TLDR: This glossary defines over 110 agentic commerce terms across categories including product data, payments, protocols, authority, governance, and measurement. Key protocols covered include ACP, UCP, MCP, and AP2, alongside concepts like delegated authority, prompt injection, and audit logging that underpin safe and accountable agent-led transactions.
Tip: Hit Cmd+F (Ctrl+F on Windows) to quickly search for terms on the page.
Agentic commerce is the use of AI agents to research, compare, recommend, and sometimes buy products or services on behalf of a person or business.
This glossary explains the terms that merchants, ecommerce teams, marketers, product teams, and AI adoption leads need to understand, organised by category.
For the operational detail behind several of these terms, my seven-part agentic commerce series covers execution, infrastructure, preferences, judgement, authority, liability, and governance in depth, alongside a dedicated piece on agentic commerce tax and current statistics and benchmarks.
Agentic commerce terms
Agentic commerce
Agentic commerce is a form of buying and selling where AI agents act on behalf of consumers or business buyers. The agent may search for products, compare sellers, apply stated preferences, check policies, build a basket, request approval, and complete a purchase.
IBM defines agentic commerce as an approach in which intelligent agents research, negotiate, and complete purchases, often without direct human intervention. For a full walkthrough of what agent-readable stores look like in practice, see Agentic Commerce: How to Make Your Store Readable to AI Agents.
AI shopping agent
An AI shopping agent is software that helps a buyer find, compare, and buy products from a machine-readable product catalog. A basic shopping assistant answers questions. A stronger shopping agent makes choices, calls tools, remembers preferences, and acts across several systems without constant prompting.
Agentic checkout
Agentic checkout is a checkout flow built for AI agents rather than human shoppers alone. It lets an approved agent pass basket details, buyer context, payment information, and consent records directly into a merchant’s checkout process, leading to a much faster customer experience compared to traditional ecommerce.
Agentic storefront
An agentic storefront is the version of a merchant that an AI agent can read and use. It includes product data, feed quality, availability, returns information, delivery details, checkout access, and brand information formatted for machine consumption rather than human browsing.
Agent-ready commerce
Agent-ready commerce describes a merchant with the technical, commercial, and policy infrastructure AI agents need to discover products, understand offers, build baskets, and complete or prepare purchases.
Agentic commerce optimisation (ACO)
Agentic commerce optimisation, or ACO, is the work of making a merchant’s products and systems easier for AI shopping agents to find, understand, recommend, and transact with. It includes product feed optimisation, structured data, checkout readiness, returns data, accurate attributes, and agent-facing policies.
See the comparison table for how ACO sits next to SEO, AEO, and GEO within the wider digital marketing sphere.
ACO is important for traditional ecommerce brands looking to benefit from online shopping mediated by agentic payments and transactions.
SEO, AEO, GEO, and ACO overlap, but each optimises for a different output and a different reader.
| Discipline | What it optimises | Reader | Typical output |
| SEO | HTML pages and rankings | Human searcher on Google or Bing | Copy rewrite, link building |
| AEO | Citation inside AI-generated answers | ChatGPT, Perplexity, AI Overviews | Clear, retrievable answer content |
| GEO | Visibility across all generative AI outputs | Any generative AI system | Broader content and source strategy |
| ACO | Product feeds, attributes, and protocol coverage | AI shopping agents (ChatGPT, Rufus, Gemini) | Feed enrichment, schema, protocol compliance |
Most retailers need all four running at once for full agentic commerce readiness. Agentic Commerce Optimisation (ACO) is the narrowest and the newest, and is the one most directly tied to revenue, since it determines whether an agent can recommend and transact a specific product rather than simply mention a brand.
AI commerce
AI commerce is the broader use of artificial intelligence across buying and selling. It includes product recommendations, customer support, search, personalisation, dynamic pricing, content generation, demand forecasting, and agentic shopping as one subset.
Conversational commerce
Conversational commerce means buying through chat, voice, or messaging interfaces. It can involve human agents, scripted chatbots, and an AI assistant facilitating a purchasing decision.
Agentic commerce goes a step further: the software takes independent action rather than only responding to a prompt.
This has implications for product discovery.
Autonomous shopping
Autonomous shopping is buying with little or no step-by-step human involvement. The buyer states a goal, rule, or budget, and the autonomous AI agent handles the search, comparison, and transaction flow on its own.
Zero-click buying
Zero-click buying describes a buying pattern where the shopper never visits a merchant site or manually works through checkout. The personal or business agent finds the item, checks the terms, requests approval when its mandate requires it, and completes the order through a connected payment or checkout system.
Machine customer
A machine customer is a non-human economic actor, an AI agent or automated system, that researches, selects, and purchases on behalf of a person or organisation. Gartner coined the term to describe this emerging buyer class, distinct from the consumer it represents.
AI search and visibility terms
AI visibility
AI visibility measures how often a brand, product, or source appears in AI-generated answers, shopping recommendations, and agent responses. It is the AI-era equivalent of search discoverability in digital marketing.
Answer engine optimisation (AEO)
Answer engine optimisation, or AEO, is the practice of structuring content so it gets cited inside AI-generated answers. AEO focuses on clear answers, definitions, source credibility, and retrievable content fragments.
Generative engine optimisation (GEO)
Generative engine optimisation, or GEO, is the broader practice of improving visibility inside generative AI systems, including AI search, chat assistants, and answer engines.
AI search
AI search is search where the system produces a generated answer, recommendation, comparison, or summary rather than only a list of ranked links. Google AI Overviews, AI Mode, ChatGPT Search, and Perplexity all fall under this umbrella.
AI shopping surface
An AI shopping surface is any place where a user can discover, compare, or buy products through an AI system.
Current examples include ChatGPT Shopping, Google AI Mode, Gemini, Perplexity Shopping, Amazon Rufus, Microsoft Copilot Shopping, retailer-built assistants, and voice agents.
Coverage varies sharply by surface, so a merchant visible on one may stay invisible on another without separate integration work.
Found rate
Found rate is the share of tested prompts where an AI system finds or recommends a brand, product, or seller. It tells a merchant whether intelligent agents can discover their products at all, before recommendation quality enters the picture.
AI share of voice
AI share of voice is the share of AI-generated recommendations, mentions, or citations a brand receives compared with competitors, measured across a defined set of prompts.
Citation rate
Citation rate is the share of AI answers where a brand, page, report, or product appears as a cited source. This metric matters most for publishers, analysts, and B2B sites that compete on authority rather than product listings.
Recommendation rate
Recommendation rate is the share of shopping prompts where an AI system recommends a merchant’s product as one of the buying options it surfaces. Important as digital commerce goes agentic.
Merchant ranking
Merchant ranking is the order in which sellers appear when several merchants offer the same or a similar product. OpenAI has said ChatGPT may rank merchants using signals including availability, price, quality, whether the merchant is the maker or primary seller, and whether Instant Checkout is enabled.
Product data terms
Product feed
A product feed is a structured file or data stream that tells platforms what a merchant sells. It usually includes product titles, descriptions, prices, images, stock status, categories, identifiers, and URLs.
OpenAI’s product feed specification defines field names, data types, constraints, and accepted values for product discovery, pricing, and availability inside ChatGPT shopping experiences.
Product schema
Product schema is structured product information written in a format search engines and AI systems can parse. It typically includes the product name, image, description, brand, SKU, offer, price, availability, shipping, and returns.
Merchant listing structured data
Merchant listing structured data is product markup that makes products eligible for richer listings in Google Search, including price, availability, shipping, and return information. It is important for product discovery.
See Google’s structured data documentation for the full field reference.
SKU
A SKU, or stock keeping unit, is a merchant’s internal product identifier. It supports stock tracking, fulfilment, variant management, and reporting.
GTIN
A GTIN, or global trade item number, is a standard product identifier used to identify products across retailers, marketplaces, and data providers. Common GTIN formats include UPC, EAN, JAN, and ISBN.
MPN
An MPN, or manufacturer part number, is the product identifier the manufacturer assigns. It helps agents and marketplaces match the same underlying product across different sellers.
Canonical product ID
A canonical product ID is the preferred identifier used to treat different listings as the same underlying product. It reduces duplicate recommendations and prevents agents from comparing the wrong items against each other.
Product variant
A product variant is a version of a product that differs by size, colour, material, model, storage, or another attribute. Agents need clean variant data because the wrong variant means the wrong purchase.
Product family
A product family is a group of related products sharing a brand, model line, or use case. This grouping lets agents compare similar products without mixing unrelated items into the same comparison.
Offer
An offer is the commercial version of a product from a particular seller. The underlying product may stay identical across sellers, but the offer can differ by price, availability, delivery speed, warranty, return terms, and seller trust signals.
Attribute completeness
Attribute completeness measures how much of the expected product information a listing actually contains. In agentic commerce, missing attributes can stop a product from being recommended at all, particularly in categories where fit, compatibility, or restrictions decide the purchase.
Attribute confidence
Attribute confidence is the level of certainty that a product attribute is correct. An agent may need to know whether a claim such as “waterproof” or “compatible with iPhone 16” came from the manufacturer, a marketplace, a review, or an inferred source.
Product grounding
Product grounding means an agent’s recommendation rests on actual product data rather than invented or outdated information. It stands as one of the main defenses against hallucinated products, especially for conversational commerce surfaces like chatbots.
Hallucinated product
A hallucinated product is a product an AI system invents or misstates. It may carry the wrong price, the wrong features, the wrong availability, or may not exist at all. See product grounding above.
Availability, delivery, and fulfilment terms
Inventory freshness
Inventory freshness measures how current a merchant’s stock data stays. Agentic commerce depends heavily on fresh stock information, since agents may build baskets or make recommendations without a human checking the live product page.
Price freshness
Price freshness measures how current a product’s price stays across feeds, product pages, checkout, ads, and AI shopping surfaces. Stale prices break buyer trust and cause failed transactions.
Stock status
Stock status tells an agent whether an item sits in stock, sits out of stock, qualifies for pre-order, sits on backorder, or carries limited availability.
Backorder
A backorder is an order for an item not currently in stock but still purchasable for later fulfilment. Agents need this information clearly flagged, since some buyers accept slower delivery while others reject it outright.
Delivery promise
A delivery promise is the stated delivery date or delivery window offered to the buyer. It differs from vague shipping language because an agent needs a machine-readable commitment it can compare across sellers.
Estimated arrival date
An estimated arrival date is the predicted date an order reaches the buyer. It can depend on destination, stock location, carrier, order cut-off time, and fulfilment method.
Shipping policy
A shipping policy explains shipping costs, destinations, delivery times, restrictions, and carrier options. Schema.org’s OfferShippingDetails property represents shipping costs and delivery times in structured form.
Return policy
A return policy explains when and how a customer can return a product. Schema.org’s MerchantReturnPolicy property covers return conditions, methods, fees, and refund options in a format agents can parse directly.
Cancellation policy
A cancellation policy tells the buyer and the agent whether an order can be cancelled after it has been placed, and under what conditions.
Substitution rule
A substitution rule tells an agent what to do when the ideal product sits unavailable. The agent may have permission to buy another brand, choose a different size, wait for restock, or ask the user. My piece on preference layers covers how substitution settings work alongside budget caps and brand lists.
Post-purchase agent
A post-purchase agent handles work after checkout. It may track delivery, process returns, chase refunds, claim warranties, compare final invoices, or update expense systems on the buyer’s behalf.
Payment and checkout terms
Agentic payment
An agentic payment is a payment an AI agent starts or completes on behalf of a user or business. It depends on prior consent, spending rules, identity checks, and payment controls agreed in advance. Making an agent pay for a product might also be called an agentic transaction.
Agentic Commerce Protocol (ACP)
The Agentic Commerce Protocol, or ACP, is an open standard from Stripe and OpenAI, later joined by Meta, that defines how AI agents complete purchases with merchants on a buyer’s behalf.
Stripe’s documentation describes ACP as a structured commerce flow purpose-built for agent-led checkout. ACP currently powers ChatGPT Shopping’s Instant Checkout.
See the protocol comparison table below, and my breakdown in Agentic Commerce Pt. 2: Infrastructure, for how it differs from UCP and AP2.
Universal Commerce Protocol (UCP)
The Universal Commerce Protocol, or UCP, is an open standard from Google and Shopify that handles the full commerce journey for AI agents, from discovery through checkout to post-purchase service.
Where ACP focuses narrowly on the transaction step, UCP coordinates discovery, basket management, checkout, order tracking, and returns within a single standard.
My infrastructure piece covers how UCP and ACP can run side by side for merchants who need both.
Agent Payments Protocol (AP2)
The Agent Payments Protocol, or AP2, is an open protocol from Google for secure, reliable, agent-initiated payments. AP2 connects agents, merchants, users, and payment systems through verifiable payment flows, and it introduces the Intent Mandate, a signed record of price limits, timing, and conditions an agent must honour before it spends.
See Agentic Commerce Pt. 3: Preferences for how the Intent Mandate enforces budget caps in practice.
Model Context Protocol (MCP)
The Model Context Protocol, or MCP, is an open standard, originally created by Anthropic, that gives AI agents a universal way to connect to external tools, files, databases, and data sources in real time.
Anthropic introduced MCP in November 2024, and the standard now sits under the Linux Foundation. MCP forms the data layer that ACP and UCP can both build on. See my piece on MCP.
Instant checkout
Instant checkout is a checkout flow that lets a buyer complete a purchase inside the AI shopping experience rather than moving through the merchant’s standard website checkout. OpenAI’s Instant Checkout in ChatGPT runs on ACP and Stripe.
Shared payment token
A shared payment token is a Stripe mechanism, used within ACP, that lets an AI agent pass a buyer’s tokenised payment method to a merchant for processing without exposing the underlying card number to every party in the flow.
Tokenised payment credential
A tokenised payment credential is a substitute for a card number or payment account. It lets payment systems process a transaction without exposing the buyer’s underlying payment details to every party that touches the flow.
Strong customer authentication
Strong customer authentication is a payment security process that asks the buyer to prove identity using more than one factor: something they know, something they hold, or something they are.
Payment authorisation
Payment authorisation is the approval step where the payment network, card issuer, or payment provider confirms a transaction can proceed.
Settlement
Settlement is the process of moving money from the buyer’s payment method to the merchant after a payment has been authorised and captured.
Chargeback
A chargeback is a payment reversal a buyer starts through their card issuer or payment provider.
In agentic commerce, chargebacks raise harder questions, since the merchant, the agent, the user, the payment provider, and the platform may all share some part of the dispute.
My piece on agentic commerce liability covers how chargeback rules, Section 75, and emerging AI insurance products handle disputed agent purchases.
Merchant of record
The merchant of record is the legal seller responsible for processing the payment, handling tax obligations, and managing certain customer service duties.
This role becomes harder to pin down when a purchase happens inside a third-party AI interface rather than the merchant’s own site.
See Agentic Commerce Pt. 6: Liability for how responsibility splits across the parties in an agent-led purchase.
Proof of purchase
Proof of purchase is evidence that a transaction happened. In agentic commerce, proof needs to include the product, the merchant, the price, the approval, the mandate, the timestamp, and the agent’s action log, not just a receipt.
Authority, consent, and control terms
Delegated authority
Delegated authority is the permission a user grants an agent to act on their behalf. It can cover research only, basket creation, purchase preparation, or full payment execution, depending on the mandate the user sets.
Agent mandate
An agent mandate is the instruction, permission, and boundary a user gives an agent. It typically includes the task, the product type, the budget, seller limits, a deadline, the payment method, and any approval requirement.
Agentic Commerce Pt. 5: Authority covers how mandates, scoped tokens, and signed credentials work together to prove what an agent is allowed to do.
Human-in-the-loop approval
Human-in-the-loop approval means the agent must ask a person before completing a sensitive action. A buyer may let the agent compare products freely but require approval before any payment goes through.
Spend limit
A spend limit is the maximum amount an agent can spend without further approval. It can apply per transaction, per day, per category, per merchant, or per user.
Approval policy
An approval policy defines when an agent needs permission before it continues. It may require approval above a set amount, for a new merchant, for a subscription, or for a product outside the user’s usual rules.
Mandate revocation
Mandate revocation is the process of removing or ending an agent’s permission to act. Speed matters here, since an agent could otherwise keep using a permission the user already withdrew.
Consent receipt
A consent receipt is a record showing what the user approved, when they approved it, and what the agent had permission to do. It supports audits, dispute handling, and buyer trust.
Escalation rule
An escalation rule tells the agent when to stop and ask a human. Triggers can include unclear product claims, missing return terms, unusual prices, restricted goods, conflicting preferences, or payment risk.
Agent identity
Agent identity is the verified identity of the software agent taking action. Merchants and payment providers need to confirm whether a request came from a real user’s agent, a trusted platform, an approved business system, or an unknown bot.
Agentic Commerce Pt. 5: Authority covers the identity, authentication, and delegation questions a seller has to answer before money moves.
Agent trust layer
The agent trust layer is the verification, consent, and authorisation infrastructure that confirms an AI agent acts legitimately on behalf of a genuine buyer.
It sits underneath delegated authority, agent identity, and mandate enforcement as the combined system that makes all three checkable in real time.
Agent permissions
Agent permissions define what an agent can read, write, change, buy, cancel, refund, or approve. Strong permission design sits at the centre of safe agentic commerce.
Preference and judgement terms
Preference profile
A preference profile is the structured record of what a buyer wants. It can include brands, sizes, colours, dietary needs, delivery rules, payment preferences, budget, sustainability preferences, and loyalty scheme details.
Preference memory
Preference memory is an agent’s ability to remember user preferences across sessions. This improves shopping accuracy but raises privacy, fairness, and control questions that merchants and platforms still need to resolve.
Budget rule
A budget rule tells the agent how much the buyer accepts spending. It can set a hard ceiling, a preferred range, or a trade-off rule such as paying more only when delivery arrives faster.
Brand preference
A brand preference tells the agent which brands the buyer favours, avoids, or trusts. In agentic commerce, brand preference carries new weight, since the agent may filter options before the buyer ever sees them.
Compatibility rule
A compatibility rule tells the agent whether a product works with something the buyer already owns. This matters for electronics, car parts, software, furniture, appliances, and B2B procurement alike.
Constraint
A constraint is a hard rule the agent must obey. Examples include a gluten-free requirement, a delivery deadline, a laptop size limit, or a requirement that the seller appear on an approved supplier list.
Trade-off
A trade-off is a choice between competing preferences. The cheapest item may not offer the fastest delivery, the best-rated item may sit out of stock, and the preferred brand may cost more than an alternative.
Digital purchasing twin
A digital purchasing twin is a structured model of how a person or business buys. It goes beyond a simple preference list. It captures rules, habits, risk tolerance, budget logic, approval boundaries, and judgement patterns.
Gartner describes the seller-side version as a digital twin of the customer; Agentic Commerce Pt. 4: Judgement covers how this model decides what an agent does when explicit rules run out.
Infrastructure and protocol terms
Comparing the four commerce protocols
| Protocol | Backed by | Layer it handles | Primary use today |
| MCP | Anthropic, now Linux Foundation | Data connectivity | Lets agents access tools and live data from any connected system |
| ACP | Stripe and OpenAI, joined by Meta | Checkout and payment | Powers ChatGPT Shopping’s Instant Checkout |
| UCP | Google and Shopify | Full commerce journey | Discovery through checkout to post-purchase service |
| AP2 | Payment authorisation | Signed Intent Mandates for agent-initiated payments |
Most retailers will need more than one of these over time. MCP supplies the data layer underneath both ACP and UCP, so it tends to come first.
Tool use
Tool use is an AI agent’s ability to call external functions, APIs, databases, browsers, or payment systems. In commerce, tool use might include checking stock, adding an item to a basket, retrieving a quote, or creating a return.
Tool permissioning
Tool permissioning defines which tools an agent can use and under what conditions. A shopping agent might have permission to search the web but stay blocked from placing an order without separate approval.
API
An API, or application programming interface, gives software systems a structured way to exchange data and perform actions. Agentic commerce relies on APIs for product data, stock, orders, payments, returns, customer accounts, and support workflows.
Webhook
A webhook is an automated message one system sends to another when an event happens. A retailer might use webhooks to tell an agent that an order has shipped, a refund has gone through, or a product has come back in stock.
Order status API
An order status API lets an agent check where an order sits in the fulfilment process. This supports delivery tracking, refund handling, and customer service, and supports the customer experience.
Policy engine
A policy engine is software that evaluates rules before an action happens. In agentic commerce, it decides whether an agent may buy, escalate, substitute, refund, or reject a seller.
Audit log
An audit log records what happened, when it happened, who or what caused it, and what data informed the decision. Agentic commerce needs audit logs because agents take commercial actions that someone may need to explain later.
Action log
An action log records the steps an agent took. It can include searches, product comparisons, tool calls, rejected options, selected items, approvals, and checkout events.
Dry run
A dry run is a test run where the agent completes the research and planning steps without finishing the transaction.
Teams use dry runs to test agent behaviour before allowing real purchases (e.g., having an agent pay for a mock product within a sandbox to text out a new payment provider or headless commerce feature).
Simulation mode
Simulation mode lets teams test agentic shopping flows using fake users, fake payments, test products, and controlled scenarios.
Risk and governance terms
Guardrail
A guardrail is a rule, check, or control that limits what an agent can do. Guardrails can stop unsafe purchases, block unknown merchants, force approvals, or detect manipulative instructions before they reach checkout, and are crucial to the safety and security of any agentic system.
Prompt injection
Prompt injection is an attack where hidden or visible text tries to manipulate an AI system into ignoring its rules or taking an unsafe action.
This poses a significant risk for digital commerce, since agents read web pages, product descriptions, reviews, and emails before making a decision, and any of those sources can carry a hidden instruction.
Tool poisoning
Tool poisoning happens when an external tool, connector, or data source feeds an agent misleading, malicious, or manipulated information.
Malicious merchant
A malicious merchant is a seller that tries to exploit the agentic commerce flow. This can include fake products, fake reviews, misleading policies, counterfeit goods, hidden fees, or prompt injection embedded inside product pages. These ill-advised, short-term hacks damage the customer relationship.
Brand impersonation
Brand impersonation happens when a fake seller poses as a real brand. AI shopping agents need stronger merchant verification than human shoppers, because a convincing fake storefront can fool an autonomous AI agent just as easily as it fools a person.
Data poisoning
Data poisoning is the manipulation of data an AI assistant or system relies on. In commerce, it can distort product rankings, recommendations, reviews, pricing, or seller trust scores.
Red teaming
Red teaming is structured testing where people or systems try to break, mislead, or exploit an AI system. Commerce red teaming should test payment abuse, fake merchants, unsafe substitutions, prompt injection, refund abuse, and authority failures.
Evaluation set
An evaluation set is a collection of test prompts, products, edge cases, and expected outcomes used to measure how well an agent performs.
Retrieval failure
A retrieval failure happens when an agent can’t find the right product, policy, seller, or source, even though the information exists somewhere in the system.
Unsupported claim
An unsupported claim is a statement the agent makes without enough evidence from product data, merchant policies, reviews, or verified sources. Can be solved by grounding (RAG).
Agentic commerce governance
Agentic commerce governance is the layer that decides whose rules bind an agent before it spends a buyer’s money: the user’s stated instructions, the platform’s defaults, the ranking function that decides which products an agent sees first, or commercial pressure from merchants and developers.
Agentic Commerce Pt. 7: Governance covers why this question remains unsettled across every major platform.
Agentic AI autonomy spectrum
The agentic AI autonomy spectrum classifies AI systems by how independently they act, from low-autonomy automation that follows fixed rules to fully autonomous agents that plan, reason, and act with minimal human input.
Where a commerce agent sits on this spectrum determines what controls and disclosures it needs. See The Agentic AI Autonomy Spectrum for the five-level framework regulators increasingly reference.
Measurement terms
Agent conversion rate
Agent conversion rate is the percentage of agent-led shopping sessions that end in a purchase, an approved basket, a checkout handoff, or another defined conversion event.
Agent-assisted revenue
Agent-assisted revenue is revenue an AI agent influenced. It can include sales where the agent researched the category, recommended the product, built the basket, or completed checkout directly.
Checkout handoff rate
Checkout handoff rate is the share of agent shopping sessions where the agent sends the buyer to checkout or passes checkout data into a connected system.
Approval completion rate
Approval completion rate is the share of approval requests a human accepts. A low approval completion rate can mean the agent keeps choosing poor options, asking too often, or failing to explain its reasoning clearly.
Feed error rate
Feed error rate is the share of products carrying missing, invalid, inconsistent, or rejected product feed data.
Catalog coverage
Catalog coverage is the share of a merchant’s full product range available in structured, agent-readable form.
Attribute coverage
Attribute coverage is the share of required or useful product attributes present across the catalog. High attribute coverage helps agents compare products accurately rather than guessing at missing fields.
AI performance insights
AI performance insights are reports showing how products or brands perform on AI shopping surfaces. Google’s Merchant Center AI performance insights are designed to show how products get discovered across AI Mode, AI Overviews, and Gemini.
B2B and procurement terms
Read more about this topic in my guide to agentic procurement.
Procurement agent
A procurement agent is an AI agent that helps a business buyer research suppliers, compare offers, check policy, create purchase requests, or prepare purchase orders.
Approved supplier list
An approved supplier list is the set of vendors a company allows employees or agents to buy from. Agents need this list to avoid purchases from unapproved sellers.
Purchase order
A purchase order is a formal buying document a buyer sends to a supplier. It usually includes products, quantities, prices, delivery terms, and approval details.
Contracted pricing
Contracted pricing is pricing a buyer and supplier agree in advance. A procurement agent should check contracted pricing before defaulting to public list prices.
Punchout catalog
A punchout catalog is a supplier catalog a buyer accesses from inside their own procurement system. It lets employees or agents shop from approved suppliers while keeping procurement controls in place.
Approval workflow
An approval workflow is the process a business uses to approve a purchase before it completes. It may involve a budget holder, a manager, a finance team, a procurement team, or a legal team.
Invoice matching
Invoice matching is the process of checking whether an invoice matches the purchase order, the goods receipt, and the agreed terms.
Three-way match
A three-way match checks the purchase order, the goods receipt, and the supplier invoice before payment goes out. This step matters for B2B agentic commerce, since an agent may help create the purchase, but finance still needs independent evidence before it pays.
Agentic commerce tax attribution
Agentic commerce tax attribution is the question of who accounts for tax when an AI agent completes a purchase on someone’s behalf.
Under current UK and US law, tax consequences still follow the person or company behind the agent rather than the software itself, but cross-border transactions, stacked payment instruments, and incomplete transaction logs complicate the record a buyer needs at tax time.
See my article on tax in agentic commerce for the full breakdown.
The main idea behind agentic commerce
Agentic commerce asks more than whether an AI tool can recommend a product. It asks whether the whole buying system can be understood and used by software.
- Product data has to be structured.
- Prices and inventory in a product catalog have to stay fresh.
- Shipping and returns have to be readable.
- Payments need consent and clear authority.
- Agents need limits.
- Merchants need audit logs.
- Buyers need control.
- Platforms need to know which seller, product, price, policy, and mandate applies to any given transaction.
Merchants that prepare for this shift early will not only improve their AI answer visibility, but also make their products easier for autonomous agents to evaluate, recommend, and buy.
Get started on your website’s agentic commerce readiness checklist today.