14,000 Recipes, 10 Dimensions: How AI Enrichment Makes Product Catalogs Agent-Ready
Betty Bossi is the first partner in the Nexbid ecosystem. 14,391 recipes were enriched with 10 targeting dimensions in under 24 hours — initially via local LLMs (Ollama), ongoing maintenance via Gemini Flash. A blueprint for catalog readiness.
Less than 10% of product catalogs in Europe are AI-ready. This means AI agents cannot meaningfully process the data — no structured signals, no targeting information, no machine-readable semantics. For agentic commerce, that's a dealbreaker. With Betty Bossi, we've completed the first full enrichment run in the Nexbid ecosystem. The result: 14,391 recipes enriched with 10 targeting dimensions — in under 24 hours, fully automated. The Problem: Unstructured Product Data A typical recipe entry at a food publisher contains: title, description, ingredient list, preparation instructions. That's enough for a human — but not for an AI agent searching for a dinner for a family with two kids, vegetarian, under 30 minutes, budget-friendly. This kind of contextual matching requires structured signals. That's exactly what Nexbid Enrichment delivers. 10 Targeting Dimensions Each recipe is enriched with the following dimensions: Cuisine Style — Swiss, Italian, Asian, Fusion, etc. Meal Type — Main Course, Appetizer, Dessert, Snack, Breakfast Difficulty — Easy, Medium, Advanced Budget — Budget-friendly, Medium, Premium Time Category — Quick (<30 min), Medium (30-60 min), Long (>60 min) Cooking Method — Baking, Frying, Boiling, Grilling, Steaming, Raw Equipment Needed — Oven, Stove, Mixer, Grill, No special equipment Season — Spring, Summer, Autumn, Winter, Year-round Audience — Families, Singles, Couples, Kids, Guests Occasions — Everyday, Celebration, BBQ Party, Christmas, Easter The Tech: Local LLMs + Gemini Flash Maintenance The initial enrichment run was performed entirely with local open-source LLMs (via Ollama) — no cloud dependency and no external API costs. Ongoing maintenance and re-enrichment uses Google Gemini Flash — optimized for speed and cost at under CHF 5 per run. Processing runs distributed with automatic retry and cross-model evaluation. This hybrid approach — local for the initial effort, cloud for maintenance — makes the technology affordable even for smaller publishers while remaining privacy-compliant. Result: Bilingual, Complete, Immediately Usable The Betty Bossi catalog comprises 7,194 German-language and 7,197 French-language recipes — matching the bilingual Swiss audience. Enrichment quality was verified using a cross-model evaluator: hard dimensions (budget, difficulty) achieve over 85% agreement; more subjective dimensions (occasion, audience) are expectedly lower. The result is live: At nexbid.dev/recipes, all enriched recipes can be browsed and filtered by all 10 dimensions. What This Means for Agentic Commerce Enrichment is Layer 3 in the 6-Layer AI Commerce Ecosystem — Catalog & Data Readiness. Without this layer, agents can find products (Layer 1: Visibility) but cannot intelligently filter and …
- Author
- Holger von Ellerts
- Published
- 2026-04-02
- Topics
- Enrichment, Agentic, Agent Discovery