Supplements, medical devices, financial products, insurance, child safety. If your content hub reads like journalism and covers a YMYL topic, apply the three-part test properly and document the result. DLA Piper’s analysis of the deployer timeline and Bird & Bird’s guide to the final Guidelines are the best secondary sources for this assessment. The EU has published an optional icon set if you determine a label is warranted.

Am I in scope at all if I’m not in the EU?

Yes, if your output reaches EU audiences. The obligations follow where the content is consumed, not where the business is incorporated. A US Shopify store selling to German customers is a deployer in scope. Anthropic applies watermarking worldwide regardless of operator location.

Dates: Article 50 obligations apply from 2 August 2026. Pre-existing AI systems have until 2 December 2026 for the provider-level marking duty. Watermark-detection interoperability is due 2 February 2027. Penalties reach €15m or 3% of worldwide turnover — directed at systemic non-compliance, not a missing badge on a product page. California SB 942 became operative the same day and covers image, video, and audio. Not text. This is consistently misreported.

Reviews: The Genuinely Risky Area

On translation specifically. Review-authenticity rules require that what you present to consumers accurately reflects genuine consumer experience. Machine translation fails on informal, idiomatic, typo-laden review text at a rate that matters commercially. It softens complaints, loses qualifiers, and changes emphasis. The translation now carries the translating model’s watermark; the original reviewer’s words are gone. You have manufactured a provenance signal pointing at your AI vendor for text a customer wrote.

Amazon’s KDP policy treats AI translations as AI-generated content requiring disclosure. That is a book-publishing rule, not ecommerce law — but it is the clearest existing statement of the principle and the direction of travel. The PowerReviews analysis of Omnibus compliance and KG-Legal on fake reviews under EU law provide the regulatory context.

Practical recommendation: show original-language reviews alongside translations, label translations as machine-translated, and never let a translated review be the only artefact. Good UX and good compliance simultaneously.

What You’re Worrying About for Nothing

Retroactive: What to Actually Go Back and Fix

Ordered by real risk, highest first.

1. Any AI-generated reviews in your corpus. Not a labelling issue — remove them. This is the only genuinely urgent item and it has nothing to do with watermarking.

2. Your Merchant Center feed. A feed is not a historic artefact. It is a live declaration re-submitted every day. Descriptions bulk-generated in 2024 or 2025 that are still active are being submitted undeclared today. There is no grandfathering. Audit which SKUs have AI-origin copy and migrate those to structured_title / structured_description. This is the most impactful action in this article for the largest number of stores.

3. AI-generated images still in the feed. Confirm your image pipeline — CDN, resizer, compression — is not stripping IPTC DigitalSourceType. Most tooling strips metadata by default. Verify and fix the pipeline, not the individual images.

4. Your chatbot disclosure. Not retroactive, but almost certainly missing. Add the Article 50(1) notice this week.

5. Content generated before 2 August, published after. For text that falls under Article 50(4), the relevant date is publication, not generation. Check any scheduled YMYL content hub posts.

What not to do: do not add AI disclaimers to product pages as a precaution, do not rewrite copy to remove watermarks, do not embark on archive-wide relabelling. All three are wasted effort.

Practical Checklist

Feed / ecommerce ops

• Audit which SKUs have AI-origin titles and descriptions
• Migrate to structured_title / structured_description with digital_source_type: trained_algorithmic_media
• Verify your feed tool does not emit both title and structured_title simultaneously
• Confirm your image pipeline preserves IPTC DigitalSourceType

Content / SEO

• Apply the three-part Article 50(4) test to your content categories once; document the result
• Write down your editorial review process — named reviewer, sign-off authority, record of changes. This is the artefact the exemption rests on
• Review localisation workflow: is anyone reviewing machine-translated copy on YMYL topics?

Product / engineering

• Add the Article 50(1) AI disclosure to every conversational surface at first interaction
• Show source language alongside machine-translated reviews
• Add a provenance field to your PIM or CMS — human, ai-assisted, ai-generated, ai-translated — per content object. A boolean costs nothing to add now and is impossible to reconstruct later

Legal / leadership

• Confirm territorial scope if you sell into the EU
• Write the “a mark means processed, not authored” caveat into any internal policy before a detector result is ever treated as contractual evidence

The One Thing Worth Building

Record provenance at write time. Every content object in your PIM or CMS gets a field: human, ai-assisted, ai-generated, ai-translated. It costs almost nothing to add now. It is essentially impossible to reconstruct afterwards — no detector can tell you which of your 40,000 descriptions came out of a model in 2024.

Merchant Center already needs this. The AI Act’s editorial-control exemption is easier to evidence with it. Platform rules are converging on it. If you ever need to answer “which of our content is machine-written” — for a feed audit, a regulatory inquiry, a client contract, or an internal policy — this field is the only answer that doesn’t involve guesswork.

This article covers text only. AI-generated product images, IPTC preservation in detail, and C2PA for ecommerce are a separate topic. The parent article on watermarking mechanism and regulatory structure covers the full technical and legal background.