Every Platform Speaks a Different Language

The Illusion

Product data is product data. A title is a title. A description is a description.

Every platform stores it slightly differently, but surely that's just a formatting problem?

It isn't.

The Tension

Shopify uses metafields. WooCommerce uses custom attributes. Salesforce Commerce has its own object model. Feed systems flatten everything into TSV columns.

Point an LLM at raw Shopify data and ask it to "fix" your catalog. It will hallucinate attributes. It will invent field names. It will produce inconsistent output across products because it doesn't know what "material" means in your catalog versus someone else's.

Monday's enrichment uses different attribute names than Friday's. The model doesn't remember. It doesn't have to. It generates plausible output, and plausible isn't the same as correct.

The Turning Point

The solution isn't a smarter model. It's a smarter constraint.

EKOM normalizes every product from every platform into a single, typed product object. Every attribute has a defined key, a validation rule, and a Schema.org mapping.

When an agent proposes a change, it proposes a change to a known attribute, not a freeform string. The schema makes the difference between "the agent is guessing" and "the agent is operating within defined boundaries."

The Risk

Without schema constraints, LLM output drifts. Not dramatically. Not in ways you notice on day one. But over hundreds of products and dozens of enrichment runs, the drift compounds.

Inconsistent attribute names. Conflicting values for identical products. Fields that exist in one enrichment batch but not another.

This is catalog entropy. And it's invisible until a downstream system (Google Shopping, an AI search engine, your own product pages) exposes the mess.

The Principle

The LLM is the brain. The schema is the spine.

Without structure, an AI agent is just a copywriter with API access. The canonical schema prevents hallucinated attributes, inconsistent enrichment, and catalog drift.

Twenty-four attribute rules across core and vertical modules. Required versus optional. Commercial versus structural. Confidence thresholds on every proposed fill. An agent can't propose a change to an attribute that doesn't exist in the schema.

This isn't over-engineering. It's the minimum viable constraint set for operating on production product data at scale.

The Future

Every platform will eventually need to speak the same language. Not because standards bodies decree it, but because AI systems that consume product data will reward consistency and punish ambiguity.

The schema isn't just internal plumbing. It's how your catalog becomes legible to machines.