Kuzushi Ads

Problem

AI agents (ChatGPT, Claude, Gemini, custom copilots) are becoming primary interfaces for information discovery, but there is no native advertising infrastructure for these conversational contexts. Traditional display ads don't work inside chat — they break the conversation flow and erode user trust. Merchants have no way to surface their products when an AI agent is recommending solutions.

Solution

Kuzushi Ads is a self-serve advertising marketplace and SDK that enables contextual, non-intrusive advertising inside AI agent conversations. Merchants list products, set budgets, and define targeting. The SDK integrates into any AI agent and uses semantic matching (pgvector embeddings) to surface relevant product recommendations within the natural conversation flow.

Key Technical Decisions

pgvector for semantic ad matching — Traditional keyword-based ad matching fails in conversational contexts because users express intent in natural language, not search queries. pgvector stores product embeddings alongside PostgreSQL data, enabling similarity search at query time. A user asking "what's a good running shoe for flat feet?" matches against product descriptions semantically, not via keyword overlap.

SDK-first architecture — Rather than building a single integration, the platform ships as an embeddable SDK that any AI agent developer can drop in. This means the ad marketplace scales with the ecosystem rather than requiring bilateral integrations with each agent platform.

Clerk for merchant auth — Merchant onboarding needs to be frictionless. Clerk handles OAuth, email verification, and organization management out of the box, letting us focus on the marketplace logic rather than auth infrastructure.

Impact

• Novel product category — first advertising platform purpose-built for AI agent conversations
• Semantic ad matching via pgvector embeddings with sub-100ms query latency
• Self-serve merchant dashboard with campaign management, budgeting, and analytics
• SDK architecture enabling integration with any LLM-powered agent
• Attribution pipeline tracking impression-to-conversion across conversation sessions