You took the money. Your name is on the card statement. You'll handle the return. On the narrow question of merchant of record, the customer is unambiguously yours, which is exactly what makes the term so dangerous to lean on.

A shopper asks ChatGPT for a linen shirt that runs true to size and ships before Friday. Ten minutes later an order lands in your admin. Six days after that she wants a medium instead, and your support team handles the exchange.

So whose customer is she?

The instinctive answer is "ours, obviously." And on the specific question of merchant of record, that instinct is correct. Merchant of record is a payments term that has quietly been drafted into doing strategy work it was never designed for. It tells you who pays for a chargeback. It does not tell you who owns the customer.

Three questions live inside "who owns the customer." In agentic commerce they now have three different answers, and only the first one is settled in your favor.

What Does Merchant of Record Actually Mean?

It is worth separating MoR from the roles it gets confused with, because agentic commerce coverage conflates them constantly and the differences decide who eats the loss.

An MoR-as-a-service provider like Paddle or FastSpring charges a premium over pure processing precisely because it absorbs that stack. Apple and Google do the same on their app stores. So when a protocol designer decides who the MoR will be, they are deciding who eats tax exposure, fraud loss and dispute cost. That decision has been made consistently, and it has been made in your favor.

Question One: Who Is Responsible for the Transaction?

OpenAI and Stripe open-sourced the Agentic Commerce Protocol under Apache 2.0 on September 29, 2025, and OpenAI's production documentation is blunt about it. The business selling the goods and taking payment is the merchant of record. OpenAI is not. The customer sees the merchant's name on their statement as though they had bought direct, and refunds and chargebacks belong to the merchant for the stated reason that the merchant accepted the payment.⁴

The mechanism that makes this work is the Shared Payment Token. The shopper stores a card with the AI platform. At confirmation, Stripe mints a single-use token scoped to one merchant and one cart total, with a maximum amount and an expiry. The platform hands that token to the merchant, who charges it through their own processor. Money flows merchant-ward, card data does not flow platform-ward, and PCI scope stays roughly where it was.

Google took the same position with the Universal Commerce Protocol, announced at NRF in January 2026 and co-developed with Shopify: businesses remain the seller of record. Gemini's agentic checkout is more conservative still, adding the item to your cart on your site and completing with Google Pay rather than hosting a checkout of its own.

The instructive exception is Perplexity, because it reversed. The original Buy with Pro made Perplexity itself the intermediary and merchant of record. The PayPal-powered Instant Buy that replaced it in November 2025 puts retailers back at the center as merchant of record, handling purchase, service and returns.

Where Responsibility Is Still Genuinely Unresolved

Tax law may not care what the protocol says. US marketplace facilitator statutes, which spread after Wayfair in 2018, shift sales tax collection and remittance onto a platform once it crosses economic nexus thresholds, commonly around $100,000 in sales or 200 transactions in a state. The EU's deemed supplier rules, live since July 2021, do something similar for VAT.

Both hinge on control. The EU explicitly excludes platforms that only process payments, list products or redirect customers, and the European Court of Justice has held that a platform which authorizes the charge, controls payment or sets the terms is irrebuttably the supplier. Today's protocols are engineered to sit on the excluded side of that line, and that is not an accident. Note the asymmetry, though: a platform can become a marketplace facilitator for tax purposes without becoming the merchant of record under card network rules. They are separate determinations.

Agent error has no clear owner. Google's AP2, announced in September 2025 with more than sixty launch partners, wraps each purchase in three cryptographically signed Mandates carried as W3C Verifiable Credentials: an Intent Mandate for the shopper's standing instruction, a Cart Mandate locking exact items and price, and a Payment Mandate carrying the minimum the network needs.⁵

It produces a non-repudiable audit trail, which is real progress. What it does not produce is an answer. The mandates bind to the user's signing key, not the agent's; agent identity is a separate problem being solved separately by Visa's Trusted Agent Protocol and Mastercard's Verifiable Intent. Analysts at Everest Group have made the point directly: AP2 gives you audit trails but does not yet establish who is liable when an agent executes a mistaken or fraudulent transaction. Until case law or network rules settle it, the party holding MoR is the party holding the bag by default.

Question Two: Who Receives the Data and the Consent?

Then there is the part that surprises people. OpenAI's merchant and consumer documentation states that merchants should not sign ChatGPT buyers up for marketing emails from those orders. You hold the legal responsibility of the seller and a contractual constraint on the most basic thing a seller does with an email address.

The law says something adjacent but not identical, and the distinction is the practical one:

Under GDPR you are generally the controller. You decide why and how the data is processed, so subject access requests, erasure and lawful basis land on you. The platform is typically a processor, or a separate controller for its own purposes. An AI agent cannot consent or bear obligations, so responsibility falls back to the humans and companies on either side of it.

Which exposes the structural problem nobody has solved. Consent was designed for a person looking at a screen, and agentic commerce removed the screen. If an agent acts for a logged-in customer who already consented on your site, does that consent travel with the agent to a third-party surface? If a shopper exercises a CCPA opt-out with you, does it propagate to the platform that brokered the sale? There is currently no established mechanism for carrying consent across interfaces, and the party carrying the controller obligations is you. We went deeper on the adjacent version of this question in AI agent memory: personalization versus surveillance.

Question Three: Who Keeps the Relationship?

Work through the asymmetry. The shopper asks the AI. The AI shapes the consideration set. The order arrives. Six days later the shirt is too small, and there is no agentic surface in existence that will run your exchange, quote your restocking policy or tell her where the parcel is. WISMO, returns, exchanges, sizing disputes and warranty claims all route to you. You pay the support cost for a relationship you were only partially allowed to form.

Consultancies have been circling this for a year and the language is unusually sharp. Bain has warned that agents could relegate retailers to drop shippers or commoditized fulfillment mechanisms. BCG describes the risk of retailers being sidelined into background utilities in agent-controlled marketplaces. Adyen, quoting a merchant, put it most economically: whoever owns intent owns the customer.

The counterweight is that the traffic is genuinely good, when you can see it.

Three honest caveats, because these numbers get quoted badly. The volume is still small, so high-converting and low-volume are both true at once. AOV signals conflict: Shopify has reported AI-referred sessions with roughly 14% higher AOV, while a 94-brand study found ChatGPT-referred AOV around 14% lower than organic. And lifetime value data on AI-acquired customers does not meaningfully exist yet. If a vendor quotes you one, ask how they got it.

The strongest evidence for what ownership is actually worth came from Walmart. Its ChatGPT Instant Checkout pilot converted at about one-third the rate of Walmart.com, a figure its EVP of AI acceleration disclosed in March 2026.³ Walmart ended the pilot and embedded its own assistant, Sparky, into the AI surfaces instead. Same demand, same channel, materially better performance, and the only variable that changed was who controlled the transaction layer. We unpacked the full sequence in Two Front Doors.

What March 2026 Changed, and Why It Helps You

OpenAI wound down in-chat Instant Checkout in March 2026, conceding in its own announcement that the first version lacked the flexibility it wanted to offer.² Adoption had been thin. Fewer than about 30 Shopify merchants ever went live, against the millions once implied.

A lot of coverage read this as agentic commerce stalling. It isn't. Discovery volume kept climbing, and Shopify switched Agentic Storefronts on by default for eligible US stores that same month. The full story of the rise and retirement is in what happened to ChatGPT Instant Checkout.

What actually happened is that the industry converged on discover in AI, buy on site, and that is the most favorable configuration a brand could have asked for. Your site is back in the flow, which means your merchandising logic, your bundles, your loyalty enrollment, your consent capture and your assistant are back in the flow with it.

The window is narrow. Consumer trust has not caught up with the protocols: Visa's Trust Index, published in September 2026 from a survey of 2,065 US consumers, found only 23% trust generative AI to handle payment transactions. That gap is your runway.

Nine Questions Before You Switch It On

If you want the full cross-platform strategy, including where Gemini, Perplexity and Copilot diverge from OpenAI, our agentic commerce playbook covers the ground above this article, and what is agentic commerce covers the definitions beneath it. What follows is the diligence layer.

• 1. Who is the merchant of record, in writing, and whose name appears on the card statement?
• 2. Who handles chargebacks, refunds, fraud liability and tax remittance? Get this before you get excited about the traffic.
• 3. Exactly which customer fields do we receive, and which does the platform retain?
• 4. Are we permitted to invite marketing consent from these buyers, and through which channel?
• 5. Does checkout complete on our site or inside the platform? This one decides who controls upsell, loyalty and consent capture.
• 6. What is the all-in take rate once platform fee and processing are stacked?
• 7. How are these orders tagged so we can attribute them later?
• 8. Which SKUs, markets and account types are actually eligible? The answer is usually narrower than the announcement.
• 9. If the surface shuts down next quarter, what do we keep? No longer a hypothetical.

Own the Layer You're Already Liable For

If you are going to carry merchant-of-record responsibility regardless, the strategic move is to make sure it arrives attached to a relationship rather than just a cost. This is the layer we build, so read what follows as an interested party's view and hold it to the standard you would apply to anyone else.

Get visible in the discovery layer at SKU level. Being cited by ChatGPT, Gemini and Perplexity is an acquisition channel now, and most brands have no instrumentation for it. Alhena's AI Visibility tracks how individual products surface across AI engines, flags feed and structured data gaps, and ties visibility to revenue rather than to brand mention counts.

Convert on the surface you control. A shopper arriving from an AI answer lands mid-decision, several exchanges into a conversation you never saw, carrying questions the agent could only half-answer from your feed. An owned on-site assistant is where you capture the intent data the platform withheld and where consent can be asked for properly. Alhena's AI Shopping Assistant works for your brand rather than a third-party surface. Tatcha reported 3× conversion and 38% AOV uplift after taking that layer seriously. Customer-specific results, not promised outcomes.

Keep the post-purchase relationship instead of just absorbing it. WISMO, returns and exchanges come to you whichever surface sold the item. Alhena's AI Support Concierge handles those as agentic actions rather than tickets, with order lookups, tracking and refunds executed rather than described. Manawa cut response times from 40 minutes to 1 minute with 80% inquiry automation. Every one of those interactions is a retention opportunity the platform is not competing for.

The brands that come out of this period intact will be the ones that stopped assuming the first answer covered the second.

Frequently Asked Questions