Overview

Amazon has begun replacing its long‑standing A9 search engine with a new model called COSMO, which relies on large‑language‑model semantics rather than simple keyword matching. The rollout started in early 2024 and now covers roughly ten percent of U.S. shopper traffic, forcing sellers to shift from keyword stuffing to context‑rich, intent‑focused content.

Key Points

• Semantic over keyword — COSMO evaluates the meaning behind a query, so a product can appear even when the exact search terms are absent from the listing.
• Knowledge‑graph layer — The algorithm links items to attributes, use cases and related concepts, enabling “wedding shoes” to surface without the phrase appearing verbatim.
• Personalization factor — Browsing history, demographic signals and recent behavior influence the order of results for each shopper.
• Two‑stage ranking — First, COSMO checks semantic relevance; second, it applies traditional performance metrics such as conversion rate and sales velocity.
• Early performance — In its initial test, COSMO delivered a 0.7 % lift in overall product sales and an 8 % rise in navigation engagement across the sampled traffic.

How COSMO Works

1. Query interpretation — When a buyer types “men’s shoes for a wedding,” COSMO parses the request to infer formal style, dress‑appropriate material and a hard‑sole silhouette, even if the listing only mentions “black leather dress shoes.”
2. Knowledge‑graph matching — The system consults a product graph that ties attributes (e.g., waterproof material) to real‑world scenarios (e.g., rainy outdoor use). A search for “home workout equipment for apartments” can therefore highlight foldable yoga mats, because the graph knows “foldable” aligns with limited space.
3. Personalized ranking — If the shopper previously purchased athletic gear and frequently browses fitness categories, COSMO boosts fitness‑related items higher in the results, tailoring the mix to that individual’s preferences.
4. Semantic relevance filter — Before any sales‑velocity or price‑competitiveness signals are applied, the algorithm discards listings that do not meet a minimum relevance threshold, ensuring only context‑aligned products proceed to the next stage.