Agentic Commerce Is Bigger Than the Checkout

On September 3, 2026, I stood on the Meet Magento Ukraine stage, speaking for DataArt, with a slide that had five words on it: mobile, voice, omnichannel, internet of things, social. Five waves of e-Commerce, each one announced as the thing that would replace the traditional store.

None of them did.

Mobile made mobile-first a default instead of a feature we billed for. Voice pushed us toward structured content. Omnichannel forced one stock, one customer profile, and APIs between systems. The internet of things normalized multi-channel and hyper-personalization, and social media made shoppable content table stakes.

Each wave quietly rewrote what a "normal" store must have and then became invisible. Nobody puts "mobile-first" on an invoice for development or design work anymore.

Agentic commerce is doing it again, with one twist that changes the whole brief: the new user isn't a human on a new device. It's a machine acting on a human's behalf, and it can only buy what it can read.

Every Part of the Store Now Has to Be Legible to Machines

That is the core difference between agentic commerce and traditional eCommerce. If the shopper's first interface becomes an AI agent, the store's job shifts from persuading a human to being legible to a machine. Every part of the store must pass that test, from discovery to retention.

Not because the checkout stopped mattering. Because the agent shows up long before it.

The shopper has already moved. One in four customers now use AI-powered platforms as their primary source for research, purchase decisions, and recommendations, ahead of brand websites and online reviews, according to Adobe's 2026 AI and Digital Trends report, which surveyed 4,000 customers and 3,000 executives in late 2025. On the other side of the counter, only 13% of organizations have embedded agentic AI for brand discovery and search. That gap is the whole opportunity.

What Is Agentic Commerce? Start With the Agent

Agentic commerce is commerce in which AI agents search, compare, configure, and increasingly buy on a shopper’s behalf, which makes the agent, not the human, the first reader of a store’s products, policies, and data.

Think of the agent as a personal shopper who has never set foot in your store and reads only what is written down. Its answer is the new storefront. Its shortlist is the new shelf. Most people hear "agentic commerce" and picture the checkout: a bot paying with someone else's card.

True, that is the endgame. But it is the smallest part of the story, and the part that moves slowest.

Two Kinds of AI Agents in eCommerce

Not all agents read the same sources, so where your product data lives decides which agents can find you. Broadly, there are two kinds: agents that read the open web, and agents that read only a retailer's own catalog.

General-purpose agents sit outside any store: ChatGPT, Google’s AI Mode and Gemini, Microsoft Copilot, Perplexity, Claude. They know nothing about you until they read you; every merchant competes on the same terms, and the currency is the public, machine-readable web plus what others write about you.

Shopping-specific agents belong to a retailer and work inside its catalog. Walmart launched Sparky in June 2025; by the quarter ending April 2026, units bought through Sparky had more than quadrupled quarter over quarter, weekly active users were up over 100%, and Sparky users' average order value was about 35% higher than other shoppers. In May 2026, Amazon retired Rufus and replaced it with Alexa for Shopping, which can now buy on shoppers’ behalf from other retailers. If you sell through those marketplaces, your listing is the only thing these agents ever read about you. There is no PR to lean on.

The two layers are already merging. In March 2026 Walmart brought its shopping experience into ChatGPT: discovery starts in the general-purpose agent, then hands over to Walmart’s own account, loyalty, and checkout. “Shopping can start anywhere now,” said Daniel Danker, Walmart’s EVP of AI acceleration, product and design, “whether that’s Walmart or a question in ChatGPT.”

Either way, one enriched catalog must serve both shelves. But no agent can read a catalog it can't reach, and that starts with who you let through the door.

Sidebar: Three Bots at the Door

Three kinds of bot now visit a store, and they want different things:

• Training crawlers: read your pages to teach a model what your category looks like
• Retrieval indexers: keep an answer engine's memory of your prices and policies fresh
• Live agent fetches: act for one person right now: reading a product detail page (PDP), a delivery policy, a stock figure before answering them

Block the first, and you disappear from the model picture of the world. Block the third and you disappear from a sale that is happening at this minute. Most CDN and WAF defaults were tuned before this distinction existed, so it's very much worth reviewing and reconfiguring your web infrastructure to optimize Agentic Commerce.

Discovery Won. Checkout Stalled.

Google announced the Universal Commerce Protocol (UCP) at NRF in January 2026, co-developed with Shopify and backed by Etsy, Wayfair, Target, and Walmart, letting shoppers check out inside AI Mode and Gemini with Google Pay. Google is explicit that merchants remain the merchant of record. OpenAI and Stripe had shipped the Agentic Commerce Protocol (ACP) and Instant Checkout in ChatGPT in September 2025; Google's Agent Payments Protocol (AP2), a standard for agentic payments, arrived the same month with more than 60 payments and technology partners. Shopify, PayPal, and Stripe now publish official Model Context Protocol (MCP) servers—standard sockets an agent plugs into.

Then, in March 2026, OpenAI sidelined Instant Checkout. Shoppers asked ChatGPT plenty of product questions but rarely bought inside it, and merchants struggled to keep prices and stock in sync through an intermediary. Checkout moved back to merchants' own apps and stores. The same month, Shopify switched on Agentic Storefronts, and every eligible store's catalog became discoverable inside ChatGPT with no integration work by the merchant.

Read those two events together. The agent as a buyer is early and unsettled. The agent as the first filter is already here, at scale, and it decides which stores humans ever see.

So the sale is increasingly won or lost before your site loads—and the loss never shows in your analytics, because there was no visit to lose.

Here's the good news. Everything the agent needs to find you, understand you, and eventually buy from you is the same groundwork, whichever protocol wins.

Agentic Commerce Is a Legibility Project

My take: Agentic commerce is not a checkout project. It's a legibility project to be readable by machines, be operable through APIs, and the checkout is the last mile of it.

That's the whole brief. It has four acts, and most merchants have only thought of the third.

Four Acts, One Foundation

1) Discovery

The agent has to find you. Nothing you say about yourself matters if the model can't read it. And the model quotes what others say about you more than what you say.

Operationally, that is three checks and one program. Check that robots.txt and your CDN or WAF are not blocking AI crawlers. Check that your prices are public and current; an agent won't recommend what it can't see. Check that delivery, returns, and warranty policies exist as plain, machine-readable pages, not PDFs behind a login.

Then start the program. GEO (generative engine optimization) and AEO (answering engines optimization) are the new reincarnations of SEO. Best practices here really depend on the platform: Google Search, Gemini, ChatGPT, Perplexity, Amazon, Walmart and others — all have their tools and playbooks. Research your case to define the strategy

Shopify's March 2026 switch-on is the example to study. Stores were syndicated into ChatGPT automatically, but only those meeting basic product data standards.

The same gate applies to marketplaces. Alexa for Shopping and Sparky read your listing (attributes, reviews, Q&A) not your website, so the enrichment you do for the open web is the enrichment that gets you recommended inside Amazon and Walmart.

Measure it the way you would measure a shelf: ask ChatGPT, Claude, Gemini, and Perplexity 10 real questions from your category every month and count how often you appear, and how accurately. Palmata by Contentful, launched in June 2026, turns that into a standing report and starts with a free one.

2) Configuration

Now the agent is inside your store. Your catalog structure and on-site search are its main interface. Treat them like one.

An agent configuring a purchase behaves like a very literal user of your site. Where a human would squint at a lifestyle photo, the agent reads the alt text or reads nothing.

Semantic markup, complete attributes, and accessibility stop being a compliance line and become the agent's user interface. What helps a screen reader helps an agent.

The measures are already in your analytics: the zero-result rate of on-site search, the share of PDPs with every mandatory attribute filled, and a monthly run of a top product page through Google's Rich Results Test.

3) Checkout

The agent completes the purchase. In agentic checkout, it buys through a protocol and pays with a token, and you stay the merchant of record.

UCP, ACP, and AP2 each define a piece of how an agent buys—the cart, the payment mandate, the order. Know which one your stack can speak today, and don't bet the roadmap on any of them; OpenAI's March pivot is the reminder that the surface can change under you in a quarter. What does not change is the requirement underneath: stable, documented APIs, sub-second stock and price confirmation, tokenized payments so no raw card data ever reaches an agent, and order status and returns exposed through the same API the agent used to buy.

Target's ChatGPT app, live for the 2025 holiday season, is the pattern that survived the pivot: the retailer owns the checkout, the agent owns the conversation.

Measure the channel like any other: agent-originated orders by surface, agent-to-checkout conversion per surface, and—the one your CTO will care about—time to integrate a new surface. If that number is months, the foundation isn't there yet.

4) Retention

You're back with a human. The agent may have placed the order, but a person unpacks the parcel.

The first sale is getting impersonal, so the game shifts to retention, where you hold most of the data and control. Adobe's report puts numbers on both sides: 70% of organizations expect agentic AI to handle at least half of post-purchase support within 18 months, while the disclosure consumers value most is the ability to switch to a human at any time. Design the handoff before you automate the service.

Operationally: personalize onboarding, service, and communication for the person, not for the agent who bought it. Expose order tracking and returns to the agent surface and make the route to a human one tap long.

Measure the second-purchase rate of agent-originated customers against your baseline, and time-to-human when a customer asks for one.

Most headline examples come from consumer retail, but agents are reaching B2B too, with three complications.

• Pricing is rarely public: it depends on contracts, volumes, and accounts, so “make prices visible” becomes “make pricing logic queryable.” Publish list prices or ranges in the open, and serve account-specific prices through an authenticated API an agent can call on the buyer’s behalf
• Products are configured, not picked: compatibility rules, units of measure, and lead times need to live in data, not in a sales rep’s head
• The buyer’s agent sits inside procurement: it works from ERP and purchasing tools, not a chat window, so quotes, approvals, and purchase orders need the same documented APIs as checkout

None of these changes into the four acts. It raises the bar for the foundation under them.

Under All Four: One Foundation

Structured, connected, governed product data and documented APIs sit under every act. They are protocol-independent, so waiting for the standards to settle buys you nothing. And the advantage of compounds: AI learns slow-moving signals, so it's more likely to recommend the same brands tomorrow. We see the same pattern when companies move AI agents from pilot to production: an agent is only as good as the data it can read.

Only 44% of organizations say their data quality and accessibility are adequate for AI at all, per the same Adobe study. Catching up is measured in months, not sprints.

Agents Are Already the First Filter

Agents haven't become the only buyers. They've become the first filter. My prediction: by the 2027 holiday season, "assistant-originated" will sit next to "organic" and "paid" on every commerce dashboard worth reading and, as with mobile ten years ago, nobody will remember it being optional.

The old question was: How do we get the customer to our site?

The new one is: When an agent describes us, is it right and does it know how to buy?

You can answer half of that this week, with no budget and no platform change:

• Unblock the right bots: check robots.txt and your CDN or WAF settings
• Test your visibility: ask four AI assistants 10 real questions from your category
• Check your markup: run a top product page through Google’s Rich Results Test
• See how AI describes you: pull a free Palmata report on your brand
• Brief your sources: reach the media and creators the models cite

If you want a second pair of eyes on the other half, our team at DataArt runs a fixed-scope agentic commerce audit in one to two weeks. We agree first on what success means for your business, then deliver a visibility scorecard, a prioritized gap plan, and the first pilot to run. You leave with a clear decision on where to invest.