Your Next Customer May Never See Your Website
Assistants can now buy from your store without a site visit. A practical readiness check for ecommerce teams: feeds, ACP checkout APIs, payment rails, and what breaks first.
On this page
- What actually shipped, minus the hype
- Surface one, the feed that became a storefront
- Surface two, the checkout API you probably don't have
- Surface three, payment rails without a card form
- Surface four, order events after the buy button
- What breaks first, ranked from the field
- The part nobody prices in
- Sources
Somewhere right now a shopper is telling ChatGPT to reorder a moisturizer, and the transaction will settle without a single pageview hitting the brand's analytics. That is the whole story. On September 29, 2025, OpenAI switched on Instant Checkout for U.S. Etsy sellers, with more than a million Shopify merchants queued behind them, and open-sourced the Agentic Commerce Protocol (ACP) it built with Stripe to make it work. Thirteen days earlier, Google Cloud had announced the Agent Payments Protocol (AP2) with more than 60 payments and technology partners, including Mastercard, PayPal, and American Express.
Two open protocols, both card networks, and the largest assistant on earth all converged on the same idea inside a month: the agent completes the purchase, and your site becomes a backend.
What actually shipped, minus the hype
A reality check helps here, because the demos outran the deployments. CNBC reported in March 2026 that roughly 30 Shopify merchants were live on Instant Checkout, while Walmart had made about 200,000 products purchasable inside ChatGPT. Small numbers. But the demand side is moving much faster than the merchant side: Adobe measured a 693.4% year-over-year jump in traffic from generative AI sources to U.S. retail sites during the 2025 holiday season, part of a record $257.8 billion online haul. And Morgan Stanley projects agentic shoppers could drive $190 billion to $385 billion in U.S. ecommerce spending by 2030, roughly a 10% share of online retail in the base case.
So the honest framing for 2026: the pipes are real, the volume is early, and the merchants who wire up now get to make their mistakes while the stakes are low.
| Layer | Who maintains it | What it covers | Status |
|---|---|---|---|
| ACP (Agentic Commerce Protocol) | OpenAI + Stripe, Apache 2.0 | Product feed spec, checkout sessions, delegated payment tokens | Live behind ChatGPT Instant Checkout |
| AP2 (Agent Payments Protocol) | Google + 60+ partners | Signed intent, cart, and payment mandates for any agent-led transaction | Spec published, pilots running |
| Network agent tokens | Mastercard (Agent Pay), Visa (Intelligent Commerce) | Agent-scoped card credentials on existing tokenization rails | Rolling out through issuers |
Surface one, the feed that became a storefront
Under ACP, the assistant never scrapes your PDP at purchase time. It reads your product feed, ranks candidates, and presents a buy button. Which means the feed you have treated as a Google Shopping afterthought is now the entire merchandising layer for a channel with real intent.
The failure modes are boring and brutal. A price that drifted since the last sync. An availability flag that says in_stock for a variant you sold through yesterday. A parent product with no GTIN and six unlabeled size variants the agent cannot disambiguate. Every one of these either kills the recommendation or, worse, creates a cart the checkout API has to reject.
Feed spec, minimum viable:
- id, title, description agent-readable, no keyword stuffing
- gtin / mpn required for cross-catalog matching
- price, sale_price must match checkout totals exactly
- availability synced <= 15 min, per variant
- enable_search / explicit opt-in flags per product
enable_checkout
- image_link, variants full attribute set, no orphans
Check this week: pull 50 random SKUs, compare feed price and availability against your live checkout. Any mismatch above 1–2% of SKUs means you are not ready for an agent that treats your feed as ground truth. Structured data on the PDP still matters too, since discovery surfaces lean on markup long before purchase intent forms; run your templates through our schema generator while you are in there.
This is also where segmentation strategy quietly shifts. When the agent picks the product, your influence moves upstream into which products and offers exist for which audiences, which is exactly the territory covered in our ecommerce ad targeting ROAS playbook.
Surface two, the checkout API you probably don't have
ACP expects you to expose a small set of REST endpoints. The specification, published open source on GitHub under Apache 2.0, is short enough to read over coffee, and the shape looks like this:
POST /checkout_sessions create cart, return line items + totals
POST /checkout_sessions/{id} update address, shipping option, items
POST /checkout_sessions/{id}/complete finalize with delegated payment token
GET /checkout_sessions/{id} session state
Required: Idempotency-Key header, request signature
verification, API-Version pinning, order webhooks
Two things trip teams up. First, totals latency: the agent needs final tax and shipping fast, so if your tax service or shipping-rate call takes eight seconds, the conversation moves on. Second, idempotency: agents retry. Without idempotency keys honored end to end, a network blip becomes a double charge and a support ticket you cannot see coming, because there was no session to replay.
Shopify merchants inherit most of this from the platform. Headless and custom stacks have to build it, which is a contained project (weeks, a couple of engineers) if your commerce APIs are sane, and a re-platforming conversation if they are buried inside a monolith. That build-versus-wait call is most of what we discuss in early headless and AI-dev scoping sessions right now.
Check this week: time a full quote (items + address to final totals) on your current stack. Under 3 seconds, you are fine. Over 10, that is your first engineering ticket.
Surface three, payment rails without a card form
Nobody types a PAN into a chat window. Under ACP, Stripe issues a Shared Payment Token scoped to a single merchant, a single cart, and a capped amount; you charge it through your existing PSP flow like any other tokenized payment. AP2 goes further on the trust layer, wrapping the transaction in cryptographically signed Intent, Cart, and Payment mandates so there is a verifiable answer to "did the human actually authorize this" when a dispute lands. The card networks are layering their own agent-scoped credentials underneath both.
What this means practically: you stay merchant of record. Chargebacks, refunds, and fraud screening remain your problem. Your fraud rules, though, were trained on human signals (device fingerprints, session length, mouse movement) that agent traffic simply does not produce. Expect your risk vendor to need new allowlists for verified agent signatures, and expect a few weeks of false declines while everyone recalibrates.
Surface four, order events after the buy button
The purchase happens in chat, but the relationship afterward is entirely on you, and it runs on webhooks. The agent expects order_created and order_updated events so it can tell the shopper "shipped, arriving Thursday." Slow or missing events mean the assistant shrugs at your customer on your behalf.
The subtler issue is data. You receive the order, the fulfillment details, and the customer contact needed to deliver, but the upstream conversation (what they compared, what they almost bought, why they chose you) stays inside the assistant. Your post-purchase flows become the primary place you learn anything at all, which makes first-party data capture and enrichment jobs like the ones in our data enrichment playbook considerably more valuable than they were when you could watch the whole session yourself.
What breaks first, ranked from the field
- Feed sync latency. The number one killer. Overselling through an agent is worse than overselling on-site because the cancellation happens inside someone else's UI, attached to your brand.
- Price mismatch between feed and checkout. ACP-style flows treat this as a hard error. The agent quotes one number, your API returns another, the sale dies silently.
- Slow totals. Tax and shipping calls that were tolerable behind a spinner are fatal in conversation.
- Variant ambiguity. Missing GTINs and lazy variant naming make the agent either guess or skip you. It usually skips.
- Webhook gaps. No order events, no post-purchase trust, no reorder next month.
Notice that four of the five are data problems before they are engineering problems. Feed discipline, catalog completeness, and inventory truthfulness decide whether the fancy protocol work ever matters, the same way clean segments decide whether personalization works, a point we have hammered before in our ecommerce personalization blueprint.
The part nobody prices in
Agentic checkout compresses your brand into three moments: whether the agent selects you, whether the transaction completes cleanly, and whether the package arrives as promised. Everything you used to do between click and conversion (urgency banners, cross-sells, exit intent, the whole CRO toolkit) evaporates for this channel. What remains is product data quality, operational reliability, and post-purchase experience.
That is uncomfortable for teams whose growth engine lives in on-site persuasion. It is a genuine gift for operators whose catalogs, inventory, and fulfillment are already tight. The protocols do not reward the best marketer in the room; they reward the most trustworthy system of record. Get your feed honest, your checkout API fast, and your webhooks reliable, and the assistants will treat your store the way they are designed to: as infrastructure that simply works.
Sources
- OpenAI, "Buy it in ChatGPT" announcement of Instant Checkout and ACP (openai.com, Sept 2025)
- Google Cloud, Agent Payments Protocol (AP2) announcement (cloud.google.com, Sept 2025)
- CNBC, reporting on early Instant Checkout merchant adoption (cnbc.com, March 2026)
- Adobe, 2025 holiday shopping season data via Adobe Analytics (news.adobe.com, Jan 2026)
- Morgan Stanley Research, agentic commerce market outlook to 2030