What AI-ready product data is, and what the gap costs

There is a version of 2026 where nobody finds your products. The products are fine. The systems reading your catalog can’t make sense of them.

Search engines read structured data and product feeds alongside HTML. Marketplace search takes listing completeness into account. AI purchasing agents compare products by querying specifications, not by reading your marketing copy.

Regulators are starting to ask for machine-readable product data, and many catalogs couldn’t produce it. This isn’t a data quality problem. The systems that read your product data changed, and the systems that produce it didn’t.

AI commerce protocols exist: Google’s Universal Commerce Protocol (UCP) powers checkout in AI Mode for eligible US merchants, and OpenAI’s Agentic Commerce Protocol (ACP) now runs behind merchant apps in ChatGPT. Both lean on structured product data. The EU battery passport applies from February 2027 to EV and light-transport batteries, and to industrial batteries over 2 kWh; textile passport rules are expected to be adopted around 2027 and apply later. And the Model Context Protocol (MCP) lets an AI system call a catalog as a set of structured tools, so the tooling exists and the data is often what holds it back.

Your catalog was built for one reader.

Many catalogs exist to render a product page for a person. They hold just enough to fill a template: title, description, price, a few images. That was enough when the only reader was your webshop CMS.

AI systems read differently. They compare typed attributes across products, follow the relationships between them, and weigh completeness and consistency. Here is what they find in many catalogs:

None of these is a content gap. Better descriptions won’t close them, and a one-time cleanse won’t hold. What is wrong is the data model, the boundaries between systems, and the way work moves through them.

Clean data and AI-ready data aren’t the same thing.

“Clean data” suggests the problem is dirt: duplicates, typos, stale records. Scrub it once and you are done. AI-readiness asks for something else.

An LLM, an autonomous agent, a recommendation model, a purchasing bot: each one needs typed, structured, contextual product data it can reason over, and something that checks and controls what enters it. Specifically:

• Agents depend on context they can trust. One that recommends a product needs to know the record behind it is current and checked, not a spreadsheet export from 6 months ago. Trust is a property of the system, not of the row.
• LLMs break on inconsistent data. The same product can carry different specifications in your webshop feed, your Amazon listing and your ERP. A system that reads all three either invents one answer or refuses to surface the product. Consistency has to be enforced by the system, not by people remembering.
• Nothing reasons over prose. A language model can read “lightweight, great for running, comfortable fit.” It can’t reliably pull weight=240g, use_case=running, cushioning=high out of that sentence. Structured attributes aren’t optional enrichment. They are the input format.
• Multi-step runs need machine-readable data. An agent that checks whether a product meets a marketplace’s compliance rules, writes locale-specific copy, or holds a release until the record is complete can’t work from free-text fields. It needs typed schemas, validation rules, and relationships between products.

The cost shows up under other names.

Many companies pay for unready product data every day and file it under something else. Here is what the symptoms trace back to:

Listings come back rejected.

Say your team submits 120 products to Amazon DE and a third come back rejected: missing EAN, wrong category node, a description over the character limit, no material and no care instructions. The team fixes them by hand in Seller Central.

The cause: nothing in the PIM checks a product against Amazon DE’s required fields before export.

CataZenta checks Amazon and Shopify pushes for required fields, allowed values, lengths and units before sending; products that fail stay in the job, with the reason.

The AI writing tool produces filler.

You buy an AI content generation tool. The descriptions come back generic, the specifications invented, the brand voice inconsistent. The vendor tells you to write better prompts.

The tool has nothing structured to work from. It is writing from a product name and a price, because that is what your catalog reliably holds.

Compliance can’t be checked by a machine.

The EU General Product Safety Regulation (GPSR) has applied since December 2024. Your compliance team has to show that 8,000 products carry responsible person information, safety warnings and the right classification.

The data exists, in dozens of PDF documents, 3 spreadsheets from different suppliers, and one email thread from legal. Nothing can validate it programmatically, because none of it is structured or tied to a product identifier.

The launch date slips.

Say a new collection is meant to go live in a fortnight and takes three times as long. Content comes from 3 teams (product, marketing, compliance), translation into 4 languages runs as a chain of handoffs, and media assignment needs a creative brief per product.

Each channel gets its formatting by hand, in that channel’s own admin. Nothing runs in parallel, and nothing starts on its own when a product becomes complete.

You are missing from the AI answer.

Your competitors appear in Google AI Overviews for category queries. You don’t. Your products have titles and descriptions but no typed battery_life_hours, no quantitative noise_reduction_db, no machine-readable connectivity_type.

The overview compares products on named attributes. Yours has none it can pull out.

Recommendations make no sense.

Your recommendation engine puts budget phone cases under premium headphones as “related products.” Nothing in the catalog says these two go together, or that one is the higher-tier version of the other. The model is guessing from purchase correlation, because your catalog gives it no structure to read.

This is a design problem, not a content problem.

The instinct is to enrich the catalog. Hire more editors. Run a data cleanse. Buy an AI content tool. That misses what is broken.

Your data isn’t dirty. Your systems were never built to produce, validate and maintain structured product data as a standing job. AI-readiness asks for five layers, set out below.

So this is a system design, not a project. Treat it as a cleanup and you will clean it up, watch it decay, and do it again. Build the structure and the catalog improves on its own, because the system holds the model, checks changes, and flags gaps for someone who can close them.

Many PIMs were designed for a person filling in a form. An AI-native PIM is built the other way round: an assistant’s tool calls run inside the signed-in user’s permissions, and by default agent suggestions wait for a person.

What the working version looks like.

Moving from a catalog people edit to product data AI systems can operate on takes these layers:

1. A structured model, built on families. Every product belongs to a family that sets its attribute schema: required fields, types, units, validation rules. A Wireless Headphone family asks for driver_size_mm (number), noise_cancellation (boolean), battery_life_hours (number) and connectivity (select: bluetooth_5.0, bluetooth_5.3, multipoint). Typed fields are what make a product comparable by machine.
2. One authoritative system, reachable by program. Changes land there, whether a person, an AI enrichment agent or an import pipeline makes them. Channels, AI systems and export jobs all read from it. It exposes that data over an API and over MCP, so nothing has to scrape a screen or wait for a CSV export.
3. Completeness scored per channel, with a gate before publish. The system should know what “ready for Amazon DE” means: German title, 5 bullets, EAN, category node, material, care_instructions, 5+ images. It should score every product against every target channel, continuously. Below the threshold the product shouldn’t publish, and the gate should sit in the system rather than on someone’s checklist.
4. Validation and review on the way in. Attribute changes from editors, AI and bulk imports should be checked against the attribute schema: types, units, business rules. By default, AI-generated suggestions should wait in an approval queue until a person reviews them.
5. Multi-step runs that need nobody in the middle. When a supplier file arrives, an agent should extract the attributes and the validation pipeline should check quality; exceptions should become review tasks, and approved products should publish to the channels they qualify for. People should define the rules and handle the exceptions, and the system should do the running.

Find out where you actually are.

Before you invest, measure your current state honestly:

Level 1-2: the work is structural. Build the data model, and settle on one system that everything else reads from. No AI investment holds up without it.

Level 3: the work is operational autonomy. Build the enrichment pipelines, the gates before publish, and the completeness scoring that let AI do work you can rely on.

Level 4-5: the work is keeping it that way as channels and rules change.

Start with one family, one channel, 90 days.

You don’t have to restructure the whole catalog at once. Start with a wedge that proves the architecture:

1. Pick your highest-value product family, the top 50 SKUs by revenue
2. Import it into a structured PIM, with proper family modeling and typed attributes
3. Write down what complete means for one priority channel: Amazon DE, your webshop, whichever one rejects the most
4. Run AI enrichment on the structured data, with validation and a review step
5. Measure completeness before and after, listing acceptance rate, time to publish, and channel conversion
6. Use those numbers as the business case for the next family, and the one after

The pilot answers two questions. Whether structured product data measurably changes the numbers. And whether the model, the review rules, the validation and the AI enrichment hold together as a system, rather than as one more cleanup.