Instant Checkout: Merchant of Record, Refunds, Disputes

Purchase-in-chat is easy. The Merchant of Record problem is not

Agentic commerce is moving from “interesting” to “measurable” fast. Adobe reports traffic to US retail sites from generative AI sources grew 4,700% YoY in July 2025. These visits also show higher intent: more time on site, deeper navigation, and lower bounce rates. For BCG, this data is the signal: agentic shopping is no longer a side quest.

Now the hard part: getting from “the assistant recommended it” to “the order completed” without breaking trust.

And that is where the Merchant of Record question becomes the whole game.

The MoR decision defines the product decision

In Instant Checkout, OpenAI is explicit: OpenAI is not the merchant of record. The buyer’s purchase is from the merchant shown, and that merchant handles payment, shipping, returns, and support.
On the technical side, OpenAI’s delegated payments spec repeats the same idea: settlement, refunds, chargebacks, and compliance stay with the merchant and their PSP.

This isn’t a checkout choice. It’s a choice about accountability: who takes payment and who owns refunds, disputes, and the customer experience when things go wrong.

In practice, MoR means: who takes payment, who appears on the statement, who owns chargebacks, who must honor refunds, and who is responsible for tax and compliance in the flow.

The uncomfortable truth: trust is operational

Buyers don’t experience “protocols”. They experience whether the price at checkout matches what they were shown, whether delivery is actually available for their address, whether refunds happen when promised, and whether support is straightforward instead of turning into a week-long email thread.

If those things fail, the platform’s reputation takes a hit, even if legally the merchant owns it. That tension is exactly why MoR becomes such a loaded choice.

Three models platforms are converging on

You can see the industry clustering around three approaches. None is “free”.

1) Merchant is MoR, platform stays light

This is the model OpenAI positions for Instant Checkout: merchants keep control of orders, payments, fulfillment, and the customer relationship.

The upside: the platform avoids becoming a payments, tax, disputes, and support company.

The catch: your experience is only as strong as the merchants’ data quality and operational discipline. Even if you stay light legally, you still carry reputational risk. You need guardrails, monitoring, and clear failure handling, or the whole channel feels unreliable.

2) Platform becomes MoR (or delegates it to a MoR partner)

This buys you consistency. One set of rules. One support posture. One refund policy surface.

It also buys you work you probably did not want: the MoR is on the hook for payment acceptance, disputes, refunds, and often tax compliance responsibilities.
That is a second business line, not a feature.

3) Merchant remains MoR, but the “agentic checkout layer” is shared infrastructure

This is the middle path teams quietly like.

Merchants keep funds and legal positions. The platform (or an infrastructure layer) standardizes what is hardest to get right quickly:

• product data that is complete enough to rank and convert
  
• checkout behavior that is predictable across merchants
  
• order state that stays consistent end to end

This matters because Instant Checkout relies on two synchronized layers: the Product Feed (for discovery) and the ACP-based checkout endpoints (for transaction).

If your feed is thin, your products may never even enter the conversation. OpenAI calls out that required fields make products display correctly, while recommended attributes (rich media, reviews, performance signals) improve ranking, relevance, and trust.
That is not a marketing problem. That is an infrastructure problem.

Where “purchase-in-chat” tends to break in real life

Most failures happen after the buyer has already made a choice and clicked into checkout. The product looks available in the feed but turns out to be out of stock at checkout. The price shown in the recommendation doesn’t match the price returned by the checkout flow. Shipping constraints only appear at the last step. Order status updates arrive late or not at all. And refunds may be processed on the merchant side without the assistant experience ever reflecting the final state.

OpenAI’s own production guidance is blunt: the merchant is responsible for refunds and chargebacks as MoR, and order state needs to stay synchronized.

If you are a platform, you cannot outsource the buyer’s confidence. You can only decide how you will build it.

A practical way to choose your model

Before you talk about protocols, pick your answers to five questions:

• Whose name should the buyer see on the charge?
  
• Who can approve, deny, and refund with one consistent policy?
  
• Who is equipped to handle disputes at scale?
  
• Who owns the “truth layer” for catalog, price, availability, and shipping?
  
• Who is accountable when the assistant experience promises something the merchant cannot deliver?

If you cannot answer these cleanly, the protocol work will not save you. It will just make failure faster.

The actual opportunity

Agentic commerce is not just “new distribution”. It is a new contract between three parties: buyer, merchant, and assistant platform.

Platforms that win will not be the ones with the prettiest chat UI. They will be the ones that make purchase-in-chat feel boring in the best way: accurate, predictable, and safe.

And boring, in commerce, is how trust scales.