Overview

Amazon has introduced COSMO, an AI‑driven knowledge graph that replaces traditional keyword matching with a network of product, intent, and usage relationships. Launched in 2023, COSMO now powers search, recommendation, and discovery across the marketplace, making it essential for sellers to understand its mechanics to keep their listings visible and attract organic traffic.

Key Points

• Scale of the graph — COSMO contains more than 6.3 million nodes and roughly 29 million edges, dwarfing earlier e‑commerce knowledge structures.
• Behavioral signals — The system relies on co‑buy and search‑buy data, plus every view, add‑to‑cart, and browse event, to map how products relate in real time.
• Three relationship tiers — Product‑to‑use‑case, product‑to‑customer, and product‑to‑context connections drive how items appear in search and recommendation slots.
• Unified text analysis — Title, bullets, description, A+ content, and backend keywords are merged into a single text block for graph ingestion, demanding consistency across all fields.
• Intent‑stage coverage — COSMO evaluates awareness, consideration, interest, and purchase stages, rewarding listings that speak to each funnel point.
• Actionable shift — Sellers must move from isolated keyword stuffing to building a coherent web of entity relationships that describe what the product does, who benefits, and where it fits.

How COSMO Works

1. Data ingestion — Amazon extracts structured triples (behavior, relation, tail) from every seller listing, turning titles, bullet points, and backend keywords into a continuous string. For example, a listing titled “Lightweight Hiking Backpack for Long‑Distance Trails” generates a triple linking the product to the “hiking” use case and the “long‑distance” context.
2. Relationship mapping — Using 15 predefined relation types such as capableOf, usedOn, and isA, COSMO connects each node to others. A portable charger might receive edges like “usedOn → travel” and “capableOf → fast charging,” linking it to travel accessories and power‑need scenarios.