Your Shopping Agent Is Working For Someone Else

Ask Rufus whether the blender you’re looking at is cheaper on Walmart.com.

I’ll wait.

Amazon built Rufus. It runs on Amazon’s inventory, Amazon’s pricing data, Amazon’s margins, Amazon’s Prime conversion goals. Rufus is a genuinely useful assistant — for shopping on Amazon. It’s not trying to find you the best blender in the world. It’s trying to find you the best blender Amazon sells.

That distinction is going to matter a lot.

ChatGPT did a pivot last month that I think got read wrong. OpenAI pulled back from native checkout and moved toward merchant apps — Target, DoorDash, Instacart. The coverage framed it as a retreat. I think it’s a clarification. ChatGPT is telling you what it actually is: a surface where brands build agents that sell their products. Nike’s app inside ChatGPT is not going to recommend Hoka. That’s not a bug, it’s what the product is.

Google UCP is dressed up as an open standard but it’s still an ad product. There’s a “Direct Offers” feature baked into it where brands pay to surface discounts to buyers. The agent layer is just the new ad unit.

So we have: Amazon’s agent. Each brand’s agent. Google’s agent. None of them are your agent.

The thing nobody has built yet is a buyer’s agent. One that holds your preferences, knows your purchase history, has your payment credentials, and shops the whole internet for the best answer for you. Not the best answer for the platform’s margins. Not the best answer for the brand’s conversion rate. The best answer for you.

Why hasn’t it been built?

Part of the answer is incentives. Building an honest buyer’s agent means you make money on the transaction without controlling the inventory or the brand relationship. You’d need to be indifferent to which merchant wins, as long as your user gets a good outcome. Very few companies are structurally positioned to do that.

The ones that actually are: Mastercard and Visa.

This is not an obvious answer, so stay with me. The payment networks make money on every transaction regardless of merchant. They have no inventory to push, no brand deals to protect, no Prime membership to convert you into. They’re the one layer in commerce with no inherent preference for who wins the sale — just that a sale happens.

Mastercard dropped an open-source authorization standard called Verifiable Intent on March 5. It’s cryptographic proof that an agent is authorized to act on your behalf — not the platform’s behalf, yours. That’s the underlying architecture problem a buyer’s agent has to solve. And a payment network just solved it, for free, for anyone who wants to build on it.

Visa has been moving the same direction. Visa Intelligent Commerce has 100+ ecosystem partners and 30+ actively building in their sandbox. Both networks are laying payment rails that are explicitly agent-native.

But here’s what neither has done: ship a consumer-facing product.

Apple is the other name that keeps coming up. Apple Pay works everywhere. Apple has your identity, your biometrics, your device trust. An Apple Shopping agent would have no conflicted interests — Apple doesn’t sell TVs or dog food. Reportedly exploring. Nothing shipped.

And there are probably five startups building this right now. None have broken through.

So we’re sitting in an interesting moment. The infrastructure for a real buyer’s agent exists — the payment rails, the authorization standards, the AI capability (Google’s Project Mariner runs 10 parallel web tasks at 83.5% accuracy; that’s the spec for a shopping agent that actually compares). But the product doesn’t exist yet.

The Google parallel is the useful frame here. Search looked neutral. It got gamed. Google Shopping looked neutral. It became an ad product. The agent layer is going to be the biggest monetization surface in commerce history, and every platform knows it. They’re all racing to establish their agent as the default before a genuinely unaffiliated one can.

76% of consumers say they want AI help shopping. Only 30% are willing to let an AI complete the payment. That gap isn’t a trust problem with AI in general. It’s a trust problem with whose AI it is. People can feel that the agent recommending a product might have a reason to recommend it. They’re right.

What should you actually do with this if you’re running a brand?

The current agents — Rufus, the ChatGPT merchant apps, Google UCP — are your customers right now. You optimize for them the way you used to optimize for Google’s search algorithm. Clean product data, accurate inventory, GTINs that work. That’s the immediate project.

But keep one eye on who builds the buyer’s agent, because when it arrives, the game changes completely. A buyer’s agent with no agenda can’t be gamed the way a platform algorithm can. You either make the best product for a given need or you don’t. The whole edifice of keyword strategy and listing optimization becomes a lot less important.

That agent is probably 2–3 years out at real scale. Mastercard and Visa are laying the rails. Apple has the distribution. Some startup nobody’s heard of yet might ship it first.

The question is whether your brand is the right answer when an agent with no conflicts goes looking.