Overview

Amazon has introduced COSMO, an AI‑powered knowledge graph that replaces the legacy keyword‑centric A9 search with intent‑driven product discovery. The system, built on more than 6 million product nodes and tens of millions of relational links, began rolling out to shoppers in early 2024. Sellers who adapt their listings to the new intent model can capture higher visibility and conversion rates.

Key Points

• Scale of the graph — COSMO maps over 6.3 million distinct items and 29 million typed relationships, creating the largest semantic network Amazon has ever deployed.
• Intent over exact terms — Products are surfaced based on the problem a shopper wants to solve, even when the listing lacks the exact search phrase.
• Commonsense reasoning layer — A dedicated language model, COSMO‑LM, trained on 30 000 human‑validated annotations, infers logical connections such as “isFor” or “suitableFor.”
• Behavioral signals integration — Co‑purchase histories, browsing sequences, and review sentiment directly influence ranking decisions.
• Multi‑factor ranking — Relevance, contextual alignment, product quality, and commonsense appropriateness are combined to determine placement in search results.

How COSMO Works

1. Data ingestion — Amazon aggregates massive behavioral datasets, including items frequently bought together, click‑through paths from search to purchase, and time‑spent on product pages. Example: a shopper who views a hiking boot, then a waterproof jacket, and finally purchases a trail map contributes a sequence that COSMO records.
2. Hypothesis generation — COSMO‑LM processes the ingested data and proposes semantic links, such as “waterproof jacket → protects → rainy conditions.” Human annotators review these hypotheses to ensure logical consistency. Example: the model suggests that a portable charger is “capableOf → extending device battery life,” which annotators confirm.
3. Graph construction — Validated links become typed edges connecting product nodes. Common edge types include “isFor,” “capableOf,” and “suitableFor.”