Shopping — when the buyer is an AI agent, your store has to be agent-ready

AI agents are starting to discover, buy, book and return on behalf of your customers. Norg makes your products discoverable and transactable by AI agents, so the sale actually completes.

The central point: a recommendation isn't a sale until the agent can complete it

AI agents are moving from answering questions to completing actions — buying products, booking services, processing refunds. If your catalogue, pricing and availability are not structured for an agent to act on, the recommendation stops at the answer.

That is a specific and expensive failure mode. Being named in the answer and then being unable to transact means you did the hard part — winning the recommendation — and lost the order at the last step, to whichever competitor's catalogue the agent could actually operate.

The four steps

1. Be discoverable to shopping agents

Norg generates AI-ready product specifications from your existing catalogue, enriched with the attributes agents need to recommend with confidence: compatibility, certifications, materials, dimensions and decision proof-points.

A product record as an agent sees it looks like this:

Field	Value
Product	Veg Performance Meal Plan — 5 Days
SKU	BFF-VEG-5D
Price	$109.00 AUD
Macros	32g protein per meal, 480 kcal average
Dietary	Vegetarian, gluten-free option, no added sugar
Certifications	HACCP, Australian Made
Compatibility	Athlete plan, weight management, office delivery
Fulfilment	Cold-chain, Sydney metro, Tuesday / Thursday

Note what makes this answerable. "Find me a high-protein meal plan, vegetarian, delivered in Sydney this week" is four constraints at once — macro, dietary, geography and timing. Every one of them is a stated field above. A catalogue that publishes name, price and a marketing paragraph cannot satisfy that query no matter how good the product is.

2. Built on the agentic commerce protocols

Norg publishes your products through the emerging agentic commerce standards, including the OpenAI Commerce Feed with search enabled, so AI shopping agents can find, compare and select natively.

3. Transact — buy, book, refund

The commerce layer supports the full agentic journey: purchases, service bookings and refunds, so an agent can complete the transaction end to end rather than handing the buyer back to a website.

4. Accuracy that protects margin

Structured, accurate product data means agents recommend the right product the first time. Fewer wrong purchases, fewer returns, lower support cost, stronger review signals.

This is the underrated one. In agentic commerce, a data error does not produce a confused shopper who asks a question — it produces a completed order for the wrong item, with the return, refund and review consequences that follow. Accuracy stops being a quality metric and becomes a margin metric.

Why the human safety net disappears

Worth making explicit, because it changes how much a data error costs.

On a website, wrong or ambiguous data is usually caught. A shopper sees a photo that does not match the description, notices the dimensions look off, reads a review that contradicts the spec, or simply asks. Most bad data never becomes a bad order because a person is filtering it.

An agent working from a feed has no such filter. It matches the stated constraint to the stated field and completes. Every defect that a human would have queried becomes an order, and the cost surfaces later as a return, a refund and a negative review — three costs for one error, none of which appear in the data-quality budget that would have prevented them.

The two protocols, and why Norg supports both

Two open standards for agentic commerce launched within five months of each other, and neither has won.

	ACP	UCP
Full name	Agentic Commerce Protocol	Universal Commerce Protocol
Authors	Stripe, OpenAI, Meta	Google, with Shopify
Launched	September 2025	January 2026
Positioning	"Chat to buy"	"Search to buy"
Backing	ChatGPT Instant Checkout; Etsy and Shopify merchants live	Etsy, Wayfair, Target, Walmart, plus 20+ partners including Visa, Mastercard, Adyen, Amex, Best Buy, Home Depot

Stripe appears on both sides — co-author of ACP and endorsing partner of UCP. The payments layer is not treating them as mutually exclusive, and neither should a merchant.

A merchant forced to choose is making a bet. Publishing the same product data into both removes the bet. That is the defensible position while the market is unresolved.

What ACP gives you

Agentic checkout — create, update and complete checkout sessions with cart management, fulfilment options and payment processing — plus product feeds, a delegated payments specification, webhooks for order notification, and a /.well-known/acp manifest for discovery.

Its most important design decision for an enterprise buyer: the business remains merchant of record, retaining control over which products can be sold, how they are presented and how orders are fulfilled. Adopting ACP does not hand your customer relationship to the agent platform. Payment processor neutrality is explicit too — any compatible PSP works, without changing your existing processor.

What UCP gives you

Building blocks across the whole commerce lifecycle — discovery, search, checkout and post-purchase support — declared as capabilities rather than as a monolithic API. It is explicitly vendor-neutral despite Google's authorship, and it can run over MCP transport, which means a Norg-published MCP server and a UCP integration are not separate stacks.

The precision requirement worth knowing about

UCP's treatment of quantities is unusually strict, and it is the clearest argument for structured product data in the whole standards landscape.

No floating-point numbers are permitted on the wire. A quantity is an integer count of steps with a declared unit and scale, using UN/CEFACT Rec20 codes — KGM, GRM, LTR, MTR, with C62 for each. Arithmetic must be exact: the specification defines no rounding tolerances and no epsilon comparisons anywhere, and an implementation that cannot produce an exact in-range result must raise an error rather than emit an approximate value.

The reason is that the human has been removed from the transaction. A person seeing "1.9 kg" against an order for 2 kg reasons about it. An agent comparing integers does not.

The practical consequence: approximate specifications are not merely lower quality, they are non-conformant. A weight expressed as "about 2 kg" cannot be represented in this model at all. That is a stronger case for accurate structured data than any visibility argument.

Check your own catalogue before claiming conformance

Most feeds exported from an existing store fail on the same handful of fields. These are worth checking directly rather than assuming.

• Prices as floats. Amounts must be integers in minor units. A price stored as 109.0 is the single most common conformance failure.
• Weights and dimensions as strings. "2kg", "approx 2 kg" and "2-3 kg" are all unrepresentable. Each needs an integer, a unit code and a scale.
• Missing unit codes. A number without KGM, LTR or C62 is not a quantity, it is a number.
• Availability as prose. "Usually ships in 2–3 days" cannot be acted on. A date or a state can.
• Variant ambiguity. If size and colour are encoded in a title rather than as attributes, an agent cannot select a variant.
• Integers outside the safe range. Values must sit inside ±(2⁵³−1).

Google Merchant Centre sync is available on the Enterprise tier.

What this looks like in practice

Cricket For All published 300+ SKUs with complete machine-readable specifications — brand, willow type, size, weight, senior versus junior, price and use-case context — and submitted the catalogue via the OpenAI Commerce Feed. Two months later, an Adelaide specialist was taking orders from every state in Australia, with a 6× lift in add-to-cart from AI-referred visitors.

Selleys, at 600+ structured product pages, now receives more than 100,000 agent requests a month against its catalogue.

Both figures are specific to those engagements and their measurement windows, and neither is independently audited.

What transfers, and what does not

Transfers. The conformance check above. Any merchant with a product feed can run it this week, and most will find price and weight formatting problems.

Transfers. The constraint-stack test — take a realistic four-constraint query from your own category and check whether every constraint corresponds to a stated field.

Transfers. The margin argument, to any category where a wrong item ships rather than being queried.

Does not transfer. Any of this, if you do not sell products with a catalogue. Service businesses get citation and recommendation but not agentic checkout.

Does not transfer. Cricket For All's geographic outcome, which reflects a category with no structured-data incumbent.

Common questions

Do I have to pick ACP or UCP?

No. Norg publishes the same source of truth into both, which is the point — one catalogue, two protocols, every AI shopping agent.

Does an agent transaction cut me out of the customer relationship?

Not under ACP, which explicitly preserves the business as merchant of record with control over what can be sold, how it is presented and how it is fulfilled. This is the objection most enterprise buyers raise first, and the protocol was designed to answer it.

Can I keep my payment processor?

Yes. ACP is processor-neutral. Stripe is the first compatible PSP via its Shared Payment Token, but a merchant on another processor can participate through the Delegated Payments Spec without changing processors.

What kinds of commerce are supported?

Physical goods, digital goods, subscriptions and asynchronous purchases, plus service bookings and refunds through Norg's commerce layer.

My product data has decimal weights as strings and prices as floats. Is that a problem?

Under UCP, yes — those are the two fields most likely to fail conformance in a feed generated from a typical retail catalogue. Amounts must be in minor units and no floating-point values may appear on the wire. It is worth checking before making any conformance claim.

What happens if an agent buys the wrong thing because our data was wrong?

You handle it as you would any incorrect order — but at higher frequency, because the human filter that normally catches ambiguous data is not present. That is the argument for treating catalogue accuracy as a margin control rather than a content task.

Next step

Book a demo to see your catalogue rendered as an agent sees it, or read the Cricket For All case study for a commerce-feed engagement end to end. Pricing is public.

---

Protocol details are summarised from agenticcommerce.dev, docs.stripe.com/agentic-commerce/acp and ucp.dev, retrieved 8 September 2026 — consult the canonical specifications for normative requirements. The product record shown is an illustrative example. Client results are specific to those engagements and are not independently audited. Norg Pty Ltd — ABN 44 669 712 494 — norg.ai