The Protocol War Under AI Shopping

The real fight in agentic commerce is not whether AI will sell. It is who gets to define the rails, the rules, and the leverage beneath the sale.

The Principal

26 min read

Published 11 May 2026

The Protocol War Under AI Shopping

There is a slightly boring version of the AI-commerce story, and then there is the real one.

The boring version says AI shopping is coming. Agents will help people discover products, compare options, and complete checkout. Merchants should get ready. Consumers will enjoy less friction. Conversion goes up. Everyone claps.

Fine. True enough.

But that is not the interesting part.

The interesting part is that a protocol war is breaking out underneath all of this, and that war will decide who owns the leverage in the next era of commerce.

That is what seems to be getting real heat right now across X and the broader operator crowd: not just “AI shopping”, but the standards, connectors and control layers being rushed into place beneath it. Shopify is pushing Universal Commerce Protocol. OpenAI is extending MCP deeper into ChatGPT and workspace connectors. Anthropic is trying to lock in MCP as neutral infrastructure by donating it into the Agentic AI Foundation under the Linux Foundation. Stripe is circling the checkout rail. Google is embedding commerce into AI Mode and Gemini. Microsoft is doing the same in Copilot.

If you strip away the launch language, the argument is simple: if commerce moves into conversations, whoever defines the machine-readable rules of those conversations wins a lot more than traffic.

They win positioning.

They win defaults.

They win integration gravity.

They win data exhaust.

And eventually, they win margin.

This is not a feature race. It is a control race.

Most people still talk about agentic commerce as if it is a UX upgrade.

It is not.

It is infrastructure capture disguised as convenience.

When a human shops on the web, brands spend fortunes fighting for attention. They optimise landing pages, tune paid search, buy affiliates, obsess over CAC, and perform ritual sacrifices to the gods of conversion rate.

When an agent shops, a lot of that theatre collapses.

The agent does not care that your hero section won design awards.

It does not care that your founder wrote a heartfelt manifesto.

It does not care that your PDP has clever microcopy.

It cares whether your product can be discovered, interpreted, compared, trusted, bought, paid for, fulfilled, returned, and supported through machine-readable systems.

That means the old fight for attention becomes a new fight for legibility.

And legibility is always political.

Who decides which product attributes matter? Who determines how discounts are represented? Who gets to express loyalty logic, fulfilment rules, delivery constraints, payment handoffs, identity checks, return conditions, and post-purchase workflows? Who sets the negotiation model between agent and merchant? Who owns the namespace? Who approves extensions? Who benefits when one protocol becomes the easiest default for everybody else?

These are not implementation details. These are the new chokepoints.

Shopify is making the clearest power move

Of all the players, Shopify is being the least coy.

Its line is basically: commerce is messy, we already understand the mess, and if agents are going to buy things, they should do so through a standard designed by people who have survived real retail complexity.

That is the pitch behind UCP, the Universal Commerce Protocol, co-developed with Google.

On paper, it is sensible. Commerce is not one API call. It is a swamp of taxes, discounts, fulfilment options, loyalty logic, stock states, subscriptions, edge cases, market-specific rules, and merchant-specific weirdness. Shopify is correct about that. Anyone who thinks agentic commerce will work through a clean, elegant universal checkout abstraction has not spent enough time near real operators.

Shopify’s argument is that this complexity needs a layered, extensible protocol rather than brittle one-off integrations. That is a strong argument. Frankly, it is the adult argument.

It is also a power grab.

Again: not an evil one. Not even necessarily a bad one. But a power grab all the same.

If merchants become automatically sellable everywhere an AI conversation happens, the platform coordinating that legibility becomes far more important than the storefront itself. The storefront turns into one rendering of the catalogue, not the centre of gravity.

That is a profound shift.

For years, ecommerce software fought to help merchants build better destinations. In agentic commerce, the destination matters less than the transaction grammar.

Shopify sees that early. Credit where it is due.

OpenAI and Anthropic are fighting a different battle

While Shopify is trying to standardise commerce behaviour, OpenAI and Anthropic are trying to standardise tool access and context flow.

This matters because before an agent can buy anything, it has to understand the world it is operating in and safely use tools inside it.

That is where MCP comes in.

Anthropic launched Model Context Protocol as a universal standard for connecting models to external systems. OpenAI, to its credit, did the rational thing and adopted support rather than pretending it needed a bespoke universe. Now ChatGPT workspaces are getting deeper MCP support in beta, including connectors and remote servers. Anthropic, meanwhile, is trying to strengthen the neutrality claim by moving MCP into the Agentic AI Foundation under the Linux Foundation, alongside other founding projects.

This is smart for two reasons.

First, it reduces the fear that one lab will privately control the interface layer for agents.

Second, it acknowledges the obvious: no one serious wants to rebuild tool integrations separately for every model vendor forever. That is wasted motion.

But let’s not romanticise it. “Open” does not mean “powerless”. Open standards often become the most effective method of power concentration because they create dependency through adoption rather than lock-in through force.

If your standard becomes the assumed way agents reach tools, data and actions, you do not need to own every endpoint. You just need everyone to keep speaking your language.

That is softer power. It is still power.

The stack is splitting into layers

One reason this moment feels hot is that the market is beginning to see the layers more clearly.

There is:

The context/tool layer — how agents discover tools, load context, request actions, and handle permissions.
The commerce transaction layer — how agents express intent to buy, negotiate capabilities, build carts, pass identity, handle checkout, and manage post-purchase flows.
The distribution layer — where the customer conversation happens: ChatGPT, Copilot, Gemini, AI Mode, whatever comes next.
The trust layer — who is allowed to act, under what approval model, with what safety constraints, liability, and auditability.

Most commentary still bundles all of this into “AI shopping”. That is lazy.

The real contest is about who dominates which layer, and whether one company can successfully bundle multiple layers together.

Shopify wants weight in the transaction layer.

Google wants distribution plus commerce relevance.

OpenAI wants distribution plus tool/control surface.

Anthropic wants trusted agent plumbing.

Stripe wants payment inevitability wherever agents transact.

Microsoft wants to quietly turn Copilot into a buying interface for enterprise and consumer workflows.

Everybody is smiling and talking about openness because no one can afford to say the quiet part out loud.

The quiet part is this: if agent-mediated purchasing becomes normal, the winners will not necessarily be the brands consumers love most. They will be the infrastructure companies that become impossible to route around.

The big losers may be brands that confuse brand with distribution

This is where it gets uncomfortable.

A lot of consumer brands are still behaving as if “being good on the internet” means being good at human persuasion.

That still matters. But it matters less if the first selector is an agent.

In an agentic world, brand does not disappear. It just gets filtered.

Your brand has to survive translation into structured attributes, comparative reasoning, trust signals, delivery certainty, price/value positioning, and machine-consumable policy clarity. If it cannot survive that translation, it becomes decorative.

That is brutal, but useful.

The winners here may not be the loudest brands. They may be the most operationally legible ones.

Clear catalogue data.

Consistent availability.

Reliable shipping promises.

Unambiguous policies.

Strong product reputation across signals agents can actually use.

Checkout systems that do not break when another machine touches them.

In other words: a lot of “brand moat” is about to get audited by software.

Good.

A great many businesses have been hiding weak operations behind expensive acquisition and glossy creative for years. Agents will be rude about that.

The standards war is not just technical. It is economic.

Here is the bit people underplay: protocol design determines who gets paid, who gets seen, and who gets commoditised.

If one standard makes it trivial for agents to compare ten functionally identical products on price, delivery and return policy, then any premium that depended on shopper confusion starts collapsing.

If another standard preserves richer merchant-specific logic and loyalty structures, maybe incumbents keep more leverage.

If checkout happens natively inside the AI interface, the distribution layer grows stronger.

If the final step kicks out to the merchant, the merchant keeps more control.

If payments become abstracted behind agent-facing credentials, payment providers with the safest and easiest agent primitives gain share.

Every protocol choice contains an economic philosophy whether its authors admit it or not.

That is why this is heating up now. The market is realising the standards are being drafted before behaviour is settled.

Once that happens, you do not merely document the market. You shape it.

My contrarian take: “open” will win, but not in the way people think

The fashionable position is to say open standards will obviously win because no one wants vendor lock-in.

I think that is half true and half nonsense.

Open standards usually win at the infrastructure layer when enough heavyweights decide duplicated effort is stupid. That appears to be happening with MCP-style tooling and probably with commerce protocols too.

But “open wins” does not mean the market becomes egalitarian.

It usually means a handful of companies become the best at implementing, extending, hosting, securing, governing, and commercially exploiting the open layer.

Linux was open. That did not eliminate power. It redistributed where power sat.

The same thing is likely here.

The winners will be the firms that combine three things:

enough scale to influence the standard,
enough credibility to get others to adopt it,
and enough product surface to monetise the standard once it spreads.

That is why Shopify, OpenAI, Anthropic, Google, Microsoft and Stripe all matter in this conversation.

Not because they are the only clever people in the room. Because they are the ones positioned to turn a protocol into gravity.

So what should operators actually do?

First, stop treating this as distant future theatre. It is moving from speculative to operational.

Second, stop asking whether agentic commerce is “real”. That is now the wrong question. The right questions are:

Which interfaces are your products becoming legible in?
What structured data do agents actually see?
Can your checkout, policy and fulfilment logic survive machine mediation?
Which protocols are emerging in your ecosystem?
Where are you accidentally becoming dependent on someone else’s default?

Third, understand that this is not just an ecommerce issue.

The same logic is coming for software buying, B2B procurement, travel, financial services, healthcare admin, and any category where decision-making can be partially delegated to software.

Every market that looks “search-like” today eventually becomes “agent-like” tomorrow.

And when that happens, the control points move down the stack.

Why this now

Because the conversation has shifted from “agents are interesting” to “which standard gets embedded first”. In the last few hours, the highest-signal chatter and supporting announcements have converged on the same underlying story: AI buying is no longer just a demo problem. The real contest is over protocols, connectors, checkout rails and trust boundaries.

That is the story under the story.

And if you miss it, you will spend the next two years optimising the shop window while someone else quietly buys the land underneath it.