Google, OpenAI, Visa & Mastercard Are Racing to Let AI Agents Shop for You — Only 3% of Purchases Actually Do

Five major companies have shipped competing protocols to let AI agents shop and pay on your behalf, all within the last twelve months. The infrastructure race is real and well-funded. The consumer adoption behind it isn't — yet. Here's the data on both sides, and what it actually means if you're not Amazon or Etsy.

In the last twelve months, Google, OpenAI, Visa, Mastercard, and PayPal have all shipped a protocol that lets an AI agent hold your payment details and complete a purchase without you clicking “buy.” That’s five of the largest companies in tech and finance, racing each other to build the rails for a future where your assistant does the shopping.

Here’s the number that doesn’t fit the hype: only 3% of transactions today actually involve an AI agent.

What Is Agentic Commerce, and Why Is Everyone Building for It Right Now?

Agentic commerce means an AI agent — not a human clicking through a checkout flow — completes some or all of a purchase on a person’s behalf: finding the product, comparing options, and paying for it within a set of rules the person approved in advance.

The reason five major companies shipped competing protocols inside the same twelve-month window isn’t coincidence. Google announced its Agent Payments Protocol (AP2) on September 16, 2025 (Google Cloud Blog). Thirteen days later, OpenAI launched Instant Checkout and the Agentic Commerce Protocol (ACP) inside ChatGPT, co-built with Stripe (OpenAI). Visa introduced its Trusted Agent Protocol the same year and has been pushing toward what it calls “mainstream adoption in 2026” (Visa). Mastercard followed in June 2026 with Agent Pay for Machines (Mastercard). Nobody wants to be the payment network or platform that AI agents can’t transact on when this actually takes off — the same pressure that’s pushing Google to start scoring sites on how agent-ready they are in the first place.

The Protocol Race: Who’s Building What

Each protocol solves a slightly different piece of the same problem — how does a merchant, a bank, or a card network know that an AI agent is actually authorized to spend a real person’s money?

The most technically interesting piece is how Google’s AP2 actually proves authorization. It uses cryptographically-signed “Mandates”: an Intent Mandate captures what you asked the agent to do (say, “find me white running shoes under $80”), and a Cart Mandate is generated and requires your approval once the agent has an exact cart — items, prices, and all — creating a non-repudiable record of what you actually agreed to buy (Google Cloud Blog). For purchases you pre-authorize without being present — a recurring reorder, a price-drop buy — the agent generates the Cart Mandate itself once your conditions are met.

PayPal’s own read on the landscape is worth noting because it’s coming from a company with no protocol to sell: it groups the field into commerce protocols (ACP, Google’s UCP), payment/trust protocols (AP2, Visa Trusted Agent Protocol, Mastercard Agent Pay), and infrastructure protocols (Google’s A2A, Anthropic’s MCP) — and argues merchants shouldn’t bet on one winner, because “support for multiple protocols will be needed to enable commerce experiences with the leading platforms” (PayPal Newsroom).

The Market Is Moving Fast — On Paper

The growth numbers explain the urgency. The global agentic commerce market was worth $5.7 billion in 2025, is estimated at $7.7 billion for 2026, and is forecast to grow at a 35.7% CAGR through 2033 — reaching $65.5 billion (Grand View Research). North America currently holds 38.2% of global revenue share, while Asia Pacific is growing fastest. Retail and e-commerce dominate the use cases so far.

But Consumers Aren’t Racing With Them

Here’s where the infrastructure race runs ahead of actual behavior. Checkout.com’s June 2026 survey across six markets, including the US and UK, found that a third of consumers (33%) expect at least 10% of their purchases to be AI-driven within a year — but almost a quarter (24%) say they will never let AI handle a purchase, and 27% say they trust no organization to operate an AI shopping agent on their behalf (Checkout.com). Put plainly: right now, only about 3% of transactions actually involve an AI agent.

The same survey found consumers aren’t rejecting the idea outright — they’re asking for guardrails before they’ll use it. The top three requirements were a spending cap (30%), the ability to instantly revoke an agent’s access (29%), and easy cancellation (28%). On average, people said they’d pre-authorize an agent to spend around £177 without needing to approve each purchase individually — useful context for how small the “unsupervised” transaction ceiling actually is today.

This tracks with what I found digging into how AI chatbots are already influencing purchase decisions earlier this month: consumers are far more comfortable letting AI talk them out of a purchase than letting it complete one automatically. Trust in an agent’s judgment and trust in an agent’s spending authority are two very different thresholds, and right now most people have only crossed the first one.

So Is This Hype, or Is It Real?

Both, depending on which half of the picture you’re looking at. The infrastructure is real, well-funded, and moving fast — five major companies don’t build competing payment protocols for a market that doesn’t matter. But the consumer side is still an early-adopter phase, not a mainstream shift. A 3% current-usage number next to a 35.7% CAGR forecast means most of that growth is still ahead of us, not behind us.

What This Actually Means If You’re Not Amazon or Etsy

Most B2B and SaaS marketers reading this aren’t going to build a native “buy with ChatGPT” button this quarter, and at 3% current adoption, that’s the right call — the volume isn’t there yet to justify the engineering effort for most businesses. But there’s a more immediate, lower-cost piece of this that’s worth doing now:

Agents are already reading your site to answer questions and make comparisons, even before they’re completing checkouts on it. That means your pricing pages, product specs, and structured data need to be legible to an agent today, independent of whether that agent can actually complete a purchase yet. The same entity signals and structured data that help you show up in AI Overviews and chatbot answers are the foundation any future checkout integration would sit on top of anyway — it’s the same groundwork I covered in the Agentic Web Readiness Checklist, and it overlaps heavily with getting your Search Profile Badge and Preferred Sources status right — the same verified-entity signals that help Google trust you are what a payment protocol’s credential system will eventually check too.

My Take

The protocol race tells you where the big platforms think commerce is headed, not how fast your customers are actually getting there. I’d treat 2026 as the year to get the groundwork right — clean structured data, accurate and agent-readable product and pricing information, clear entity signals — rather than the year to build a bespoke AI checkout flow for a use case that’s still 3% of transactions. The companies racing to build these protocols are playing a longer game than most of us need to play right now.

What I’d Actually Check on My Own Site

If you want to be ready without over-building, here’s where I’d start:

• Confirm your product and pricing data is in structured, machine-readable format (Schema.org Product/Offer markup), not buried in images or JavaScript-rendered text an agent can’t parse.
• Read up on WebMCP if you’re technical — it’s the emerging standard for letting agents take structured actions on your site, not just read it.
• Keep an eye on your referral mix as this shifts — my breakdown of how much traffic actually comes from AI sources today is worth revisiting once agent-driven purchases start showing up in your analytics.
• Don’t rush to integrate a specific checkout protocol yet — with five competing standards and no clear winner, PayPal’s “stay protocol-agnostic” advice is sound for most businesses.

Written by mootkur9

17 years in digital marketing. Founder of Sadhyam. Writing about AI search, AEO/GEO, and the agentic web. More about me →