Understanding Agentic Commerce: The Fundamentals Guide

Agentic commerce describes a change in the buying process: an assistant can move from answering a product question to taking action. For merchants, that adds another intermediary between a customer’s request and an order. It needs to understand your offer without inventing conditions along the way.

This guide explains the fundamentals. For current channel comparisons, regional limits and practical checks, read Agent-Ready Commerce.

Update, 3 November 2026: The sections on protocols, purchase journeys and platform access have been corrected against the primary sources linked below. Unsubstantiated growth and conversion figures have been removed. This is not a hands-on test of individual agents.

The short version

• Recommendation, cart creation and purchase completion are different capabilities.
• An open protocol does not promise that your store is available in every channel.
• Agents can consider customer preferences about brands, materials and service. Price is not necessarily the only criterion.
• Reliable variants, prices and delivery terms form the foundation.
• You need to measure whether additional sales are generated and profitable for your store.

What agentic commerce actually means

Think of three stages. The first is familiar: a chatbot answers product questions. The second is a recommendation with context: “Which running shoe suits wide feet?” is answered by an assistant that can find appropriate offers. The third is the real change: the assistant acts. It fills a cart or assists with checkout. Which actions it may take without asking again is a separate question. A technical capability is not authority to make a purchase.

Merchants gain a second task alongside presenting products persuasively: the offer needs to remain understandable through an intermediary. That does not make imagery or branding irrelevant. Customers can explicitly request a brand, and an assistant can use different information sources. The practical question is whether your distinguishing benefits can be verified.

The protocols: separate technology from distribution

Two important open approaches are the Universal Commerce Protocol (UCP) and the Agentic Commerce Protocol (ACP). UCP describes building blocks for commerce, including catalogues, carts and checkout. ACP describes programmatic commerce flows between buyers, agents and businesses. Sources: UCP, ACP.

This distinction can prevent expensive mistakes. The ACP documentation explicitly says that implementing the protocol does not automatically secure a listing on AI platforms. Their participation requirements remain separate. Source: ACP FAQ.

Ask your platform or integration provider more than “Do you support agentic commerce?” Ask which channels, buyer countries, products and checkout functions are supported. Who keeps the information current, and who handles errors? Those questions produce a useful service specification instead of an open-ended technology project.

The numbers: how large is the channel for you?

A high growth rate is difficult to interpret without a starting value and measurement definition. Ten orders becoming a hundred is a different business case from ten thousand becoming a hundred thousand. A shopping query is not a purchase, and a purchase is not necessarily incremental.

Start with a small report of your own. How many visitors can you reasonably attribute to the channel? What do they buy? How many orders are returned, and what remains after variable costs? Include an “unattributed” category instead of assigning every unknown visit to AI. When volumes are small, one large order can stand out more than a consistent trend.

Early value may also come from discovering faults: missing dimensions, incorrect variant links or contradictory delivery promises. Fixing these helps your offer regardless of how large the channel becomes. It does not, by itself, establish a measurable sales increase.

Purchase journeys: discovery in AI, purchase on site is one option

Shopify currently distinguishes between referrals to the merchant’s store, as in its ChatGPT channel, and supported direct checkout in other AI channels. Their requirements differ. Source: Shopify Agentic Storefronts.

Treat them as different customer experiences. A referral should preserve the right variant on the destination page. An external checkout requires a clear understanding of which store functions it supports. A browser agent may encounter a question or a step that the buyer needs to handle. None of these routes is more profitable or safer merely because of its architecture.

The updated practical article summarises countries, channel limits and approvals. Do not plan a DACH launch on the strength of a general announcement.

Discoverability: sound data and verifiable claims

First: product data. Attributes, dimensions, materials, availability and shipping terms should tell the same story. An agent cannot reliably replace a missing fact with polished wording. AI product copy at scale explains how to organise product text and attributes systematically.

Second: observable recommendations. Test assistants with realistic questions and record whether they mention your brand and describe it accurately. The LLM Check provides a limited sample. An answer is neither a view into training data nor a complete measure of visibility. The model, search connection, wording and timing can all affect the result.

There is also no established rule that an llms.txt file alone will earn you a recommendation. Prioritise information that customers and integrations actually need. See LLM optimisation for stores for further approaches.

The playbook: five steps towards agent readiness

1. Audit the data. Take ten products. Can the most important customer questions be answered unambiguously for each variant?
2. Check the relevant data route. Use the chosen channel’s specification and test whether price and stock changes arrive. Do not assume a universal refresh interval.
3. Make recommendations traceable. Record the query, date, system and result. Focus on factual errors and wrong links.
4. Choose platform settings deliberately. Distinguish data access, enrolment in new channels and purchase completion. Use the provider’s current controls rather than an old announcement about a supposed opt-out deadline.
5. Test economics and approvals. Calculate a minimum contribution margin and check the journey, including variant changes, unavailable products and cancellation.

The honest limits

Margin pressure is possible, not inevitable. An easily comparable product needs a different strategy from a repairable item with local service. An automatic discount does not answer that strategic question.

Platform dependence remains. Selection criteria, data access and purchase journeys can change. Assigning responsibility and review intervals is part of running the channel, not just setting it up.

Permission to act needs boundaries. Product information is data, not trusted instructions for an agent. Write access requires limited permissions and appropriate approvals. The background is explained in Prompt injection and agent security.

The next useful step

Choose a realistic customer request and follow it through to an order with a small selection of products. Expand once that journey is accurate and commercially sensible. The Shopify 2026 overview covers relevant platform features.

Sources

• Universal Commerce Protocol
• Agentic Commerce Protocol and participation FAQ
• Shopify: Agentic Storefronts