Almost every conversion technique developed in the last twenty years acts on human perception. Headline framing, social proof placement, visual hierarchy, urgency cues, trust badges, friction reduction in checkout — all of it assumes a person is looking at a rendered page and forming an impression.
Agentic commerce removes that assumption. When a user delegates a purchase to an agent, the agent queries structured data, evaluates it against stated constraints, and transacts. Your page may never render at all.
This is early, and the honest position is that most merchants should not restructure their business around it yet. But some of the preparation is cheap, and nearly all of it pays off whether or not agentic checkout becomes a major channel.
What an Agent Actually Evaluates
Strip out everything that requires a human eye and this is what remains:
| Human buyer weighs | Agent weighs |
|---|---|
| Brand impression, design quality | Nothing — not rendered |
| Persuasive copy, urgency | Nothing — not parsed as persuasion |
| Reviews as social proof | Review score as a numeric field |
| "Free shipping" in a banner | Shipping cost and threshold as structured values |
| Perceived trustworthiness | Return window, warranty terms, stated policy |
| Checkout flow smoothness | Whether a checkout API exists at all |
The pattern: anything expressed only in prose or pixels is invisible; anything expressed as a field is decisive. A merchant whose return policy is a beautifully written page and a competitor whose return policy is a machine-readable value are not competing on equal terms in an agentic comparison, regardless of which policy is better.
The conversion levers that remain are unglamorous: price, availability, specification completeness, shipping terms, return terms, and whether the agent can transact without a human handoff.
The Protocol Landscape
Four standards matter, and they stack rather than compete.
ACP (Agentic Commerce Protocol) — OpenAI with Stripe. Merchants publish a structured product feed and implement the Agentic Checkout Spec so ChatGPT can complete purchases directly.
UCP (Universal Commerce Protocol) — Google released it as an open standard on 11 January 2026, developed with Shopify, Etsy, Wayfair, Target, and Walmart, endorsed by 20+ partners. Its stated purpose is removing the N×N integration bottleneck: rather than building a custom connection per shopping surface, a merchant integrates once. It supports integration via APIs, Agent2Agent, and MCP.
AP2 (Agent Payments Protocol) — Google's authorization layer, announced September 2025 with 60+ partners including Mastercard, PayPal, Coinbase, and American Express. It answers the liability question: three signed mandates — Intent, Cart, and Payment — carried as W3C Verifiable Credentials, establishing cryptographically that a specific user authorized a specific purchase at a specific amount.
MCP (Model Context Protocol) — the general standard for exposing tools and data to models. UCP supports it as an integration path.
Roughly: ACP and UCP handle the commerce journey, AP2 handles proof of authorization, MCP handles the plumbing.
What Is Worth Doing Now
Ordered by ratio of benefit to cost.
Get your structured product data right. Complete, accurate, current product markup — price, availability, GTIN, specifications, shipping, returns — is the foundation for every agentic surface, and it independently improves traditional search visibility and AEO. This is the highest-return item on the list and it would be worth doing if agentic commerce never materialized.
Make policies machine-readable. Return windows, warranty terms, shipping thresholds, and stock status should exist as structured values, not only as prose on a policy page.
Audit for ambiguity. Agents do not infer. A price that requires selecting three variants to reveal, stock status shown only as a color, or shipping cost disclosed at step four of checkout are all data an agent cannot resolve. It will pick the competitor whose data it could read.
Log and identify agent traffic. You cannot reason about a channel you cannot see. Server-side logging with signature-based identification tells you whether this is 0.01% of your traffic or 3%.
Watch your platform. If you are on Shopify or another platform that participated in UCP's development, most protocol support will arrive as a platform feature. Waiting is often the correct move.
Do not rebuild your storefront, restructure your catalog, or divert optimization budget from human conversion for this. The volume does not justify it yet.
The Measurement Trap
A specific warning, because this one is already causing bad decisions.
Agent sessions that browse but do not transact behave like agentic browser traffic generally: they enter your analytics and your experiments as ordinary sessions and almost never convert. That dilutes conversion rates and distorts A/B tests — covered in full in are AI agents breaking your A/B tests.
Meanwhile, agent-completed purchases may arrive through an API path that your front-end analytics never observes at all. So the same trend can simultaneously depress your measured conversion rate and hide the revenue that offsets it. If you are seeing unexplained conversion-rate decline alongside flat revenue, this is worth ruling out.
How This Changes Optimization Work
The useful way to think about it: you are increasingly optimizing for two audiences with nothing in common.
For humans, everything in the existing CRO playbook still applies, and will for a long time — this is where essentially all revenue still is.
For agents, the work is closer to data engineering than to design. Completeness, accuracy, and machine-readability replace persuasion. There is no multivariate test to run on a JSON feed; there is only whether the field is present and correct.
Those are different disciplines, and the mistake to avoid is letting the second one — currently a rounding error in revenue — pull resources from the first.
The Bottom Line
Agentic commerce is real, standardizing quickly, and still small. The protocols consolidated faster than most people expected, which reduces the risk of betting on the wrong one, and much of the merchant-side work will arrive through platforms rather than custom builds.
The rational posture is preparation, not transformation. Clean structured data, unambiguous machine-readable terms, and visibility into agent traffic are all things that pay for themselves through conventional search and AI citation regardless of what happens next. Everything beyond that can wait for the volume to justify it.
For the human side of the funnel — still where the revenue is — Surface AI runs continuous autonomous experiments so that work keeps compounding while you prepare for the rest.