Amazon's AI Image Tag Doesn't Survive Your Export Pipeline
๐Ÿ“ข
← Back to Blog

Amazon's AI Image Tag Doesn't Survive Your Export Pipeline

John Aspinall · · 8 min read

If you use AI-generated people in your listing images, your compliance step is not a checkbox in Seller Central. It is a metadata keyword that has to be embedded inside the image file before you upload it. And here is the part almost nobody is saying out loud: that metadata is one of the most fragile things in a creative pipeline. Save for Web drops it. Export presets drop it. Resize scripts drop it. Third-party upload tools re-encode and drop it.

Which means you can do the compliance work correctly at the design stage and still ship a non-compliant file โ€” because a preset somebody set up two years ago quietly deleted the only evidence that you complied. Nothing errors. Nothing warns you. The image looks identical.

That is the whole post. Everything below is the detail.

What happened

Amazon now requires sellers to tag product images and A+ Content containing photorealistic AI-generated people, using an IPTC-compatible metadata editor to add the keyword contains-synthetic-performer to the dc:subject (XMP) field before upload โ€” reported by CNBC on July 23, 2026 and Forbes on July 25. The requirement follows New York General Business Law ยง396-b, the synthetic performer disclosure law that took effect June 9, and Amazon says it will display an indicator to shoppers on affected listings.

I covered the rule itself and the creative tradeoff on July 26. I'm not re-explaining the mechanics. This is about the two things that post didn't cover: where the tag actually dies, and what the shopper-facing indicator does to a frame you spent money controlling.

Why most brand owners will read this wrong

The dumb take is "Amazon is cracking down on AI images."

It isn't. Read the exclusions, because they're wide and they're the most useful part of the whole rule. You do not need the tag for images containing only real people, even if you used AI tools to edit them. You do not need it for non-photorealistic people. You do not need it when there are no people in frame. Background generation, lighting correction, and generating a lifestyle scene around a real product photo are all fine, untagged.

So the vast majority of how brands actually use AI in listing creative right now โ€” cleaning up a background, building a kitchen scene around a real product shot, fixing the lighting a photographer got wrong โ€” is explicitly outside the rule. If you were about to re-shoot a catalog out of caution, stop. You're solving a problem you don't have.

The real signal is narrower and less comfortable: the one group genuinely in scope is the group least likely to catch a silent metadata failure. Brands using photorealistic synthetic models are, by definition, brands that replaced a photoshoot with a generation-and-export workflow. That workflow is exactly the kind that runs files through three or four tools between creation and upload, and every one of those handoffs is a place where an XMP block gets discarded.

I want to be precise about what I'm claiming here, because I can't verify the other half. I have no visibility into what Amazon's ingest pipeline does with metadata, whether it re-encodes, or how it handles a file that arrives stripped. Nobody outside Amazon does, and I'd be suspicious of anyone who tells you they do. What I can verify โ€” and what you can verify in five minutes โ€” is whether the keyword survives your pipeline. That's the part you own, and it's the part that's broken in more shops than anyone realizes.

What actually changes if you're running $200K/mo

Two things, and they land in different places on the P&L.

The compliance step moved to the end of the process, and nobody moved it. In most shops the person who knows the image contains a synthetic model is the designer, and the person who uploads is an ops coordinator or a VA working from a folder of finals. If the tag gets applied in the working file and the export strips it, the only human who could catch it is three steps upstream and never sees the uploaded asset. Across a catalog with 40 lifestyle images, you don't find out by auditing โ€” you find out when something gets flagged.

The indicator is an element in your frame that you did not design. I have optimized more than 14,000 hero images. Every element in the frame got argued about โ€” the crop, the type size, whether the logo earns its corner. Amazon is now going to render a disclosure indicator on affected listings, and you had no input on its placement, size, or wording.

I'm not going to give you a CTR number for that, because there isn't one and I'd be making it up. What I will say is that it changes a decision that used to be purely about cost and speed. Synthetic models got adopted because a lifestyle shoot with real talent runs $3,000โ€“$8,000 and takes three weeks, and a generated model takes an afternoon. That math was clean. It's now: an afternoon, plus a per-asset metadata step your pipeline may silently undo, plus a shopper-facing indicator on a page where you control everything else. For a lot of brands that's still the right trade. For apparel, beauty, and anything where the model is the aspiration being sold, it's worth re-running the comparison rather than assuming last year's answer holds.

What I'd do this week

1. Run the survival test. Take one image. Add contains-synthetic-performer to dc:subject with an IPTC-capable editor. Now push it through your actual production path โ€” the real export preset, the real resize step, the real upload tool. Re-open the exported file and check whether the keyword is still there. Five minutes, and it either survives or it doesn't. Most people have never checked because it never occurred to them there was anything to check.

2. Inventory what's genuinely in scope. Not "which images used AI" โ€” which images contain a photorealistic AI-generated person. Given how wide the exclusions are, this list is usually much shorter than the panic suggests. Some brands will find it's zero.

3. Move the tagging step after final export, not before. If your pipeline strips metadata, tag the file that actually gets uploaded. This is a workflow change, not a tooling purchase.

4. Put the question to your creative vendor in writing. "Does your delivery process preserve XMP metadata?" If you use an agency or a freelance retoucher, the files land in your hands already exported by someone else's preset. A vendor who can answer that has thought about it. A vendor who answers with a paragraph about their commitment to compliance has not.

5. Make the synthetic-model decision deliberately rather than by inertia. If you're in scope, you're now choosing between an indicator on your listing and a real shoot. Choose it on purpose.

What I'd ignore

The "label everything by August 15" claim. There's a version of this story circulating that says Amazon requires labeling of all AI-generated or AI-assisted content โ€” images, A+ copy, bullets, product descriptions โ€” with hard enforcement starting August 15 and escalation to forced removal. I went looking for it and I could not source it to Amazon. It traces back to aggregator posts that cite an "official announcement" while reproducing no Amazon language and linking to no policy page. It also contradicts the well-sourced version on the specific points that matter: the reported rule explicitly exempts AI-edited images of real people, and Amazon's general content policy does not require disclosure that copy was AI-generated.

I'm not telling you it's false. I'm telling you I couldn't source it, that the version with CNBC and Forbes behind it says something narrower, and that the correct move is to open your own Seller Central notifications rather than reprice your creative roadmap off a blog post. If Amazon sent you something with a date on it, that's your source. If they didn't, you don't have one.

The "is AI creative dead" discourse. It isn't, the rule barely touches most of what brands do, and posts arguing about it produce zero decisions.

Reverse-engineering how the indicator will look. Nobody has seen it at scale. Speculating about its visual weight before it renders is the same energy as optimizing for a device that ships in 2027.

The rule here is small and narrow and mostly exempts you. The failure mode is neither of those things: it's a metadata block getting quietly discarded by a preset, on the one category of asset where you're actually on the hook. Go check whether yours survives.

Put AI to work inside the business you already run.

The Aspi OS Bootcamp is a 4-week live build: second brain, Claude Code workflows, Codex execution — on your real business. The next cohort is forming now.

Get first access →

Not ready? Get the free newsletter — the AI workflows I actually ship, when they're worth your inbox.