How Does Agentic Commerce Work?

Behind Agentic Commerce lies a complex interplay of AI models, protocols, APIs, and payment systems. This article explains the technical functionality — from the user's request to the completed transaction.

The Typical Process in 6 Steps

1. Request: The user describes in natural language what they need. "Find me a birthday gift for my sister — she likes tea and ceramics, budget max $50."
2. Planning: The AI agent analyzes the request, identifies relevant product categories, and creates a search plan: Which shops to search? Which criteria to prioritize? What information is still missing?
3. Research: The agent searches product feeds and product catalogs from multiple shops via standardized interfaces (ACP, UCP, MCP). It filters, compares prices, checks reviews and availability.
4. Recommendation: The agent presents the user with a curated selection — with reasoning for why each product fits. "I found three options: A handcrafted tea set for $45, a Japanese teapot for $38..."
5. Confirmation: The user selects a product and confirms the purchase. The agent creates a checkout session with the merchant, displays the final summary (price, shipping, delivery time), and waits for purchase approval.
6. Transaction: The agent completes the purchase — including tokenized payment — and delivers the order confirmation. The user never left the chat.

Technical Layers

Agentic Commerce is built on multiple technological layers that work together:

Layer 1 is the existing e-commerce infrastructure: online shops, inventory management, payment providers, logistics. This layer does not fundamentally change.

Layer 2 consists of the new protocols — ACP, UCP, AP2, MCP. They standardize how the layers above communicate with the commerce infrastructure.

Layer 3 is Tool-Use: The ability of modern LLMs to call external APIs via Function Calling. An LLM "knows" that a checkout API exists, what parameters are needed, and how to interpret the response.

Layer 4 is the agent itself — an LLM with planning, memory, and decision-making capabilities. It orchestrates the entire purchasing process and decides which tools to use in which order.

Layer 5 is the interface to the user — currently primarily chat (ChatGPT, Google Gemini), prospectively also voice assistants or embedded agents in other applications.

Structured Data as the Foundation

Agents don't read HTML. They need machine-readable data. The most important formats:

• Schema.org / JSON-LD: Standardized product information that can be embedded in any website. Already used by Google today — for Agentic Commerce it becomes mandatory.
• Product Feeds: Structured product catalogs in ACP or UCP format. Similar to Google Shopping Feeds, but specifically designed for AI agents.
• API Endpoints: REST or GraphQL interfaces through which agents can query product data and execute actions.

Shops that don't provide structured data are invisible to AI agents. This is the fundamental difference from traditional SEO: Google could index websites even without Schema.org — an AI agent cannot shop without machine-readable data.

MCP: The USB Port for AI

The Model Context Protocol (MCP) by Anthropic plays a special role in the Agentic Commerce architecture. MCP is not a commerce protocol — it is the infrastructure layer that connects AI models with the outside world.

The USB metaphor explains it best: Just as USB serves as a universal port connecting various devices to a computer, MCP connects various tools and data sources to an AI model. A shop provides an MCP server — any MCP-compatible agent can use it, regardless of the LLM being used.

Shopify already uses MCP: The Shopify MCP server gives agents access to product catalogs, checkout, and order management. For merchants, this means: One integration, reachable by every AI system.

Multi-Agent Systems

In the future, Agentic Commerce will not be handled by a single agent, but by specialized agents working together:

• Research Agent: Searches product catalogs and collects options
• Comparison Agent: Systematically compares prices, reviews, and features
• Negotiation Agent: Searches for coupons, negotiates terms (in B2B scenarios)
• Checkout Agent: Handles the transaction via ACP or UCP
• Monitor Agent: Monitors prices after purchase and reports price drops

Google's Agent-to-Agent Protocol (A2A) and UCP already account for this multi-agent architecture in their design.

Agents Learn: Reinforcement & Feedback

AI agents improve over time — not through one-time training, but through experience:

• Implicit feedback: Which recommendations does the user accept? Which does the user reject? The agent learns preferences.
• Quality signals: Which shops deliver reliably? Where are there problems? Agents could build shop ratings from their own experience.
• Collective learning: Anonymized data from millions of transactions helps identify the best products and shops.

For merchants, this means: Reliability and data quality are directly rewarded. A shop that offers correct prices, fast delivery, and hassle-free returns will be recommended more frequently.

Generative Engine Optimization (GEO)

While SEO optimizes content for search engine rankings, GEO (Generative Engine Optimization) aims to be used as a source by AI systems and leveraged for recommendations.

In practice, GEO means:

• Clear, citable definitions: Opening paragraphs that an LLM can directly use as an answer
• Structured data: Schema.org, JSON-LD, Product Feeds — machine-readable, not just human-readable
• Factual accuracy: LLMs prefer sources that are consistent and verifiable
• Authority: Backlinks, mentions in trade media, and consistent information across multiple sources increase the likelihood of being cited

For merchants, GEO becomes the second optimization discipline alongside SEO. Those who want to appear in the recommendations of ChatGPT, Gemini, or Claude must optimize their product data for AI systems.

Example: A Concrete Purchase

Here is what a typical Agentic Commerce process could look like in practice:

Total duration: under 2 minutes. No tab switching, no checkout form, no account creation.