The Search Is Over Before It Begins

Tobi Lütke thinks AI isn’t changing how people shop. It’s changing when they start — and that breaks the business model underneath the ads.

On a podcast released October 28, Alex Heath asked the CEO of Shopify why his company welcomed Meta’s Muse while Amazon blocked it outright. Tobi Lütke’s answer was not about AI policy. It was about what each company sells.

Amazon runs a digital department store. Customers arrive, search, compare, and buy — and the platform earns from both the transaction and the attention, because merchants pay for placement in a store the platform controls. Shopify sells software to independent merchants and takes a cut of what they sell. “The more places merchants can sell their products, the greater the value of Shopify,” he said. Whether the order comes from a human or an agent is, from that vantage, a detail.

The same logic explains the two responses. Muse threatens Amazon’s control of the entrance. It offers Shopify another door.

But the more consequential claim in the 84-minute conversation was not about who lets agents in. It was about what agents are being asked to do.

From product demand to life demand

The standard mental model of e-commerce is a funnel: a person has a product need, searches, compares, decides. Twenty years of optimization has been spent shortening the distance between a query and a checkout.

Lütke described a different starting point. Instead of searching for a treadmill and then separately researching dumbbells, benches and mats, some consumers now hand the whole problem over: buy everything I need to set up a home gym. He extends the example to a weekend 3D printing project with a child — the agent assembles the project, works out which electronics, magnets and filament are required, checks delivery deadlines, and sources locally where needed. He imagines a customer starting from a photograph of the room or a floor plan.

The prompt change looks small. The work behind it is not. An agent handling a home gym is not matching a query to a product. It is solving a constrained project: budget, floor space, fitness goal, compatibility between items. It has to decide what is worth buying before it decides what to buy.

That relocates the commercial moment. In the old model, the platform competed for attention at the point of product comparison. In the new one, the comparison has already happened by the time a platform is called — and it happened inside a conversation the platform cannot see. Lütke acknowledges this directly: Shopify cannot observe the planning conversation that produced the order.

Some care is warranted here, and it belongs in the record rather than a footnote. These are Lütke’s illustrations of a direction, not measurements of platform behavior. He reports that agent-driven transactions remain on a low base and says growth rates at that scale are not meaningful. He offers no counts, no share of GMV, and no causal test separating AI from everything else changing in commerce. What he is describing is an observed shift in the composition of demand — toward considered, higher-value, multi-item projects — not a proven migration.

The commission question nobody has answered

If agents do more of the deciding, the obvious business model is a cut of the transaction. Several agent startups have pitched exactly that.

Lütke’s response was arithmetic. Shopify’s take rate is roughly three percent, and he noted that most of that goes to credit card companies. Whatever is left is not a pool of profit waiting to be divided. When Heath raised the two-percent-of-transaction assumption common among agent founders, Lütke’s answer was that the conversation belongs in basis points — a rounding error on a round of funding, but a real cost on a P&L.

He did not foreclose the model. He listed three possibilities: charging users directly, metering usage the way utilities do, and taking a share of transactions. His test for any of them was unusually concrete — build something better than a cup of coffee, and you will probably endure.

The unresolved part is the one that matters most for brands.

The trust problem hiding inside agent advertising

Asked how advertising survives in an agent world, Lütke described a conflict rather than a format. An assistant that is supposed to work for the buyer but is paid by the seller to steer the purchase is “a breach of trust.”

That is a sharper statement than most platform executives have been willing to make, and it is not a hypothetical. It is the structural condition of every ad-funded assistant currently in market, including the ones selling placements inside AI answers. The buyer’s agent and the seller’s inventory are the same interface.

Lütke still sees room for a new format — something closer to a discovery marketplace than a sponsored slot. He did not specify what it would look like, and no one has built it at scale.

What he did describe is what Shopify is doing meanwhile: building more APIs and transaction infrastructure so agents can call product and checkout capabilities directly. The strategic bet is that Shopify does not need to own the agent. It needs to be the plumbing every agent has to use.

What this means if you sell direct

Four things follow, and they are less about technology than about where you spend.

Your category page is not the first impression anymore. If a household project starts as a sentence rather than a search, the products that appear are the ones an agent can assemble into a working answer. That rewards breadth of catalog, compatibility data, and clear specification over hero imagery and keyword coverage.

Assume the planning conversation is invisible to you. Shopify cannot see it, and neither can your analytics. Your first observable event is an order with unusual basket composition — five items, one project, one delivery window. If you only measure last-click attribution, that order looks unattributable and gets filed under direct.

Watch the basis-point math before you sign an agent deal. A commission on an order that already carries payment processing, fulfillment and marketplace fees is not free money. Model it against your contribution margin, not your revenue.

Treat “the agent works for the user” as a claim to verify, not a default. Lütke named the conflict himself. When an assistant takes money from sellers, its recommendation ordering is no longer neutral, and the burden of proof sits with whoever claims otherwise.

The interesting part of Lütke’s position is that it is not a technology argument. It is an observation that Shopify’s business gets better when commerce happens in more places, and Amazon’s gets better when it happens in fewer. Agents did not create that difference. They just made it visible.