What Is Agentic Commerce? Definition, Examples, and How It Works

"Agentic commerce" has a marketing problem. Everyone uses it; almost no one defines it. The result is that vendors stretch the term to mean "we have a chatbot" or "we have a recommendation engine", and the actual structural change goes unnoticed. This piece is a definition and a stack.

Definition

Three terms in that definition do real work.

• Autonomous: the agent is allowed to make non-trivial choices between candidates without consulting the human at each step.
• Goal: the input is in natural language, not a SKU or a search keyword. "A travel jacket I can wear over a suit" — not "navy travel jacket size M".
• Delegated payment: a wallet, card-on-file, or buy-with-agent flow that lets the agent complete the purchase without re-prompting the user.

Drop any of these and you are describing something else: a recommender (no agency), a search engine (no natural-language goal), or a chatbot (no delegated payment).

Three live examples in 2026

Example 1 — Shopper-side personal agent

A consumer opens ChatGPT and says "buy me a birthday gift for my brother — he is into vintage cameras, budget £200". ChatGPT, with the user's permission and a connected payment method, queries multiple retail APIs, evaluates options against the goal (vintage, photography, £180-220 range), picks one, and completes checkout. The user receives an order confirmation.

The retailer in this flow never sees a session in their analytics. The conversion is a checkout API call from an OpenAI IP range with a "buy-via-agent" tag in the order metadata.

Example 2 — Retailer-side concierge agent

A shopper lands on a retailer's homepage and is greeted by a branded agent. They describe what they are looking for ("something for a beach holiday, breathable, under £150"). The agent searches the retailer's own catalogue, returns three matched products with reasoning, and offers to add the chosen one to cart. Conversion analytics see this as a normal checkout, but the path-to-purchase is one conversation instead of seven pageviews.

This is where Swap, Klevu's RetailMedia, and Algolia's NeuralSearch sit today.

Example 3 — Marketplace-side agent

A buyer on a B2B marketplace says "source 200 units of biodegradable phone cases — RoHS-compliant, ships from EU, lead time under three weeks". The marketplace's agent issues quote requests to qualified suppliers, evaluates responses against the constraints, and returns a ranked vendor list. The buyer picks one and the marketplace handles escrow.

Alibaba's "Ask AI" and Amazon Business's "Source-with-agent" are the early production examples.

The stack underneath

Six layers, top to bottom.

1. Goal layer — the human input in natural language. Increasingly multimodal (voice + image).
2. Reasoning layer — the LLM that plans, decomposes the goal, evaluates candidates. GPT-4-class or better.
3. Tool layer — the function calls and API access the agent has. Catalogue search, payment, shipping APIs.
4. Evidence layer — the data the agent reads when evaluating candidates. Reviews, UGC, structured product data, Q&A. This is where merchants influence outcomes.
5. Wallet layer — delegated payment with spending limits and per-merchant trust scores. ApplePay-with-agent, Visa AI checkout, Stripe Agent Pay.
6. Audit layer — agent logs accessible to the user, so they can see why a particular SKU was chosen.

Protocols that matter

Three emerging protocols are worth tracking.

• MCP (Model Context Protocol) — an Anthropic-led standard for how tools expose themselves to LLMs. Increasingly used for retailer APIs.
• llms.txt — a draft standard for telling AI crawlers which pages are canonical and answer-ready. Different from robots.txt; complementary.
• OASF (Open Agent Settlement Framework) — emerging payment protocol for agent-to-merchant trust. Still early.

What it means for merchants today

Five concrete things to do this quarter.

1. Audit your robots.txt — make sure GPTBot, ClaudeBot, PerplexityBot, Google-Extended are explicitly allowed.
2. Ship an llms.txt at /llms.txt that lists your most citable pages with one-line descriptions.
3. Validate your JSON-LD on every PDP, category page, and resource page. Use the schema.org validator.
4. Tag verified-buyer reviews with a clear isVerifiedBuyer field. Agents weight this heavily.
5. Stand up a citation dashboard — at minimum, monitor referer headers from GPTBot, ClaudeBot, PerplexityBot.

We cover the full playbook on our <a class="text-primary hover:underline" href="/answer-engine">Answer Engine page</a>.

Closing

Agentic commerce is not a 2030 problem. It is a 2026 distribution channel that already moves measurable revenue for the merchants who set it up. The cost of action right now is small (config files, structured data, a few resource pages). The cost of inaction grows monotonically as more shoppers route through agents and your absence becomes structural rather than tactical.