Somewhere this quarter, a customer bought something from a merchant who never saw them. No product page load, no cart abandonment email, no session in Google Analytics — an AI shopping agent read a structured feed, checked stock and returns terms, and pushed a payment mandate through to settlement. That is the practical reality of the agentic commerce protocol era, and it stopped being a conference demo the moment card networks started underwriting the payment leg. For business owners, the uncomfortable part is that visibility in this channel has nothing to do with your theme, your hero image, or your conversion-optimized checkout flow. If your catalog is not machine-readable and your policies are not stated in structured terms, the agent picks a competitor who did the boring work.
What’s new in the agentic commerce protocol
Two stacks matter, and they solve different halves of the same problem. The Agentic Commerce Protocol (ACP), open-sourced by Stripe and OpenAI, is the merchant-facing half. It defines how an agent discovers your products, builds a cart, asks for tax and shipping quotes, and submits an order against your system of record. Recent spec revisions tightened the merchant-side contract: richer product feed attributes, a clearer session lifecycle for carts an agent holds open across turns, and explicit fields for returns and cancellation windows. That last part is not cosmetic. Agents refuse to complete purchases when the risk terms are unreadable, so an unstated return policy now functions as a hard conversion blocker rather than a footer link nobody clicks.
The second stack is Google’s Agent Payments Protocol (AP2), which handles trust rather than catalog. AP2’s core idea is the mandate: a cryptographically signed, verifiable record of what the human actually authorized. It splits into an intent mandate (“buy running shoes under $150 in size 11”) and a cart mandate (the specific, priced cart the agent assembled). When a dispute lands, the mandate chain answers the question that has haunted every agentic checkout pilot — did a person approve this, or did a model hallucinate a purchase? Card-network participation moved AP2 from interesting to inevitable. Once the networks accept mandates as dispute evidence, the liability question that kept large merchants on the sidelines has an answer.
Meanwhile, the consumer surfaces shipped. Instant checkout inside ChatGPT went live for merchants on supported platforms, Shopify and Etsy catalogs got wired in, and Perplexity, Microsoft Copilot and Google’s shopping surfaces all run some version of agent-assisted buying. The distribution exists. The bottleneck is merchant readiness — most storefronts still publish prices only inside JavaScript-rendered pages, which is precisely the format an AI shopping agent handles worst.
Why it matters
- Discovery moved upstream. An agent shortlists from feeds and structured data before it ever touches your site. If you are not in the feed, you are not in the consideration set, and no amount of on-page conversion work rescues you from that.
- Your returns policy is now a ranking factor. Agents optimize for user-stated constraints and downside risk. A clearly expressed 30-day, free-return window in machine-readable form beats a vague “hassle-free returns” banner every time.
- Price and stock accuracy became load-bearing. A stale feed price that fails at the AP2 cart-mandate step is a rejected order, not a “sorry, price changed” upsell moment. Feed drift costs revenue directly.
- Attribution breaks before you notice. Agent-placed orders arrive with thin referrer data and no session history. If you make budget decisions from last-click reporting, this channel looks like unattributed direct traffic and gets zero credit for months.
- Payment liability is being renegotiated right now. Mandates, delegated payment credentials and network-level agent flags determine who eats fraud losses. Merchants who adopt the standard flows get the protections; merchants improvising get the chargebacks.
- Small merchants gain real leverage. Agents do not care about brand recall or ad budgets. Clean data, honest specs and fast fulfillment are the ranking inputs — the most level playing field ecommerce has offered in a decade.
How to use the agentic commerce protocol today
-
Audit whether an agent can read you at all. Fetch your own product page the way a headless agent does — no JavaScript execution — and check whether price, availability and SKU survive.
curl -sL -A "Mozilla/5.0 (compatible; ShoppingAgent/1.0)" \ https://yourstore.com/products/your-best-seller \ | grep -Eo '"(price|availability|sku|priceCurrency)":[^,]*'If that returns nothing, your prices live only in client-side JavaScript and you are invisible to a large share of AI shopping agents.
-
Ship Product structured data with the fields agents actually use. Shipping details and return policy are the two most commonly omitted, and the two agents weigh most heavily.
<script type="application/ld+json"> { "@context": "https://schema.org", "@type": "Product", "sku": "AILG-1042", "name": "Merino Trail Runner", "gtin13": "0123456789012", "brand": { "@type": "Brand", "name": "Northbound" }, "offers": { "@type": "Offer", "price": "129.00", "priceCurrency": "USD", "availability": "https://schema.org/InStock", "itemCondition": "https://schema.org/NewCondition", "shippingDetails": { "@type": "OfferShippingDetails", "shippingRate": { "@type": "MonetaryAmount", "value": "0", "currency": "USD" }, "deliveryTime": { "@type": "ShippingDeliveryTime", "handlingTime": { "@type": "QuantitativeValue", "minValue": 0, "maxValue": 1, "unitCode": "DAY" }, "transitTime": { "@type": "QuantitativeValue", "minValue": 2, "maxValue": 5, "unitCode": "DAY" } } }, "hasMerchantReturnPolicy": { "@type": "MerchantReturnPolicy", "applicableCountry": "US", "returnPolicyCategory": "https://schema.org/MerchantReturnFiniteReturnWindow", "merchantReturnDays": 30, "returnMethod": "https://schema.org/ReturnByMail", "returnFees": "https://schema.org/FreeReturn" } } } </script> -
Publish an agent-readable product feed. Most platforms can emit one; if yours cannot, generate a nightly file and serve it at a stable URL. Keep it flat and boring.
id,title,description,link,image_link,price,availability,brand,gtin,condition,shipping_price,return_window_days AILG-1042,Merino Trail Runner,Lightweight trail shoe with merino lining,https://yourstore.com/p/1042,https://cdn.yourstore.com/1042.jpg,129.00 USD,in_stock,Northbound,0123456789012,new,0.00 USD,30 -
Stop blocking the agents you want to sell to. Many stores inherited a
robots.txtthat blanket-denies AI crawlers. Separate the training crawlers you may want to exclude from the shopping agents that bring orders.User-agent: OAI-SearchBot Allow: / User-agent: ChatGPT-User Allow: / User-agent: PerplexityBot Allow: / User-agent: Google-Extended Allow: / Sitemap: https://yourstore.com/sitemap.xml Sitemap: https://yourstore.com/feeds/products.csv -
Wire up an ACP checkout endpoint, or let your platform do it. Shopify, Stripe-hosted and several cart platforms expose this for you. If you run custom infrastructure, the merchant contract takes this shape:
POST /agentic_commerce/checkout_sessions Content-Type: application/json Authorization: Bearer $MERCHANT_API_KEY { "items": [{ "id": "AILG-1042", "quantity": 1 }], "buyer": { "email": "buyer@example.com" }, "fulfillment_address": { "line_one": "500 Market St", "city": "San Francisco", "state": "CA", "country": "US", "postal_code": "94105" } }Your response returns line items, tax, shipping options and a total the agent can present for human approval — exactly the artifact that becomes an AP2 cart mandate.
-
Tag and measure agent orders separately. Flag them at ingest so this channel does not disappear into “direct.”
SELECT DATE_TRUNC('week', created_at) AS wk, COUNT(*) AS orders, SUM(total) AS revenue FROM orders WHERE source_channel IN ('acp','agent','instant_checkout') GROUP BY 1 ORDER BY 1 DESC; -
Test like a buyer’s agent would. Run this prompt against a frontier model with browsing enabled, for your product and two competitors:
You are a shopping agent with a $150 budget. Find the best men's size 11 trail running shoe available today. For each candidate, report: exact price, in-stock status, shipping cost, delivery estimate, and return window. Reject any product where you cannot verify all five from the merchant's own pages or feed. Then recommend one and explain the rejections.Whatever the agent could not verify about you is your prioritized backlog. This single test is worth more than a quarter of CRO theorizing.
How it compares
| Standard | Backed by | Solves | Merchant lift | Best fit |
|---|---|---|---|---|
| ACP (Agentic Commerce Protocol) | OpenAI, Stripe | Discovery, cart, order submission against merchant system of record | Low on Shopify/Stripe; moderate for custom carts | Any merchant wanting agentic checkout inside ChatGPT-class surfaces |
| AP2 (Agent Payments Protocol) | Google plus card networks and 60+ partners | Verifiable human authorization, mandates, dispute evidence | Mostly PSP-side; merchants inherit it | High-value or regulated carts where authorization proof matters |
| Merchant product feeds (Google/Meta style) | Ad platforms, retail search | Structured catalog exposure | Low — most platforms auto-generate | Baseline. Do this first; it feeds everything else |
| MCP storefront servers | Anthropic-originated, broadly adopted | Letting agents call your business logic directly as tools | Moderate — you build and host a server | Complex catalogs, quoting, B2B, configurable products |
| Plain schema.org + clean HTML | Open web | Readability by any agent, no integration required | Lowest | Every merchant, immediately, as the floor |
These are layers, not rivals. Feeds and schema.org make you legible, ACP makes you buyable, AP2 makes the payment defensible, and MCP exposes logic a static feed cannot express. A merchant who does only the first two still captures most of the near-term upside.
What’s next for ecommerce AI agents in 2026
Expect consolidation on the payments side first. Delegated payment credentials — network tokens scoped to a single agent-initiated purchase — let an agent transact without ever holding a raw card number, and their rollout determines how fast large merchants move. Watch the agent-initiated transaction indicators the networks are standardizing. Once those flags carry defined liability rules, the risk teams currently blocking these integrations lose their main objection, and adoption goes from pilot to default.
On the merchant side, the interesting fight is over post-purchase. ACP handles the buy; returns, exchanges, order status and support are still human-shaped workflows an agent has to guess at. The specs are moving toward structured post-purchase operations, and the merchants who expose those as callable endpoints will win the repeat business, because the agent that can process a return in one turn is the agent a customer keeps using. Expect the same for subscriptions and reorders, where agents have obvious leverage.
The strategic risk worth naming: this channel is intermediated. Whoever owns the agent surface owns the customer relationship, the ranking logic, and eventually the take rate — the same arc as marketplaces and app stores. The defensive play is not to abstain, since abstaining just means invisibility. Keep a direct relationship the agent cannot disintermediate — email lists, owned content, product quality, post-purchase experience — while taking the agentic volume. Sell through the channel; do not become dependent on it.
Frequently Asked Questions
Do I need to implement ACP myself?
Probably not. If you sell on Shopify, use Stripe, or run a major cart platform, agentic checkout support arrives as a platform feature you enable. Custom or headless builds need real integration work — a checkout session endpoint, quote calculation, and order creation against your system of record. Either way, the structured data and feed work in step one benefits you regardless of which protocol wins.
What is the difference between ACP and AP2 in plain terms?
ACP is the shopping and ordering conversation between an agent and your store. AP2 is the proof that a human authorized the resulting payment. ACP gets the cart built; AP2 makes it safe to charge. They are designed to coexist, and most merchants will touch ACP directly while inheriting AP2 through their payment provider.
Will agent orders cannibalize my direct traffic?
Some, yes — for commodity repurchases especially. But the larger early effect is incremental: agents surface merchants that ad budgets never would have reached. The real cost is attribution confusion, not lost revenue. Tag agent orders at ingest so you can see the split honestly before you cut spend anywhere else.
How do I keep pricing accurate for AI shopping agents?
Regenerate your feed on the same schedule as your price changes, not nightly-by-default. Serve availability from live inventory rather than a cached value, and make your checkout session endpoint the authority. If an agent quotes a stale price, your endpoint should return the current one and let the agent re-confirm with the buyer rather than silently accepting a bad total.
Is this only relevant for physical products?
No. Digital goods, courses, subscriptions and services all benefit, and in some ways they are easier — no shipping quotes or returns logistics. The requirement is identical: a clear price, a clear scope of what is delivered, and a clear refund policy in machine-readable form.
What should I do this week if I only have a few hours?
Run the curl check in step one, add the schema.org block with shipping and return policy to your top twenty products, and unblock shopping crawlers in robots.txt. That is a few hours of work, and it covers the majority of what an agent needs to consider you a valid candidate. Everything else — ACP endpoints, MCP servers, mandate flows — can follow once you know you are readable.
Go deeper than this article
This article covers the essentials. Our Technical & Coding eguide collection gives you the full step-by-step playbooks — prompts, workflows, and copy-paste recipes built for exactly this work.
