For years "the agentic web" was a slide in a conference deck. In 2026 it is in your analytics, even if you cannot see it labelled as such. AI agents are now researching products, comparing options and, increasingly, completing checkout on a person's behalf. The shopper describes what they want, the agent does the legwork, and the human only approves at the end.

The blunt reality for most sites is this: agents already reach your cart, and most of them stall or get blocked before they buy. You are losing a customer you never even saw arrive.

This is real now, not a forecast

A quick, accurate map of what shipped:

  • OpenAI launched Operator in January 2025, then folded it into ChatGPT Agent later that year. It can browse, add to cart and prepare checkout.
  • The Agentic Commerce Protocol, from OpenAI and Stripe, powers Instant Checkout inside ChatGPT for participating merchants.
  • Google's Project Mariner brought a Gemini-powered browsing agent, with agentic features feeding into Search, and the Agent Payments Protocol (AP2) arrived with 60-plus partners.
  • Visa and Mastercard both announced agentic-payment frameworks using tokenised credentials for verified agents.
  • Amazon's "Buy for Me" lets its agent purchase from third-party brand sites, and Perplexity added shopping with a merchant programme.

One clarification, because the acronyms get muddled: MCP is a protocol for connecting models to tools and context. It is not a payment rail. The payment and checkout standards are the ACP and AP2 family. Treat any specific protocol name with care, because new ones are appearing fast.

What "agent-ready" actually means

You do not need to integrate a payment protocol to start. You need a site a competent agent can actually traverse. In our experience that comes down to a handful of things:

  • Machine-readable structured data on products: Schema.org Product and Offer with price, availability and variants.
  • Clean semantic HTML with meaningful landmarks and labels, so the page is legible without guesswork.
  • Stable, predictable selectors. Agents break when your DOM is reshuffled or obfuscated every release.
  • Feeds and APIs so an agent can read inventory without scraping.
  • An accessible, linear checkout: clear required fields, programmatic labels, fewer surprise popups.
  • Do not block legitimate agents. Some networks now block AI bots by default; if you want agent buyers, you have to let them in deliberately.
  • Fast pages with content in the HTML, not buried behind heavy client-side rendering.
  • Agent-readable docs (an llms.txt or similar map), which a growing list of technical companies have adopted.

Why most sites fail today

The irony is that the same defences that stop bad bots also stop good agents. Behavioural fingerprinting flags an agent before any visible challenge appears. Checkout is a multi-step JavaScript flow with coupon popups and surprise interstitials. Prices and stock live behind dynamic rendering. Selectors change weekly. A human pushes through all of it on instinct. An agent simply gives up.

The first move is not adopting a payment protocol. It is making sure your existing site does not fail a competent agent today.

Where to start

Our Agentic AI Readiness Audit runs real agent flows against your key journeys, from product discovery to checkout, and reports exactly where an agent gets blocked, confused or stalled, alongside a prioritised roadmap to fix it. For high-consideration purchases especially, being the site an agent can actually complete is turning into a quiet first-mover advantage.