Agentic Commerce Booms While Most Product Catalogs Can’t Talk to AI

Nearly half of online shoppers could be leaning on an AI agent to choose what they buy by 2030, Morgan Stanley predicts. Adobe scored the average US product page at just 66% machine readable, meaning roughly a third of it is invisible to the software now doing the shopping.

That gap, not the flashy new checkout protocols grabbing headlines this year, is what is slowing agentic commerce down. Jyotirmoy Dutta, cofounder and chief executive of the ecommerce AI startup Yarnit, argues the industry has the order backward: companies are racing to plug into protocols before fixing the product knowledge those protocols are supposed to carry, he told CXOToday in a recent interview.

What Agentic Commerce Means for Everyday Shopping

Payments processor ACI Worldwide defines the shift plainly, saying “AI shopping assistants act as delegated decision makers for consumers.” Instead of typing a search query and scrolling through blue links, a shopper describes what they want and an agent researches, compares and sometimes buys on their behalf.

Deloitte breaks the shift into stages: assisted discovery, assisted shopping, autonomous shopping, and agent-to-agent commerce, where one company’s AI negotiates directly with another’s. Its retail survey found 63% of global retailers expect to fall behind without AI agents within two years, and 58% think agents will handle most customer interactions within five.

JPMorgan’s payments division frames the stakes bluntly. Building the digital commerce world we shop in today took more than two decades, the bank notes, and it expects agentic commerce to move faster, though not overnight.

The Traffic Numbers Are Already Real, Even at a Tiny Scale

The volume is still small next to total retail sales, but it is growing fast. ChatGPT crossed 900 million weekly active users in February 2026, according to OpenAI, roughly double its count from a year earlier.

Shopify told investors that AI-driven traffic to its merchants grew eightfold year over year in the first quarter of 2026, with orders originating from AI-powered search up nearly thirteenfold and carrying 14% higher average order values than organic search.

• $3 trillion to $5 trillion: McKinsey’s estimate of global retail spend that could move through agentic commerce by 2030
• 393%: year-over-year growth in AI-referred traffic to US retail sites in the first quarter of 2026, per Adobe Analytics
• 20%: share of referral traffic ChatGPT now sends to Walmart and Etsy
• 50 million: shopping-related queries ChatGPT fields every day, based on OpenAI’s own research

Even the sharpest spikes still compound off a tiny base. One tracking analysis of Adobe’s own data found the headline growth figures represent a channel that remains well under 1% of total retail traffic.

The same agentic pattern is popping up well outside retail. Agentic AI tools have already surfaced more than ₹500 crore in finance errors inside Indian corporate ledgers, using the same discover-and-flag logic now aimed at shopping carts.

Two Protocols, One Turf War

Behind those numbers sits a standards fight. OpenAI and Stripe shipped the Agentic Commerce Protocol, or ACP, in September 2025 to power Instant Checkout inside ChatGPT. Etsy sellers began filling orders that originated entirely inside a chat window within weeks, and Shopify said more than a million of its merchants, including Glossier, Vuori, Spanx and SKIMS, would follow.

Google answered three months later. At the National Retail Federation’s January 2026 conference, chief executive Sundar Pichai unveiled the Universal Commerce Protocol, or UCP, co-developed with Shopify and backed by more than 20 partners including Walmart, Target, Mastercard and Visa.

A third standard, the Model Context Protocol (MCP), came from Anthropic, which donated it to the Linux Foundation’s Agentic AI Foundation in December 2025. MCP does not handle transactions itself. It gives an AI agent one way to query outside databases instead of requiring a custom build for every retailer.

“Nobody has figured it out, but everyone has FOMO,” said Emily Pfeiffer, a principal analyst at Forrester who covers AI and commerce. “Everyone is prematurely rushing to market.”

Academic researchers studying agent design say UCP’s payment model is one of its stronger pieces. A paper on delegation frameworks notes the protocol enforces cryptographic proofs for authorizations, which keeps a human accountable even when software executes the trade.

Big Tech is fighting an identical battle one layer up the stack. Meta’s Hatch and Google’s Remy have opened a parallel fight for position in everyday consumer AI agents, a rivalry this site detailed in its own report on the two companies’ competing assistants.

Product Catalogs Are Where the System Breaks First

Even the best protocol is only as good as what it reads. Large language models were trained on the open web, not on inventory feeds, and that mismatch is showing.

Screen-Scraping Was Never Going to Scale

Fast Company’s reporting on the race between Google, OpenAI, Stripe and Walmart found that OpenAI’s first attempt to get products into ChatGPT amounted to screen-scraping retailers like Dick’s Sporting Goods and Ulta, the same brittle method a browser plug-in might use.

An implementation guide written for merchants put the risk bluntly: “Protocol implementation without data readiness is equivalent to building a storefront with empty shelves.” Data, in other words, comes before any protocol.

What CatalogIQ Checks for Every Product

Yarnit’s flagship product, CatalogIQ, is built around that exact gap. It scans customer search behavior, product reviews, frequently asked questions, best-selling items, marketplace requirements and competitor listings, then flags what is missing, Dutta told CXOToday.

“Think of it like a knowledgeable salesperson,” Dutta said. A good one does not just read out a spec sheet; they figure out what a shopper actually needs and connect them to the right product, which he says is the level of judgment AI is now moving toward.

What ends up mattering, in practice, is whether an agent can actually parse the listing. Businesses preparing for that test are focused on a short set of things:

• Structured attributes such as size, material and compatibility, instead of paragraphs of marketing copy
• Direct answers to the questions shoppers keep repeating in reviews and support tickets
• Consistent pricing and inventory feeds across every channel an agent might check
• Delivery and return terms written in a format software can parse, not just a page a person can read

Yarnit’s own product page for the suite says CatalogIQ audits and enriches product listings for AI discovery platforms and traditional search alike, benchmarking a catalog against competitor listings and marketplace rules.

Can You Trust an Agent with Your Credit Card?

Not fully, and not yet. Payment networks are rolling out safeguards such as tokenized credentials and identity checks, but the people building these agents admit the technology still makes expensive mistakes, and surveys show only a small share of shoppers are comfortable letting one buy without a final human review.

The mistakes are not hypothetical. Sebastian Heyneman, founder of a San Francisco startup, asked an AI agent built by Tasklet to land him a speaking slot at the World Economic Forum in Davos. It succeeded, CBS News reported, at a cost of $30,000 he had not authorized.

Andrew Lee, Tasklet’s founder, told CBS News the trouble usually starts with conflicting instructions in a user’s prompt. “The specific use case of shopping is not a good thing to use these systems for,” he said, adding it simply is not ready yet. “The agents are fundamentally hard to trust.”

Matt Kropp, an AI expert with Boston Consulting Group, put the caution more plainly: “It could potentially go buy a car, but I wouldn’t say, ‘Here’s my credit card.’”

The numbers back up the hesitation. A Payments Association survey found 58% of UK online merchants believe agents are already active on their platforms, yet only 3% of transactions today actually involve one. Separately, an estimate from data infrastructure vendor MetaRouter puts the share of financial institutions bracing for an AI-driven fraud spike at 78%.

American Express has already moved to plug the gap. The card issuer rolled out protections this year for cardholders shopping through approved AI agents, verifying an agent’s identity before a charge clears and covering customers for errors the agent makes on its own.

Amazon Sits This One Out

Not every retailer is racing toward the same standard. Amazon has blocked OpenAI’s ChatGPT-User and OAI-SearchBot crawlers in its robots.txt file, meaning Amazon listings cannot show up in ChatGPT’s shopping results in real time.

The company receives less than 3% of its traffic from ChatGPT referrals today, a share that has been declining 18% month over month, according to Similarweb data cited by ecommerce analysts. Target pulls about 15% of its traffic from ChatGPT the same way, and eBay around 10%, brands that never blocked the crawler in the first place.

Amazon’s logic is defensive. The block protects an advertising business estimated at $4 billion by keeping shoppers inside its own search results. The trade-off is a head start for any brand that also sells through its own Shopify storefront, where the same products remain visible to AI agents.

Brands Are Already Rebuilding Around What AI Can Read

Yarnit’s answer to that fragmentation is a layer it calls Memora, a memory system meant to give an AI agent the same institutional knowledge a longtime employee builds up over years, according to Dutta.

Memora retains and updates what an agent knows about products, customers, brand rules and past interactions, then connects to whatever a company already runs, whether that is a product information management system (PIM), a digital asset manager (DAM), an enterprise resource planner (ERP) or a content management system (CMS), Dutta told CXOToday.

Not every fix gets automated. Routine updates, such as filling in a missing attribute or adapting content for a new channel, can run without a person in the loop. Anything touching regulated claims, pricing or brand positioning stays with a human reviewer, Dutta said.

The company is betting that this positioning matters more over a five-year horizon than any single protocol war. Commerce leads at Google and OpenAI told Fast Company they expect the industry to reach a tipping point within months, not years.

Ultimately, AI will become another channel through which customers discover and evaluate brands, and the businesses that succeed will be the ones whose products are easiest for AI to understand, trust and recommend.

Jyotirmoy Dutta, cofounder and chief executive of Yarnit, said in the interview.

Yarnit already bundles that philosophy into a wider suite. Beyond CatalogIQ, the company’s press materials describe a Campaign OS that automates re-engagement across email, WhatsApp and SMS, plus an AI Sales Rep that guides on-site visitors using real-time inventory and margin data.

Whoever keeps the cleanest, most trustworthy catalog wins the recommendation, no matter which protocol carries the checkout.

Frequently Asked Questions

What Is Agentic Commerce?

Agentic commerce is online shopping where an AI agent researches, compares and sometimes buys on a person’s behalf instead of the person clicking through search results themselves. Deloitte calls the broader shift a-commerce for short, and splits it into four stages running from simple product recommendations up to one company’s AI negotiating directly with another’s.

What Is the Difference Between ACP and UCP?

ACP, built by OpenAI and Stripe, focuses on the checkout moment inside ChatGPT. UCP, built by Google and Shopify, covers the full shopping lifecycle across any participating agent. The cost gap between them is real, with UCP’s transaction fee near 3.2% against roughly 7.2% for ACP, though one analysis found merchants running both protocols see up to 40% more agent-driven traffic.

Can an AI Agent Buy Something Without Asking Me First?

Technically yes, within limits a shopper sets in advance, but the agents still struggle in practice. In one comparison test called WebMall, the strongest shopping agent completed under 65% of harder tasks like finding the cheapest offer across multiple stores, and on a separate benchmark called DeepShop, the top system solved only 20% of its hardest queries.

Is My Payment Information Safe with a Shopping Agent?

The payment networks say increasingly yes. Mastercard’s Agent Pay lets verified AI shopping agents transact using tokenized credentials rather than a raw card number, and Visa offers similar AI-ready cards through its Intelligent Commerce network, so a merchant and the cardholder can trace a purchase back to the specific agent that made it.

What Does Yarnit’s CatalogIQ Check for?

Beyond attributes and reviews, CatalogIQ looks at seasonal trends, best-seller rankings and brand guidelines, then scores how a product page compares with competitor listings. Yarnit’s separate Creative OS tool exists because a brand often has under ten seconds to convince a shopper a product deserves attention, so the two tools handle discovery and persuasion as one pipeline.

When Will Agentic Commerce Become Mainstream?

Sooner than the current sliver of retail traffic suggests, though timelines vary by forecaster. Bain projects agentic channels could carry 15% to 25% of US ecommerce sales by 2030, while Gartner expects AI agents to intermediate $15 trillion in global B2B purchases by 2028, years before consumer habits fully catch up.