The agentic commerce glossary

Core concepts

Agentic commerce arrived with a vocabulary of its own, and most of it hardened between late 2025 and mid 2026: protocols with four-letter names, crawlers with hyphenated ones, and metrics no analytics dashboard shipped with. This page defines all 40 terms in plain language, grouped by where they show up in the work: core concepts, discovery, protocols and payments, and measurement. Each definition stands alone, so link to any term directly.

The seven terms everything else builds on. For the full picture of how they fit together, start with what is agentic commerce.

Agentic commerce

Online retail where an AI agent helps a person shop. The shopper states a goal, the agent researches across merchants, compares options, answers questions, and narrows the choices; on some platforms it can also complete the checkout the shopper approves. The shopper keeps the decision, the agent does the legwork. Full definition and history in what is agentic commerce.

AI shopping agent

Software that shops on a person's behalf: ChatGPT, Amazon Rufus, Perplexity, Google Gemini, Microsoft Copilot. It parses the shopper's intent, fans out into background searches, assembles candidates from feeds and indexes, evaluates them against structured data and reviews, and answers with a short recommended list. How each one works, step by step, in AI shopping agents.

AI assistant vs AI agent

An assistant answers questions and recommends; an agent takes actions: comparing across merchants, filling a cart, completing a checkout the shopper approves. The line blurred in late 2025 when in-chat checkout launched, but most of what agents do for shoppers today is still research and recommendation.

Answer engine

An AI system that responds to a query with a synthesized answer instead of a list of links: ChatGPT, Perplexity, Google's AI Mode, Copilot. Where a search engine returns ten blue links to evaluate, an answer engine returns one answer with a handful of citations, which is why being cited matters more than ranking.

Agentic storefront (AI storefront)

A machine-readable, answer-capable version of a store that AI agents can read, cite, and transact with: structured data, product feeds, plain-text answers to buying questions, and agent-checkout hooks. Building one takes five steps, covered in AI storefront setup.

Agent site

A parallel, agentic copy of a brand's store hosted on the brand's own subdomain, built so AI-referred shoppers land on a page that can answer product, price, and fit questions instead of a static one. The main site stays untouched; checkout returns to the brand. This is what Aigency builds.

Conversational commerce

The older, adjacent idea: a merchant offers a chat interface (a site chatbot, WhatsApp, Messenger) to its own visitors. The direction of loyalty is the difference. A conversational-commerce bot works for the merchant; a shopping agent works for the shopper and reads every store at once.

Discovery and citation

The vocabulary of getting found, trusted, and quoted by AI engines. This layer is live at scale today and carries the volume; the practical playbook for it is in ChatGPT product discovery.

Answer engine optimization (AEO)

The practice of structuring content so answer engines can quote it accurately: answer-first pages, clean headings, plain-text facts, FAQ markup. AEO targets the citation, where classic SEO targets the ranking.

Generative engine optimization (GEO)

A near-synonym for AEO with a broader scope: influencing how generative AI systems describe and recommend a brand overall, including what the underlying models learned in training, not just what they retrieve at answer time. In practice the two labels are used interchangeably; both end at the click.

Citation

A source reference inside an AI answer, usually a link. When ChatGPT recommends a product and links the merchant, that link is a citation. Citations are the atomic unit of AI visibility: earn them and the traffic follows.

Citation share

The share of AI answers in your category that cite or recommend you, measured across engines and across a repeatable set of buying prompts. The agentic-era counterpart of share of voice.

Retrieval-augmented generation (RAG)

The architecture behind most AI answers: the system first retrieves relevant documents (your product pages, reviews, feeds), then generates the answer from what it retrieved. RAG is why fresh, crawlable content influences answers even when the underlying model finished training long ago.

Grounding

Tying a model's answer to retrieved evidence rather than to what it merely remembers from training. A grounded shopping answer quotes your live price and stock; an ungrounded one guesses. Merchants influence grounding by making the evidence easy to fetch and parse.

Hallucination

A confident AI statement that is false: a product you never sold, a price you never charged, a return policy you do not honor. In commerce the cheapest defense is publishing unambiguous, machine-readable facts so the engine never has to fill a gap.

Knowledge cutoff

The date a model's training data ends. Anything after it (a price change, a new product line) reaches the answer only through retrieval, which is why crawl access and feed freshness matter more than the model's age.

AI crawler

A bot an AI company sends to read the web. The ones a merchant meets most, each controllable in robots.txt:

• OAI-SearchBot: indexes pages for ChatGPT search answers. Per OpenAI's bot documentation, sites that opt out are not shown in ChatGPT search.
• GPTBot: collects training data for OpenAI models. Blocking it does not affect ChatGPT search visibility.
• ChatGPT-User: fetches a page live when a user or agent asks about it.
• PerplexityBot: crawls for Perplexity's answer index.
• Google-Extended: the robots.txt token controlling whether Google may use your content for its AI models.

robots.txt

The plain-text file at a site's root that tells crawlers what they may read. In agentic commerce it is the front door: a robots.txt that blocks OAI-SearchBot removes the store from ChatGPT search answers, however good the pages behind it are.

llms.txt

A proposed standard: a Markdown file at a site's root that gives AI systems a curated, token-efficient map of what the site is and where its important pages live. The AI-era counterpart of a sitemap. This site's own copy is at aigency.ai/llms.txt.

Structured data (Schema.org)

Machine-readable labels embedded in a page that state facts outright: this is a Product, this is its price, this is its availability. Schema.org is the shared vocabulary. Structured data is how an agent parses a product page without guessing.

JSON-LD

The format structured data usually ships in: a small script block of JSON in the page's HTML, separate from the visible markup. The rule that keeps it honest: structured data must match what a visitor can see on the page.

Product feed

A structured file listing your catalog (identifiers, titles, prices, availability, images) submitted directly to a platform rather than crawled. OpenAI's ACP feed spec accepts Google Shopping-compatible files refreshed daily; Google Merchant Center feeds power the Shopping Graph.

Shopping Graph

Google's product database behind Gemini and AI Mode shopping: more than 50 billion listings, over 2 billion refreshed every hour, per Google. If your products are not in it, Google's agents cannot recommend them.

Third-party corroboration

Independent coverage an engine reads alongside your own claims: reviews, editorial roundups, community threads. Products that appear in both the merchant feed and independent coverage do better than either alone, because agents weigh evidence they did not get from you.

Protocols, payments, and checkout

The rails that let an agent transact, and the names on them. The checkout layer is the part of agentic commerce still being renegotiated, so these definitions carry dates.

Agentic checkout

A purchase completed by an agent on the shopper's approval, inside the conversation or through a protocol handoff, instead of through the merchant's human-facing funnel. Live today at Perplexity and Microsoft Copilot; OpenAI scaled its version back in March 2026 in favor of discovery.

Agentic Commerce Protocol (ACP)

The open standard from OpenAI and Stripe, launched and open-sourced September 29, 2025, defining how agents read merchant catalogs and complete checkout. The merchant stays seller of record: it accepts or declines each order, processes payment with its existing provider, and owns fulfillment and support.

Instant Checkout

ChatGPT's in-chat purchase feature built on ACP, launched September 2025 with US Etsy sellers. In March 2026 OpenAI scaled it back and pivoted to product discovery plus merchant-built ChatGPT apps, so ChatGPT-referred purchases now overwhelmingly complete on the merchant's own storefront.

Agent Payments Protocol (AP2)

Google's payments protocol, announced September 16, 2025 with more than 60 partners including American Express, Mastercard, PayPal, and Salesforce. Its job is trust: proving cryptographically that a human authorized what an agent bought.

Intent and Cart mandates

AP2's mechanism: tamper-proof, cryptographically signed records of what the shopper asked for (the Intent mandate) and what the agent assembled to buy (the Cart mandate). Together they answer the dispute question agent payments create: did the human really approve this?

Universal Commerce Protocol (UCP)

The retailer-led umbrella standard Google announced at NRF in January 2026 with Shopify, Etsy, Wayfair, Target, and Walmart, spanning discovery, buying, and post-purchase. It powers the buy buttons rolling out in AI Mode and Gemini.

Tokenized agent payments

The card networks' approach: Visa's Intelligent Commerce and Mastercard's Agent Pay, announced within a day of each other in April 2025, replace the raw card credential with a token scoped to an agent, so a merchant can verify an agent is authorized to pay without ever holding the card.

Model Context Protocol (MCP)

Anthropic's open standard, released November 2024, for connecting AI systems to external tools and data sources through one common interface. In commerce it is plumbing: the generic way an agent calls a catalog API, an inventory system, or a checkout.

WebMCP

The Web Model Context Protocol: a browser-level standard from Google and Microsoft, incubated in the W3C Web Machine Learning group and shipped in early preview in Chrome 146 Canary in February 2026. It lets a website expose structured tools (fill this form, add this to cart) to an agent working inside a visible browser session, with the human still present. Our read on what it signals.

Seller of record

The party legally selling the goods: taking payment, owing the refund, owning the customer relationship. Under ACP and UCP the merchant keeps this role even when the transaction starts inside an AI conversation. The protocols move the buy button, not the business.

Traffic and measurement

The metrics layer. Classic analytics were built for humans who browse; these terms exist because agents broke the old dashboard. The side-by-side of what changes is in agentic commerce vs traditional ecommerce.

AI-referred traffic

Visits arriving from an AI engine's answer: a shopper clicked out of ChatGPT, Perplexity, or Gemini to your site. The channel Adobe has tracked since October 2024, and the fastest-growing referral source in US retail. These visitors arrive pre-qualified, because the research happened before the click.

utm_source=chatgpt.com

The tracking parameter OpenAI appends to outbound links from ChatGPT answers. The simplest reliable way to see ChatGPT referrals in analytics: filter sessions on it, then watch the conversion rate next to your site average.

AI-referral share

AI-referred sessions as a share of total sessions. The number that tells you whether agentic commerce is a future concern or a current one for your store; it is best read per category, since AI adoption varies sharply by what people are buying.

Dark traffic

Visits whose origin analytics cannot see, lumped into direct. AI answers generate it two ways: clicks that arrive without a referrer, and influence that surfaces later as a branded search or a typed-in URL. Expect undercounting; the visible AI-referral number is a floor, not a total.

Revenue per AI-referred visit

Revenue attributed to AI-referred sessions divided by those sessions. The quality metric for the channel: volume tells you agents send people, revenue per visit tells you whether the landing experience converts them.

Prompt panel

A standing set of 10 to 20 real buying-intent prompts run weekly across ChatGPT, Gemini, and Perplexity, logging who gets mentioned, cited, and linked. The rank tracker of the agentic era, and the raw input for citation share.

Last-click attribution

The default analytics model that credits a sale to the final click before purchase. Agents break it: the comparison and the decision happen off-site in the conversation, so last-click hands the credit to whatever channel caught the already-decided shopper.

Earned default

What loyalty becomes when an agent re-runs the comparison on every purchase instead of rebuying from habit. A brand is the earned default when clean data, reliable stock, and honest pricing keep winning the re-run. It is re-earned every time, which is the point.