Everyone's debating which AI agent is the best shopper. Wrong question.
The question that actually matters: who decides which agents get to shop where?
Right now, agentic commerce is splitting into two worlds. One where agents can access product data, compare prices, and complete purchases freely. And one where a single platform controls the door. The first world is growing fast. The second world is Amazon.
I'd call this the Trust Fork. And it's going to shape how commerce works for the next decade.
Here's the thing nobody in the "consumer authorization" conversation is saying clearly enough: it doesn't matter if you give your AI agent full permission to buy on your behalf. Your agent still can't buy on Amazon unless Amazon lets it in.
Amazon updated its Business Solutions Agreement in March with a new Agent Policy. Every AI tool operating on Amazon must identify itself in HTTP requests, can't mimic human browsing patterns, must respond truthfully when asked if it's a bot, and can be blocked at Amazon's discretion. Amazon is currently suing Perplexity for operating its Comet agent on Amazon's marketplace. The message is clear: Amazon decides which agents are welcome, regardless of what the consumer authorized.
Consumer consent doesn't override platform terms of service. Your bank can authorize your agent to spend $500 on household goods. Amazon can still block that agent at the door. The platform is the bouncer. The bank is just the wallet.
This is a power that only Amazon has at this scale. Google doesn't have the same gatekeeper problem because Google doesn't own the checkout — it sends you to the merchant's site. Walmart can't really block agents without losing volume it needs. Shopify's merchants actively want agents sending traffic. Amazon can block agents because it has enough demand-side gravity that consumers will come anyway.
On the other side of the fork, everything is opening up.
Visa and Santander just completed the first end-to-end agentic payment pilot across five countries in Latin America. Agents bought products with consumer-authorized Visa tokens. J.P. Morgan committed payments infrastructure to agent-mediated transactions this month. Shopify reported AI-driven orders grew 15x year over year — and Shopify didn't build a shopping agent to get there. They made their merchants' catalogs readable by other people's agents. Millions of stores, structured for AI discovery, accessible through ChatGPT and Google and Perplexity.
OpenAI tried to own the full transaction inside ChatGPT last year. Walmart's data showed it converted three times worse than Walmart.com. Only 30 Shopify merchants went live in six months. OpenAI killed the feature and pivoted to retailer-owned apps inside ChatGPT. Walmart built Sparky, their own assistant, running at 70% of Walmart.com's direct conversion rate.
I run growth at an Amazon agency — Envision Horizons — and we track conversion across dozens of brands. The lesson from the OpenAI failure isn't that agents can't drive commerce. It's that agents are great at the front half — discovery, research, comparison — but the moment money changes hands, consumers want to be somewhere they recognize. The agent gets credit for the assist, not the close. That's the handoff model, and outside Amazon, it's winning.
So you have two worlds forming.
In Amazon's world, there's one agent: Rufus. Amazon controls what it recommends, how it buys, and which third-party tools can touch the marketplace. If you sell on Amazon, your agent strategy is Rufus. There's no other option. Amazon has the purchase history, the logistics network, the Prime relationship, and the legal infrastructure to keep it that way.
In the open world, agents compete. Perplexity's Comet, ChatGPT with Shopify integration, Google's agent checkout, whatever comes next — they all access merchant data through open protocols (UCP, ACP) or direct integrations. Merchants choose to participate. Consumers choose their agent. The payment layer is getting built by Visa, J.P. Morgan, and others. Nobody owns the whole stack, which means nobody can lock the door.
Constructor's numbers tell you how fast the open world is moving: 82% customer growth, 322 billion discovery interactions, and a 52% lift in add-to-cart rate from their PDP agent. Structured product data that an agent can actually read turns browsers into buyers at dramatically higher rates.
The fork matters because it changes what "agentic readiness" means depending on where you sell.
On Amazon, readiness means optimizing for Rufus. Structured listings, strong review signals, and accepting that Amazon controls the agent layer. You don't get to choose which agent represents your products. Amazon already chose for you.
Off Amazon, readiness means making your product data accessible to any agent. Structured attributes, clean pricing, accurate inventory, clear use-case descriptions. If an AI agent can't parse your catalog, you don't exist in the open agent ecosystem. Constructor, Lemrock (which just raised €6M building middleware that connects brand catalogs to LLMs) — they exist because brands have rich product knowledge that agents can't currently access.
The brands that treat these as one problem will get it wrong. Amazon and the open web are diverging, and the gap is going to widen. What works for Rufus won't necessarily work for Perplexity or ChatGPT, and vice versa. Different data formats, different optimization signals, different rules about what an agent can and can't do.
There's a version of the future where this fork doesn't matter — where Amazon eventually opens up and lets third-party agents operate on its marketplace. Andy Jassy said at Davos that Amazon would allow external agents "as long as there is an appropriate value exchange." But Amazon's actions since then have gone the other direction. The BSA update, the Perplexity lawsuit, the investment in Rufus as the sole Amazon shopping agent — everything points toward a closed system, at least for now.
The more interesting question is whether the open side can pull enough commerce away from Amazon to make the gatekeeper position less powerful over time. Shopify's 15x growth is early evidence that it can. But Amazon still controls roughly 40% of US e-commerce. That's a lot of gravity.
When I wrote Instant Checkout, I framed the trust question as a spectrum — how much autonomy should consumers give their agents? I was thinking about the wrong axis. The real trust question isn't between consumers and their agents. It's between platforms and agents. Amazon doesn't trust any agent it doesn't control. Everyone else is building systems that assume agents should be trusted by default, with guardrails at the payment layer.
That's the fork. Not consumer-auth vs platform-auth. Closed vs open. And the only company with enough power to keep the door shut is Amazon.
The consumer experience of AI shopping is still awkward. I use it every day and it still takes effort. But the plumbing underneath is getting built fast, by players with serious capital. J.P. Morgan doesn't enter an infrastructure bet unless they believe the volume is coming. Visa doesn't run a five-country pilot for a press release.
The checkout you don't consciously experience is still the future. What I didn't account for when I wrote the book is that the biggest obstacle isn't technology or consumer trust. It's one company deciding that its 40% of e-commerce stays behind a wall.
For brands, the move is straightforward: build for both worlds, because both are real. But stop pretending they're the same world. They aren't, and the gap is growing.