Agentic Commerce: can an AI agent buy from your store?

Agentic Commerce answers one question: can an AI agent complete a purchase on your store without a human? It verifies the actions your store declares by calling them, detects agent commerce protocols (ACP, UCP, MCP), maps the purchase journey and produces a fix pack.

Updated 2026-11-21
PlanPro and Agency
UsesAbout 25 monthly credits a run
Time2 to 3 minutes
Open Agentic Commerce →

Step by step

  1. Open Agentic Commerce #

    Find it under Agents & Commerce, Commerce, in the left sidebar.

  2. Enter your store domain #

    Optionally name a product to shop for.

  3. Run it #

    The run narrates what it finds as it goes: declared actions, protocol probes and verification.

  4. Read the journey #

    See each purchase stage and whether an agent can pass it.

  5. Apply the fix pack #

    Work through the fixes for the stages that fail.

Why it matters#

Shopping agents are starting to buy on people's behalf. A store an agent cannot complete a purchase on is skipped, however good its products. This tests the whole path, not just the markup.

What it checks#

  • Declared actions collected from the store and verified by calling them.
  • Protocols: ACP, UCP and MCP detection.
  • Delegated authentication and the store's content policy for agents.
  • The purchase journey mapped stage by stage, with coverage per stage.
  • A fix pack for the gaps.

Reading the results#

  • Stages that fail show where an agent would stop.
  • Declared but broken actions are worse than none: they waste the agent's attempt.
  • No protocol support limits which agents can transact.
Agentic Commerce Readiness run summary
Example dataWhat the agent found at each stage, from reading the store to mapping the purchase, the readiness score, and how far an agent gets through the six purchase steps.

Implementing the fixes#

  1. Fix broken declared actions first.
  2. Make product, cart and checkout steps work without hover menus, pop-ups or custom widgets.
  3. Add structured product data (price, availability, shipping) per Product Feed Audit.
  4. Adopt the commerce protocol your platform supports.
What happened when an agent tried to buy
Example dataThe agent's own account of the attempt, step by step, ending at the step where it stopped. Each failure names what the store needs to declare.

Worked example#

A coffee roaster's store is tested from its home page and one product page.

Declared actions: a SearchAction that is verified by calling it, and an OrderAction that returns 404, because the target URL template points at an old checkout path. Protocols: none of ACP, UCP or MCP. Journey: discovery and product detail pass; add to cart and checkout fail.

The fix pack: correct the OrderAction target to the live cart URL, add offers with price and availability on product pages, and list the store's return policy in plain text. The broken action is the urgent one, since it wastes every agent's attempt. Re-run: add to cart passes; checkout still needs a protocol, which the roaster's e-commerce platform has on its roadmap.

Best practice#

Re-run after platform or theme updates; checkout changes are where agent paths break.

Troubleshooting

A declared action fails when called #

The URL in your potentialAction target is wrong, old or needs a login. Open it in a browser with the same parameters. Fix the target or remove the action: a broken declared action is worse than none.

No protocol is detected #

ACP, UCP and MCP support usually comes from your e-commerce platform or an app. Check what your platform offers; it is rarely something to build by hand.

The run did not place an order #

It never does. The test uses read-only actions only, so nothing is bought or changed.

Not listed here? See Common problems for limits, checks, plans and pages that cannot be read.

Questions

Why call the actions instead of checking markup? #
Valid markup can still describe an action that does not work. Calling it is the only proof an agent could use it.
What are ACP and UCP? #
Agent commerce protocols that let AI agents browse catalogues and check out on behalf of shoppers. MCP is the general protocol for exposing tools to agents.
Will the test place an order or change my store? #
No. It uses read-only actions only, from the home page and a product page, so nothing is bought, added or changed.

Something here does not match what you see in the app? Tell us and we will fix the guide.