From Product Data to Agentic Commerce: Building an AI-Ready Catalog that Scales

The talk, by a representative of Icecat (a product content syndicator), argues that as online shopping moves toward AI agents that buy on the user's behalf, product data must be structured, complete and verified, because language models cannot be trusted to fill gaps. It opens with a hypothetical: an AI agent orders a smart fridge, hallucinates a detail, and the fridge arrives 5 cm too big for the kitchen. The speaker says the talk will cover how AI reads data differently from humans, why verified product data is becoming infrastructure, and the technical setup Icecat uses. The speaker describes four stages of online shopping: search (keywords and ten blue links), compare (summaries, tables, filters), recommend (personalized suggestions, the current stage) and act (an agent completes purchase and payment). The central claim is that AI agents do not shop like humans. People are influenced by branding and imagery, and their brains fill gaps, for example assuming a smartwatch comes with the black strap shown in the photo. An agent does not guess; it relies on facts, fields and tags. In the smartwatch example, a retailer with missing fields such as battery life would simply be left out of results when a user asks for at least 18 hours. The speaker adds that LLMs are text-prediction engines, not product databases. Wrong data can mean more returns, lost revenue and damaged trust. The proposed fix is "real time grounding," presented as a three-layer architecture: - A product truth layer of verified, brand-supplied data (Icecat's Brand Cloud or manufacturer sources). - A grounding layer using MCP, an API bridge the speaker compares to a librarian who fetches the right book rather than answering from memory. - A commerce execution layer, the Universal Commerce Protocol (UCP), covering delivery, payment and transactions. The talk does not go into this layer. The speaker lists three data requirements for AI use: attributes mapped to standard taxonomies so that, for instance, dimensions mean the same across brands; complete fields; and the ability to translate user intent into specs. In the dishwasher example, a parent wanting a quiet evening cycle in an open-plan apartment maps to noise level, installation type and dimensions. On whether to invest now, the speaker argues that centralized, standardized data already feeds retailers and marketplaces (Allegro and Amazon are named) and so supports human shoppers too. This is the speaker's own, vendor-side argument. The talk ends with a demonstration route: connect Claude to Icecat's MCP server and ask product questions answered from Icecat's grounded data, with a QR code to the documentation.