Agentic Commerce · November 8, 2026 · 12 min read

Agentic Commerce Terminology: A Working Glossary

Agentic commerce terminology is the vocabulary used to describe AI agents that research, compare and buy on a customer's behalf. The terms that matter fall into four groups: the agent itself (agent…

Quick answer. Agentic commerce terminology is the vocabulary used to describe AI agents that research, compare and buy on a customer’s behalf. The terms that matter fall into four groups: the agent itself (agent, tool call, MCP), what it reads (product feed, structured data, offer, availability), how it transacts (checkout protocol, payment mandate, ACP, AP2), and how you measure it (agent referral, conversion by assistant). Most of the vocabulary is stable; the checkout protocols are not.

What matters most

Agentic commerce vocabulary splits into four jobs: what the agent is, what it reads, how it pays, and how you count it.
Product feed quality and Offer structured data do most of the work; checkout protocols matter least today because they are not widely live.
llms.txt is a convention with no committed consumer — cheap to add, close to worthless on its own.
Agent traffic cannot be fully attributed yet; server logs are more reliable than an analytics dashboard.
Strong customer authentication and liability are the practical blockers on agent-led checkout, not product discovery.

What does agentic commerce actually mean?

Agentic commerce is the use of AI agents to research, compare, choose and sometimes buy products on a customer’s behalf. The agent works from a goal and a budget the customer gave it. Your store is one of several it can use, and it will leave quickly if the data it needs is missing.

Three different things get called an agent, and each needs different work from you. A chat assistant answers a question and links out. A browsing agent opens your storefront in a browser and clicks through it like a person. An API agent calls your endpoints directly, through tools that you or a protocol define. The first needs good written content. The second needs a storefront that survives without JavaScript surprises. The third needs documented interfaces and stable identifiers.

Agent: software given a goal and a set of tools that decides the steps itself.
Tool call: one structured request from an agent to an outside function, such as a product search filtered by size.
Agent surface: anywhere an agent can read or act — product page, feed, API or MCP server.
Agent readiness: how far an agent gets on your store without a human stepping in.
Human in the loop: the customer confirms a step the agent proposed, such as approving a substitution.

Where to go next

Which terms describe what an agent reads?

Most of this vocabulary predates agentic commerce. Product feed, structured data, offer and availability are old terms. What changed is tolerance for gaps. A shopper fills a missing size in from the photo. An agent usually drops the product and moves to the next result.

Two newer terms come up in almost every vendor conversation. llms.txt is a proposed plain-text file at the site root that points language models at your key pages. It is a convention, not a standard, and no major consumer has committed to reading it. MCP, the Model Context Protocol, is a published specification for exposing tools and data to a model. A catalogue MCP server is real engineering: authentication, rate limits, and a clear decision about which actions stay read-only.

Where this lands depends on what you run. On Shopify it is the Storefront API and product metafields. On BigCommerce, the catalog API and custom fields. On Adobe Commerce and Magento, product attributes and the GraphQL schema. On WooCommerce, product taxonomy and the REST API. The fields are the same idea in every case.

Product feed: SKU-level price, stock and attributes. Agents treat it as the source of truth.
Structured data: schema.org markup in the page HTML, mainly Product, Offer and AggregateRating.
Offer: the schema.org type carrying price, currency, availability and seller. One product can have several.
Availability: an ItemAvailability value such as InStock, OutOfStock, PreOrder or BackOrder — not a vague stock string.
Variant identity: each size, colour or pack size as its own entity, with its own SKU and GTIN.
Attribute completeness: material, dimensions, compatibility and box contents, the fields that answer comparison questions.
Merchant listing: the product result that carries price and stock in search, governed by Google’s product structured data rules.

Surface	What the agent gets	Where it lives	Effort
Product feed	Price, stock and attributes at SKU level	PIM, ERP or platform export	Medium, and ongoing
Product page structured data	Price, availability, rating, identifiers	Product template	Low to medium, one template change
llms.txt	A curated index of pages worth reading	Site root text file	Low, unproven value
Catalogue MCP server	Live search, stock checks, cart actions	Developer endpoints	High, needs auth and an owner
Checkout protocol	Order and payment on the customer’s behalf	Third-party specification	Not generally available

What do the checkout and payment terms mean, and which are live?

This is the moving part of the vocabulary. When a vendor says agentic checkout is solved, read it as a sales claim rather than a technical status.

ACP and AP2 overlap and are not interchangeable. Neither is universal. Support inside a given commerce platform is partial and shifts between releases, so read that platform’s own notes before you plan around a feature. The breakdown by platform sits in our platform notes.

The unresolved question is liability. If an agent buys the wrong variant, or spends more than the customer intended, someone absorbs the cost. That is a contract and payments problem, not a content problem, and it is not settled.

Agentic checkout: an agent completes a purchase, including payment, from a standing instruction.
Payment mandate: a signed, limited instruction allowing an agent to spend on the customer’s behalf.
Tokenised credential: a payment token the agent holds instead of a card number.
Merchant of record: whoever legally sells the goods. That stays you, whichever agent pressed buy.
SCA and 3-D Secure: authentication steps built on a human being present. An agent cannot approve a push notification in someone’s banking app.
ACP: the Agentic Commerce Protocol, published by OpenAI and Stripe for checkout inside agent surfaces.
AP2: the Agent Payments Protocol, published by Google, built around signed mandates and verifiable intent.

Which terms are real standards, and which are marketing?

Worth separating. schema.org Product and Offer are widely implemented vocabularies with documented properties. The Model Context Protocol has a published specification and reference implementations behind it. llms.txt is a proposal with no governance body. ACP and AP2 are published specifications backed by large companies, but adoption is early and neither is a recognised standard.

Then there is the marketing layer: agent-ready, AI-native commerce, autonomous storefront. None has a definition anyone enforces. Ask what the vendor means and which field, endpoint or protocol it maps to. If the answer is a slide and not a schema, the term is decoration.

How do you measure agent traffic, and what can’t you measure?

Agent traffic is hard to count. The honest position is that most of it cannot be separated from ordinary traffic today.

You cannot yet tell a customer who read about you in a chat assistant and then typed your URL from any other direct visit. A dashboard claiming complete agent attribution is inferring, not measuring. Treat those numbers as directional and check them against server logs, which is the same discipline that keeps analytics and attribution honest for every other channel.

Declared crawler: a bot that names itself in the user-agent string. Server logs show these clearly.
Assistant referral: an HTTP referrer from a chat product. Frequently absent.
Dark traffic: sessions with no referrer and no campaign. You cannot assign these to an agent with confidence.
Server log analysis: the strongest evidence available. Look for repeated fetches of product and stock endpoints.
Assisted conversion: the purchase happens later, in a normal browser. Visible only if the agent passed a readable parameter.

Which terms should you act on first?

Ignore vocabulary with no action attached. These steps cover most of the value, in the order worth doing them.

Most of this is catalogue and template work, which is where agentic delivery platform work and e-commerce SEO meet. The reading is cheap. The data hygiene is the job.

Clean the feed at variant level: price, currency, availability, GTIN and brand for every SKU.
Put Product and Offer structured data on every product template, with availability and price validity.
Load a product page with JavaScript off. Name, price, stock and variant options should still be in the HTML.
Write the comparison attributes: material, dimensions, compatibility, what is in the box.
Name an owner for the checkout question. SCA, fraud rules and liability need an answer before an agent can buy.
Defer the MCP server until a specific agent or partner needs one.

Frequently asked questions

Is agentic commerce the same as conversational commerce?

No. Conversational commerce is a chat interface on top of shopping, where a bot answers questions and hands you a link or a basket. Agentic commerce goes further: the agent holds a goal and can act on it, including comparing across stores and eventually paying. Chat is one interface an agent might use. The defining part is delegated action, not conversation.

Do I need a separate product feed for AI agents?

Usually not. Agents read the same feed you send to shopping channels, plus the structured data on your product pages. The gap is quality, not a second file. Fill price, currency, availability and identifier fields at variant level and keep them current. A dedicated agent feed only makes sense if your channel feed is deliberately trimmed for one marketplace.

Can an AI agent complete a checkout on my store today?

Rarely end to end, and never without setup. Where it works, it runs through a specific protocol such as ACP or AP2 and a payment provider that supports it. Most agents today stop at a link or a pre-filled basket. Strong customer authentication is the usual blocker, because those steps assume a person is present to approve them.

What is MCP in an e-commerce context?

The Model Context Protocol is an open specification for exposing tools and data to a model. In commerce, an MCP server could let an agent search your catalogue, check stock or build a basket through defined calls rather than scraping pages. It is real engineering with authentication and permission decisions, and worth building only when an agent you care about can use it.

How is AEO or GEO different from agentic commerce?

Answer engine optimisation and generative engine optimisation describe making content quotable by AI answers. That is about being cited. Agentic commerce is about being transacted with. The overlap is product data: the same clean price, stock and attribute fields serve both. AEO gets you mentioned; agent readiness gets you into a shortlist an agent can act on.

Sources

Start with the data an agent reads first

The audit reviews your product feed, structured data and product templates against what search engines and agents actually parse. You get a prioritised list of fixes with owners, not a score.

Request the SEO audit

5 Kings Editorial

5 Kings Marketing team