Marketing to Machines

(Was this newsletter forwarded to you? Sign up here.)

OpenAI's first attempt to let you buy things inside ChatGPT just failed.

The facts:

• OpenAI launched Instant Checkout in late 2025 and expanded it in February.

• It pulled back in March 2026 after fewer than 15 of Shopify's millions of merchants ever went live.

• OpenAI is now rerouting shoppers to retailers' own apps instead of completing the purchase itself.

Easy to file agentic commerce under "overhyped" and move on. That's the expensive mistake.

The consumer flow flopped. The rails underneath it are being laid faster than ever. Stripe and OpenAI shipped the Agentic Commerce Protocol. Google published its Agent Payments Protocol and checkout inside AI Mode. Mastercard Agent Pay and Visa Intelligent Commerce moved the payment infrastructure into place. The first product died. The standards it ran on are spreading to everyone else. Infrastructure is patient.

My read

Most coverage treats this as a new channel to optimize for. That's the comfortable read, and it's wrong.

For thirty years, marketing has run on one assumption: a human is on the other end of the decision. Every funnel, every brand campaign, every loyalty program, every dollar of retail media is a bet that if you reach the right person and move them, they choose you. Agentic commerce makes that assumption optional. When an agent does the buying, you don't lose the sale. You might win it. You lose everything around the sale that was the actual point.

You lose the data. The agent ran the query and made the call. You learn a transaction happened. You don't learn who, why, or what they'll want next. The first-party signal you spent five years and a CDP budget assembling gets intercepted upstream.

You lose the upsell. The agent bought what it was told to buy. The order bump, the bundle, the "customers also liked," the merchandising your whole site is built around: the agent never sees it.

You lose the loyalty loop. There's no repeat visit to reward and no email to send, because the relationship is now between the buyer and their agent. You're a supplier to it.

Add it up, and a brand that doesn't defend its position becomes a commodity input to a platform that owns the demand. You make the product. Someone else owns the customer. There's a word for that, and it isn't "brand." It's "vendor."

The Disintermediation Stack

Separate what an agent can take from you into three layers. They fall in order, so defend them in order.

Layer one is discovery. The agent decides which products make the consideration set at all. This is the layer everyone's panicking about, and a cottage industry already named the response: "agentic engine optimization," "share of model," the scramble to be the product the AI names. It matters. It's also the shallowest layer, because being named isn't being chosen, and being chosen once isn't being kept.

Layer two is selection. Among the options it surfaced, the agent picks. Price, specs, reviews, availability, and whatever the platform is quietly incentivized to favor. This is where most brands lose, because an agent optimizing for a stated preference is brutally rational in a way human shoppers never were. If your only edge was that people liked how you made them feel, the agent doesn't feel.

Layer three is the relationship. Who the buyer returns to. Who has the data. Who they trust enough to skip the comparison next time. This is the deepest layer and the only genuinely defensible one, because it's the thing an agent is supposed to spare the user from doing: deliberating. A buyer who tells their agent "just reorder from them" handed you the one position the platform can't easily tax.

The mistake almost everyone is making is pouring the budget into layer one while layer three erodes. Winning discovery after you've lost the relationship just means paying, forever, to be reconsidered from scratch on every purchase.

The objection

The fair pushback: agent checkout already flopped once, so maybe this never reaches scale and the relationship is safe. Could be. But read the failure again. Instant Checkout didn't die because shoppers rejected the idea. It died because merchant onboarding and real-time inventory were hard. Those are engineering problems, and engineering problems get solved. The protocols shipping right now exist specifically to solve them. Betting your customer relationship on the plumbing staying broken is not a strategy. It's a hope.

What I'd do with it

Three moves, in order of how hard they are to reverse later.

Find out if your products are already being transacted by agents, and what you can and can't see when they are. Most teams genuinely don't know. You can't defend a layer you aren't measuring.

Decide which layer you're fighting for and fund it like you mean it. Put every dollar into discovery and none into the direct relationship, and you've chosen to be a vendor without admitting it. The durable advantage is the same one that holds when AI makes building free: distribution you own, not distribution you rent. (That's the argument in Distribution Is the Last Moat.)

Take the relationship-versus-distribution tradeoff to your CEO as a strategic choice while it's still a choice. Every brand is about to face the version publishers faced with Google and merchants faced with Amazon: take the platform's distribution and surrender the relationship, or defend the relationship and pay for harder distribution. There's no option that's both free and safe.

The machines are going to do a lot of the buying. The only question that matters is whether, when they do, there's still a relationship with your name on it, or just a transaction with your product in it.