The three metadata standards inside every image
Image files can carry three overlapping metadata blocks. They were designed for different eras and different problems, which is why the field names don't line up neatly.
| Standard | Originally for | Carries | Matters for e-commerce? |
|---|---|---|---|
| EXIF | Cameras | Shutter, aperture, ISO, lens, GPS, timestamp | Rarely — mostly strip it |
| IPTC | News/stock photo desks | Description, keywords, creator, copyright, DigitalSourceType | Yes — this is the important one |
| XMP | Adobe workflows | Modern container that can wrap IPTC and custom fields | Yes — how IPTC usually travels |
EXIF is the one everyone has heard of and the one you mostly want gone. It bloats files and can leak information you didn't intend to publish — GPS coordinates from a phone shoot will happily tell the internet your home address if that's where you photographed your inventory. Strip EXIF from anything public-facing.
IPTC is where the useful information lives. It was built so a photo could travel between a photographer, a wire service, and a newspaper while carrying its own description, credit, and usage rights. That same structure turns out to be exactly what a product image needs when it travels between your DAM, your storefront, your feed, and Google.
Run a few of your live product URLs through an EXIF/IPTC viewer before doing any of this work. A large share of stores discover their images arrive at the browser completely bare — every field already gone. That changes what you should fix first.
The IPTC fields worth filling in
You don't need all 80-odd IPTC fields. For product imagery, a handful carry nearly all the value:
- Description (Caption/Abstract) — a natural-language sentence describing the product in the frame. Not your alt text verbatim, but closely related. This is the field most likely to be read by systems that index images.
- Keywords — product type, material, color, category. Think of it as structured tagging that travels with the file even when it leaves your CMS.
- Creator and Copyright Notice — your brand or studio, plus a rights statement. This is your cheapest defense when a competitor or scraper lifts your catalog.
- Credit Line — who should be attributed if the image is republished.
- Digital Source Type — the AI labeling field. Covered in detail in the next section.
The IPTC Photo Metadata Standard has also added properties specifically for generative content, including AI Prompt Information, AI System Used, and AI Prompt Writer Name. These are optional today and useful mainly for internal auditability — knowing which model and prompt produced an image is genuinely helpful when you need to regenerate a variant eighteen months later.
Write IPTC Description as a complete sentence a human could read aloud, not a keyword salad. "Charcoal merino wool crewneck sweater on a white background, front view" is worth more than "sweater wool charcoal merino crewneck white."
One caution on scope: metadata is a supporting signal, not a ranking lever. Alt text, structured data, file names, and page context still do the heavy lifting for product image SEO. Metadata is about provenance, rights, and compliance — treat claims that it will move rankings with suspicion.
Labeling AI-generated product images correctly
This is the section with real deadlines attached. The mechanism is an IPTC field called DigitalSourceType, which takes a value from a controlled vocabulary. Four values cover almost every product image scenario:
| Value | Use when |
|---|---|
digitalCapture | A camera took it. Standard retouching (color, dust, crop) doesn't change this. |
trainedAlgorithmicMedia | A generative model created the image outright. |
compositeWithTrainedAlgorithmicMedia | A real photo with AI-generated elements — the common e-commerce case. |
algorithmicMedia | Rendered by non-AI means, e.g. a 3D CGI render. |
That third row is the one most catalogs need and most get wrong. If you photographed a real product and used AI to generate a lifestyle background, swap in a model, or produce a ghost mannequin effect, the result is a composite — the product is genuine, the scene is synthesized. Labeling it trainedAlgorithmicMedia overstates the synthesis; labeling it digitalCapture understates it. Neither is accurate, and accuracy is the entire point of the field.
Not sufficient
- A disclosure line in your site footer
- "AI-assisted" in the product description
- A visible corner badge on the image
- A policy page nobody reads
Machine-readable
- IPTC DigitalSourceType in the file
- Signed C2PA manifest
- Invisible watermark (e.g. SynthID)
- All three, layered
The EU's approach is explicitly multi-layer: signed C2PA metadata, plus an imperceptible watermark, plus optional registry logging. C2PA and watermarking are the two layers most platforms are converging on. If your generation tool emits C2PA Content Credentials, preserve them — under the Code of Practice, deliberately stripping C2PA manifests from AI-generated images is prohibited.
Google Merchant Center has been ahead of the regulation here. Some AI-generated product images get disapproved specifically for missing the DigitalSourceType tag, which means this is already a revenue issue for feed-driven stores, not just a 2026 compliance project. Note that labeling and customer-facing AI disclosure are separate exercises — the metadata satisfies machines, your product page copy handles trust.
Why your CDN is deleting all of it
You can tag every image perfectly and still ship bare files. Most e-commerce image pipelines re-encode uploads to optimize file size, and metadata is dead weight to a compression pass. So it gets dropped.
Shopify merchants have reported IPTC metadata being stripped by the platform CDN, which is a known friction point for anyone trying to satisfy the Merchant Center DigitalSourceType requirement. WordPress and WooCommerce sit at the other end — WordPress stores the original file as uploaded, metadata intact.
Social platforms are the worst case and largely unfixable from your side. Instagram, X, LinkedIn, and TikTok all reprocess uploads through compression pipelines that destroy C2PA manifests, even as those same platforms announce plans to display Content Credentials. It's the standard's most conspicuous real-world limitation right now.
Practical workarounds, in rough order of effort:
- Tag before upload, always. Indexers frequently see an image in its original state before a platform pipeline finishes processing derivatives. Metadata applied after the fact may never be seen.
- Use a CDN you can configure. Services like Cloudinary let you explicitly preserve IPTC/XMP through transformations rather than defaulting to strip.
- Host the canonical original elsewhere. Keep a metadata-intact master on S3 or equivalent, and treat storefront derivatives as disposable.
- Verify in production, not locally. Fetch the actual delivered URL and inspect it. Local files pass tests that the CDN then fails.
Stripping metadata isn't purely a loss. Removing EXIF from public images is good hygiene for both file size and privacy. The goal isn't "preserve everything" — it's "strip EXIF, preserve IPTC/XMP and C2PA."
Building metadata into your catalog workflow
Metadata fails when it's a manual step at the end. It works when it's generated from data you already have.
You already store, per SKU: product title, category, material, color, brand. That's a complete IPTC Description and Keywords set sitting in your database. The workflow should template it:
- Define a template per image role. Hero, detail, lifestyle, and on-model images each get a description pattern populated from SKU fields.
- Set DigitalSourceType at the point of generation. Whatever produces the image knows how it was made — that's the only place the value can be assigned honestly. Retrofitting it later is guesswork. Tools built for e-commerce output, including Retouchable, should be evaluated on whether they emit correct provenance data, not just on image quality.
- Batch-apply with a CLI. ExifTool handles IPTC and XMP writing across thousands of files and scripts cleanly into an existing batch processing pipeline.
- Audit on a schedule. Sample live URLs monthly and assert that Description, Copyright, and DigitalSourceType survived. Pipelines change without telling you.
If you do only one thing from this article: pick ten live product image URLs, inspect what metadata actually arrives at the browser, and set DigitalSourceType correctly on anything AI-touched. That single field is the one with a regulator and a shopping feed behind it.