Agentic shopping is when the agent completes the purchase

The category has a start date. Those two specifications were Google's AP2 on 16 September 2025 and OpenAI and Stripe's Agentic Commerce Protocol on the 29th — and ChatGPT, at the 800 million weekly users reported at OpenAI's DevDay on 6 October 2025, became the first mass-market surface where a model places the order instead of handing over a link. Everything in this entry dates from that fortnight or the eleven months before it, a window that opens with Anthropic's release of MCP on 25 November 2024.

Where the term came from

The phrase got real on 29 September 2025, when OpenAI shipped Instant Checkout in ChatGPT: US users could buy a single item from a participating Etsy seller without leaving the chat, with Shopify merchants — Glossier, SKIMS, Spanx and Vuori among the names given — announced as the next wave. The plumbing underneath was published the same day as the Agentic Commerce Protocol, an open spec OpenAI built with Stripe and documented publicly at agenticcommerce.dev.

Google had moved two weeks earlier. On 16 September 2025 it announced the Agent Payments Protocol — AP2 — with more than 60 partners including Mastercard, American Express, PayPal and Coinbase, built around signed verifiable credentials it calls an Intent Mandate and a Cart Mandate, and covering cards, real-time bank transfers and stablecoins rather than cards alone.

Before either, Perplexity put "Buy with Pro" in front of US Pro subscribers on 18 November 2024: one-click checkout inside an answer. That was the first mainstream version of the thing, and nobody called it agentic shopping yet.

The rails came first. Mastercard announced Agent Pay on 29 April 2025 and Visa announced Intelligent Commerce the following day, 30 April 2025, both issuing tokenised credentials to a named agent instead of letting it type a card number into a form. AP2 runs on the same principle and is written as an extension of Google's A2A protocol and Anthropic's MCP rather than as a new stack. By the time a chat window could place an order, the payment layer had been under construction for five months.

The dated sequence, start to finish:

A year old as a category. Three months old as a standard. Treat anything that describes it as settled with suspicion.

What the two specs actually require

ACP splits into published pieces that map onto a merchant's stack: a product feed so an agent can read the catalogue, an agentic checkout API so it can open a session and receive authoritative totals, tax and shipping, and a delegated payment spec so a scoped credential can be handed to the merchant. The merchant stays merchant of record throughout.

AP2 attacks the same problem from the credential side, with signed artefacts rather than an API surface: an Intent Mandate recording what the buyer authorised the agent to go looking for, a Cart Mandate recording the exact items, totals and merchant the buyer approved, and a payment mandate signalling to the network that the transaction was agent-initiated. AP2 also separates human-present flows, where the buyer approves the cart in real time, from human-not-present flows where the agent acts later against an earlier mandate. The second case is where delegated authority actually bites.

The two are not interchangeable: ACP standardises the merchant interface, AP2 standardises the proof of authorisation. A merchant selling through both surfaces can need both — one integration against the checkout API published on 29 September 2025, one against the mandate format published on the 16th.

It is not conversational commerce

The neighbouring term people collapse it into is conversational commerce — a chatbot or AI shopping assistant that answers questions, narrows a catalogue, and then hands you a link. Every retailer has shipped one of those. They are recommenders with a chat interface.

The boundary is not how good the conversation is. It is whether the agent has authority to act.

Under AP2 the distinguishing artefact is a signed Cart Mandate. A transcript is not one. A system that recommends a product and opens a checkout page in a new tab is not agentic shopping — it is a link.

Two different things get called this

Vendors trade on the gap between the two senses, so label them.

Sense 1 — buyer-side agentic shopping. An agent acting for a consumer: it searches, compares, picks, and pays. ChatGPT Instant Checkout from 29 September 2025, Perplexity's Buy with Pro from 18 November 2024. The live question is how much authority you hand over, and for how long.

Sense 2 — merchant-side agentic commerce enablement. Making your store readable and transactable by someone else's agent: a structured product feed, checkout endpoints an agent can call, order status webhooks. OpenAI's commerce docs spell out that merchant-side surface across the three ACP components. It is an engineering project, not a shopping experience, and it is what vendors are selling when they pitch "agentic shopping readiness."

Sense 1 is the headline. Sense 2 is the invoice. When someone says they are "doing agentic shopping," ask which.

There is also a lazier usage — AI-powered product discovery rebranded. No delegated authority, no transaction. That one does not qualify under either sense.

A worked example: one Etsy listing through ACP

One US buyer, one Etsy listing, one item. That was the entire scope on launch day, 29 September 2025. The path, per the ACP spec:

1. The buyer says what they want. The model surfaces a product from the merchant's feed.
2. The agent creates a checkout session against the merchant's ACP endpoint and gets back real totals, tax and shipping — not a guess from the page.
3. The buyer confirms. This is the authority boundary, marked above, and the point AP2 calls human-present.
4. A scoped, single-use payment credential is passed to the merchant under the protocol's delegated payment flow.
5. The merchant charges it, stays merchant of record, and owns fulfilment, returns and support.

The fee on a completed Instant Checkout purchase is paid by the merchant, and OpenAI states product ranking is not for sale. Read both in the primary source rather than in a summary. They are the whole commercial argument.

How to tell whether something qualifies

Four questions. All four have to land.

1. 1
   Does the agent hold spend authority, scoped to an amount and a time window — or does it stop and ask a human for every charge? Ask-every-time is fine; it just is not agentic shopping.
2. 2
   Does it complete a transaction in a system it does not own, through a documented interface — an ACP checkout session, an MCP tool, a merchant API? Screen-scraping a checkout form counts technically and breaks constantly.
3. 3
   Is there a machine-verifiable record of what the buyer authorised? AP2's Intent and Cart Mandates are signed credentials precisely so a disputed charge can be traced to a specific delegation.
4. 4
   Can you name who eats a wrong order? Under ACP that is the merchant of record. If nobody can answer, the thing is not in production.

What is still unresolved

Three questions are open, and each one is a production risk rather than a debating point.

Who grants access. Not every merchant wants another company's agent inside its checkout. Amazon sued Perplexity in the Northern District of California in November 2025 over the Comet browser shopping on Amazon using customers' own logins, arguing that an agent signing in with a buyer's credentials is not the same thing as the buyer. The outcome sets whether buyer-side agents need merchant permission or only buyer permission. Nobody is neutral on this one. Note the structural difference: ACP and AP2 both assume the merchant has opted in by publishing a feed or accepting a mandate, while a credential-replaying browser agent assumes nothing of the kind.

Who absorbs a bad order. ACP settles the money question: the merchant is merchant of record, so refunds and chargebacks run the normal card path. It does not settle the judgement question. A Cart Mandate records what the buyer authorised, not whether the agent picked the right size — and a signed mandate makes that authorisation harder for the buyer to disown.

Who pays for placement. Instant Checkout bills the merchant, not the buyer, and OpenAI's published position is that ranking is not for sale. That pairing is the business model. It is the line to re-read in the primary source each time either company revises its commerce terms.

What the record supports so far

Across the published specifications and the launch scope, the buyer-side experience is further along than the accountability layer under it. Instant Checkout opened on 29 September 2025 with single-item purchases from one marketplace in one country; the question of who absorbs a wrongly chosen order was, within eight weeks, in front of a court. Neither September 2025 specification yet documents a dispute path for standing spend authority held across multiple merchants — AP2's human-not-present flow defines how such a delegation is signed, not who is answerable when the judgement inside it is wrong.

The merchant-side work is different in kind: a product feed and a callable checkout endpoint are useful whether or not agents ever buy at volume, because both also serve ordinary search, ads and comparison surfaces. That asymmetry is the practical reason the two senses above get sequenced differently. It is the same argument as building a three-agent system that stops when it should rather than one that is clever.

If you are deciding what to wire up first, the gating question is the one that applies to what to automate and what to gate in account-based marketing: automate the retrieval, gate the authority.