Ruy Ramos

Most e-commerce personalization happens around the edges of the experience. A “recommended for you” carousel appears below the product grid, an email suggests something based on an old purchase, or a separate recommendation engine tries to predict what somebody might click next.

Meanwhile, the actual experience remains generic, stale, and non-personal.

Every shopper entering a category sees roughly the same products in roughly the same order. Search relevance might improve the list, merchandising rules might push bestsellers or campaigns higher, but two customers with completely different needs still walk into essentially the same digital store.

Rank by Taste changes that model.

With vector ranking in the Crystallize Discovery API, products and shoppers can be described using the same meaningful vocabulary. Discovery compares those profiles and uses the resulting similarity as another ranking signal. Stock, margin, sales velocity, campaigns, recency, search relevance, and other business-critical signals can still factor into the same decision.

The result is not another recommendation block. The same Product Universe can produce a different shopping experience for every shopper without creating a different catalog for every shopper.

Ecommerce search has spent decades getting better at translating keywords into products. We added autocomplete, facets, semantic search, recommendations, and eventually generative AI. Each improvement made discovery easier, but the interaction stayed mostly the same: the shopper describes a product, and the store returns matching products.

Agentic commerce changes that relationship. A customer asking, “I need a lightweight jacket for cycling to work, preferably waterproof, under €250, and nothing too warm because I already run hot,” does not really have a search problem. They have an outcome, several constraints, and a decision to make. The interface now has to understand intent, inspect product data, reason over the constraints, and help move the customer toward a decision.

That is the shift demonstrated when Claude Commerce connects to Crystallize: the AI isn't simply sitting on top of a storefront. It is operating against the commerce backend.

AI does not need another product description explaining that Midnight means black, Pro means the expensive version, and usually ships quickly means somewhere between tomorrow and the end of the universe.

It needs product data where identity, attributes, relationships, variants, price, availability, language, and context are explicit.

That is what machine-readable product data means: product information structured so crawlers/agents/and who knows who else can reliably identify, query, compare, filter, transform, and act on it without first interpreting marketing copy.

And today, this matters beyond traditional SEO. Search engines consume structured data. Shopping platforms consume feeds. Applications consume APIs. Retrieval systems consume structured metadata and embeddings. AI agents increasingly consume tools and resources through protocols such as MCP.

The mistake is in trying to create a separate “AI catalog” for each of them.

A better architecture starts with one canonical product model and projects it into the formats different machines need: