AI-Generated Product Images & Descriptions in Feeds: 2026 Policy Guide

Every e-commerce team I speak to in 2026 has already used AI somewhere in their catalog. A background swapped out. A lifestyle scene generated from a single pack shot. Three hundred product descriptions drafted in an afternoon. The tooling arrived faster than anybody's compliance process did.

And then comes the question, usually asked with a slight wince: is this going to get us disapproved?

The honest answer is that the question is framed wrong. There is no single "AI content policy" you can comply with. There are four separate rulebooks operating at the same time, and they ask different things of you:

1. Channel content policy — what the platform allows in a listing at all
2. Disclosure and metadata — how you must declare AI involvement, usually invisibly
3. Misrepresentation — whether the asset accurately depicts what ships in the box
4. Law — as of 2 August 2026, actual statute in the EU, with real fines attached

Most merchants who get into trouble are not breaking rule 1. They are compliant on the surface and failing on rule 2 through a pipeline bug they don't know they have, or failing on rule 3 because a generative model quietly changed something about the product.

This guide walks each channel and each rulebook, and ends with the operational checks that actually catch problems.

The short version

If you only read one section:

• Google allows AI-generated images and text. It requires machine-readable declaration — IPTC metadata on images, dedicated feed attributes on text. The most common failure is metadata being stripped somewhere between your generator and Merchant Center.
• TikTok Shop is the strictest. A fully synthetic main product render is not permitted regardless of disclosure.
• Meta allows it under general accurate-representation rules, with an AI label required on ad creative.
• The EU AI Act now adds a buyer-facing layer on top of all of the above for anyone selling into the EU.
• Almost every AI-related disapproval we see traces back to an ordinary image rule — a promotional overlay, a watermark, a resolution floor — not to the AI itself.

Rulebook 1: What Google actually requires

Images: the IPTC metadata rule

Google Merchant Center accepts AI-generated and AI-edited product imagery on both free listings and paid Shopping ads. It has done since early 2024. There is no ban to work around.

What Google requires is that the image carries embedded metadata declaring its origin. The mechanism is the IPTC DigitalSourceType property, and two values matter for e-commerce:

The requirement covers image_link, additional_image_link and lifestyle_image_link. Google's own guidance is blunt about the failure mode: don't remove embedded metadata tags such as DigitalSourceType from images created with generative tools.

That instruction is the whole ballgame, and we'll come back to it, because "don't remove it" turns out to be much harder than it sounds once an image passes through a real production pipeline.

Google's built-in Product Studio writes the tag automatically. If you generate images anywhere else — and most brands at scale do, because Product Studio has no batch processing and no way to hold a consistent visual style across a large catalog — writing and preserving that metadata is your responsibility.

Text: structured_title and structured_description

This is the part almost nobody has implemented, and it has been live since April 2024.

AI-generated titles and descriptions are not supposed to go into the ordinary title and description attributes. Google created dedicated attributes for them:

<g:structured_title>
  <g:digital_source_type>trained_algorithmic_media</g:digital_source_type>
  <g:content>Merino Wool Crew Neck Sweater - Men's, Charcoal, Size M</g:content>
</g:structured_title>

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

Two details that trip people up:

Omitting digital_source_type is itself a declaration. If you submit structured_title with only the content sub-attribute, you are telling Google the content was not AI-generated. Silence is not neutral here.

If you send both, the plain attribute wins. Where title and structured_title are both present, Google uses title. So populating the structured attribute while leaving your old AI-written title in place accomplishes nothing except a larger file.

Getting this right is a mapping problem, not a copywriting problem — which is exactly the kind of thing feed rules exist for. If your descriptions are generated in one system and your feed is assembled in another, the rule layer is where you tag provenance as the value moves through.

So does Google penalize AI content?

Not for being AI. Google's position on generative content has been consistent for years and it is about output quality, not authorship. The spam policy that catches merchants is scaled content abuse — publishing large volumes of low-value material primarily to manipulate rankings — and that policy applies identically to a human writing three hundred thin descriptions and a model writing them.

The policy that actually bites in Merchant Center is misrepresentation. And misrepresentation is where AI genuinely does raise your risk, for a reason that has nothing to do with policy and everything to do with how language models work: they fill gaps plausibly.

Ask a model to write a description for a jacket and give it a thin source record, and it will produce something confident about the fabric weight, the water resistance, the fit. Those sentences may be wrong. Your landing page says water-resistant; the generated copy says waterproof. That mismatch is a policy violation and a returns problem at the same time, and it will not look like an AI problem when it lands — it will look like a product data problem.

The mitigation is unglamorous: generate from complete source data, not from thin source data. A model constrained to real attribute values invents far less than one asked to be persuasive about a product it knows nothing about. Which makes catalog completeness a compliance control, and moves the work upstream into catalog enrichment.

Rulebook 2: The other channels

Meta

Meta's Commerce Policies govern catalog listings and turn on accurate representation of the product — there is no AI-specific prohibition on catalog imagery. The requirement appears on the advertising side, where Meta expects an AI content label on ad creative that is AI-generated or substantially modified.

The practical consequence for anyone running catalog ads: your feed image and your ad creative may sit under different rules even though they depict the same product. If you are generating dynamic creative from feed data — which is the Creative Suite use case — the ad-side labeling requirement is the one to check, and it belongs in your creative template configuration rather than in your feed mapping.

See our Meta integration guide for the current catalog field requirements.

TikTok Shop

TikTok is the strictest of the major channels and the one most likely to surprise a team that assumed Google's rules were representative.

TikTok Shop's Product Listing Policy prohibits placeholders and digital renderings of a product. A fully synthetic main product image is therefore not permitted — disclosure does not cure it. AI-edited photographs of real products are allowed, with disclosure required for substantial changes under AI-generated content rules that took effect in May 2026.

Where the line sits: light AI-assisted editing — color correction, cleanup, background removal — is generally treated as ordinary post-production. Disclosure kicks in when AI generates the substance of the image: a model who doesn't exist, a scene that never happened.

The design implication is real. If your creative workflow starts from a real product photograph and generates around it, you are inside the rules on every major channel. If it generates the product itself, TikTok is closed to you.

Turkish and regional channels

For merchants selling on Trendyol, Hepsiburada, Akakçe, N11 or Çiçeksepeti, published AI-content policy is thin to nonexistent compared with the global platforms. That does not mean the risk is lower — marketplace image guidelines on watermarks, overlays, borders and accurate representation still apply, and they catch most bad AI output on their own terms without ever mentioning AI.

Our working assumption for Turkish channels: comply with the strictest applicable global standard, because it is cheaper than maintaining two asset sets, and because these policies will converge. If you export to Akakçe or the Turkish marketplaces, the standard image rules in the next section are what you should be auditing against today.

Rulebook 3: The EU AI Act

This is new enough that most published guidance predates it.

Article 50 of the EU AI Act sets transparency obligations for AI-generated and AI-manipulated content, and those obligations became applicable on 2 August 2026. They are separate from the Act's high-risk regime — they apply regardless of risk tier whenever an AI system creates or manipulates image, audio or video content.

What matters for a product catalog:

It is buyer-facing, not just metadata. Google's IPTC requirement is machine-readable and invisible to shoppers. Article 50 requires disclosure that is clear, distinguishable and delivered no later than first exposure. A faint label on an image or a line buried in your terms does not meet the threshold.

Assistive editing is generally out of scope. Upscaling, color correction and basic lighting adjustment are ordinarily treated as assistive rather than generative. The obligation attaches where AI creates or materially alters the content.

Extraterritorial reach. Deployers established outside the EU are in scope where the output is used in the EU. A Turkish or UK merchant selling into the EU is covered.

There is a standard label available. In June 2026 the Commission published a free icon set for labeling AI-generated content, distinguishing fully AI-generated from partially AI-modified material. Adoption is voluntary but it is likely to become the practical benchmark.

Fines are material. Non-compliance with Article 50 carries penalties up to €15 million or a percentage of worldwide annual turnover, whichever is higher.

Google has moved to accommodate this. AI content label settings have been rolling out across Google Ads, Display & Video 360, Campaign Manager 360, Merchant Center and Ads Editor, and Google has confirmed that these labels will not be treated as violations of its own policies prohibiting text overlays and watermarks — an important carve-out, since a visible AI label on a product image would otherwise be exactly the kind of overlay that triggers disapproval.

Google is also explicit that using its label setting does not guarantee regulatory compliance. This is a question for your counsel, not your feed tool. What your feed tool should do is make it possible to apply the right label to the right subset of your catalog — which requires knowing which assets are AI-generated in the first place, and that brings us to the part that actually breaks.

Rulebook 4: The pipeline problem

Here is the failure mode we see most often, and it is invisible until a disapproval arrives.

Your generation tool writes correct IPTC metadata. Your image then travels:

generator → asset store → resize/crop job → format conversion → CDN → feed export

Metadata does not survive that journey by default. Image resizing libraries strip it. WebP conversion strips it. CDN transformation pipelines strip it aggressively, because metadata is bytes and their entire purpose is to remove bytes. By the time the URL in your image_link attribute resolves to a file, the declaration Google requires may be gone — and nothing anywhere in your stack will have raised an error.

You are now non-compliant on every AI image in your catalog, with a fully compliant generation process.

How to check. Pull a live image URL exactly as it appears in your feed — not the source file, the URL the channel actually fetches — and inspect its metadata:

# ExifTool, most reliable
exiftool -DigitalSourceType product-image.jpg

# ImageMagick
identify -verbose product-image.jpg | grep -i digitalsource

Test one image per pipeline path. If you serve different formats or sizes through different transformations, each path needs its own check, because they fail independently.

How to fix. Configure your transformation step to preserve IPTC — most tools support a copy-metadata flag that is off by default. Where the CDN is not configurable, re-inject metadata after transformation. And add a periodic verification pass on a sample of live URLs, because CDN configurations get changed by people who have never heard of DigitalSourceType.

This is also a reason to keep provenance in the feed record and not only in the file. If your feed knows which products carry AI-generated assets, you can act on that — segment them, label them, exclude them from a channel that doesn't permit them — independently of whether the metadata survived the pipeline. That is a feed rules job, and it is the difference between "we think most of these are fine" and an auditable answer.

What actually causes disapprovals

Worth stating plainly, because the anxiety in this topic is misallocated: image violations are consistently among the most common causes of product disapproval in Merchant Center, and the overwhelming majority have nothing to do with AI.

The rules AI output tends to break — because generative tools are optimized to produce attractive marketing images, and attractive marketing images are exactly what feed policies forbid:

That last row deserves emphasis. A generated hero image that never made it to the site is a mismatch, and it is one of the more common reasons a product stays disapproved after a merchant has "fixed" the image.

On the 500×500 change: Google has said it will auto-upscale some sub-threshold images using AI as a safety net, and that upscaled images will meet the requirement without action from you. Treat that as a backstop, not a plan. You do not want a machine deciding what your product looks like.

Running these checks across a full catalog by hand is not realistic. Our free feed audit tool flags resolution, overlay and attribute issues across your live feed, and Feed Protection catches them before an export reaches the channel rather than after.

A workflow that stays inside the lines

Generate around the product, not the product. Start from a real photograph. Replace backgrounds, build scenes, add seasonal styling — but leave the product pixels alone. This single constraint keeps you compliant on every major channel including TikTok, and it eliminates most misrepresentation risk, because the thing being sold is still the thing that was photographed.

Toggle-test before publishing. Flip between the generated image and the source photograph at full size. If a logo shifted, a seam disappeared, a color drifted, a strap changed, or a feature appeared — reject it. This takes seconds and catches the errors that cost you an account.

Never let a generated scene imply an accessory you don't ship. A generated kitchen scene around a kettle that includes a matching set you don't sell is a misrepresentation, and a returns problem, and a review problem.

Templates over one-offs. A reviewed, approved template applied across a catalog gives you one thing to audit instead of ten thousand. It is also, incidentally, how creative production stops being a headcount problem — Karaca produced 10,000 branded Black Friday creatives in two hours from a single approved template mapped to feed fields. The compliance argument and the efficiency argument point the same direction.

Keep provenance in the feed. Tag which products carry AI-generated assets and which carry AI-generated text. You need this for structured_title, for EU labeling, for excluding products from TikTok, and for answering the question when someone asks. Retrofitting it across a live catalog is painful; adding it at ingestion is nearly free.

Human review on text, always. Not a full rewrite — a fact check against real attribute values. The specific things to verify are the ones a model will confidently invent: material composition, dimensions, compatibility, water resistance, care instructions, what's in the box.

Verify metadata on live URLs, on a schedule. Not the source file. The URL the channel fetches. Quarterly at minimum, and after any CDN or image-pipeline change.

When not to use AI imagery at all

An honest section, because the answer isn't always yes.

When exact color is the purchase decision. Paint, cosmetics, textiles, furniture upholstery. Generative color drift is subtle, systematic, and it shows up in your returns data before it shows up in a policy notice. Vivense's furniture catalog is a good example of a category where fabric and finish accuracy across a variant matrix matters more than creative variety.

Regulated categories. Supplements, medical devices, cosmetics with efficacy claims. Any generated image that implies a physical result is a problem on every channel, and several platforms apply heightened scrutiny to these categories specifically.

When you can't answer "what changed." If your workflow doesn't produce a record of which assets were generated and how, you cannot comply with disclosure requirements, and you cannot respond to a policy challenge. Build the record first, then scale the generation.

Where the real photograph is already good. AI is worth using where photography doesn't scale — seasonal variants, campaign restyling, lifestyle contexts for a thousand SKUs. It is not worth using to replace a hero shot that works.

Where this leaves you

The compliance burden here is real but it is not conceptually hard. It is a data-lineage problem wearing a policy costume: know which assets are generated, keep that knowledge attached to the product record, make sure the declaration survives to the channel, and check the ordinary rules that catch bad output regardless of who made it.

The merchants who will struggle in 2027 are not the ones using AI. They are the ones who used AI at scale without recording what they did, and now cannot answer a simple question about their own catalog.

Sources

• Google Merchant Center Help — AI-generated content: https://support.google.com/merchants/answer/14743464
• Google Merchant Center Help — Product data quality violations: https://support.google.com/merchants/answer/13693497
• Google Search Central — Guidance on generative AI content: https://developers.google.com/search/docs/fundamentals/using-gen-ai-content
• European Commission — Code of Practice on Transparency of AI-generated Content: https://digital-strategy.ec.europa.eu/en/policies/code-practice-ai-generated-content
• EU Artificial Intelligence Act — Article 50 transparency rules: https://artificialintelligenceact.eu/transparency-rules-article-50/