What is agentic commerce? A guide to thriving in the AI shopping era

By 2030, AI agents could mediate $3 trillion to $5 trillion in global consumer commerce, with up to $1 trillion in US B2C retail alone, according to McKinsey’s October 2025 report, The agentic commerce opportunity. McKinsey expects this shift to match the impact of the web and mobile commerce revolutions, and possibly move faster, because agents use the same digital paths to purchase that humans already use.

This is agentic commerce: buyers delegating shopping to AI agents.

If your business model depends on humans browsing, comparing, and clicking, this changes your economics. Marketplaces, loyalty programs, and ad-funded discovery were all designed to influence people. Now you have to ask a harder question: who are you selling to when the buyer is a model acting on someone’s behalf?

In this guide, you’ll learn what agentic commerce is, what the behavior data shows so far, which protocols now define how agents transact, and what your product data has to become before agents can find, trust, and buy your products.

What does the data show about AI shopping so far?

Forecasts tell you where analysts think this goes. Consumer behavior tells you what buyers are already doing. Two of the largest commerce datasets, Adobe’s and Salesforce’s, now measure agentic shopping directly, and both point the same way.

AI shopping traffic grew 393% in one year

Traffic from AI sources to US retail sites grew 393% year over year in the first quarter of 2026, according to Adobe Analytics, which based its analysis on more than 1 trillion visits to US retail sites. 

This traffic also converts better than traffic from non-AI channels such as paid search and email. That’s a reversal from a year earlier, when AI-referred visitors converted worse. Buyers arriving from an AI tool have already done their research inside it. By the time they reach your site, they’re closer to a decision than almost any other visitor you get.

AI agents influenced $262 billion in holiday sales

During the 2025 holiday season, AI and agents influenced 20% of all global retail sales, worth $262 billion, according to Salesforce data covering more than 1.5 billion shoppers. 

Salesforce also reported that retailers running their own shopper agents grew sales faster than retailers without them. The takeaway is hard to argue with: agents stopped being an experiment during the biggest shopping season of the year and became a revenue channel.

Structured feeds can make your products accessible to AI agents, but accessibility alone does not make the data reliable. See how manufacturers are approaching AI adoption, verification, and product data quality inside PIM.

Get the AI in PIM report

Why aren’t buyers letting AI agents check out yet?

The growth in sections above is real, but it’s concentrated at the top of the journey. Discovery and research have gone agentic. The final step, handing an agent the payment, hasn’t. The clearest evidence is what happened to the market’s most prominent experiment.

OpenAI launched agentic checkout, then pulled it back in five months

OpenAI launched Instant Checkout inside ChatGPT on September 29, 2025, promising products from over a million Shopify merchants. On March 4, 2026, it scaled back the feature, moving purchases into merchant apps within ChatGPT or back to merchants’ own websites. 

According to Forrester principal analyst Emily Pfeiffer, roughly 30 Shopify merchants were live with the feature at the time, far fewer than the million promised. The Information, which broke the story, reported the core reason: people were using ChatGPT to research products, but not to buy through it.

Buyers research with AI but still buy elsewhere

Forrester’s consumer data explains why. As of October 2025, only 8% of US online adults had used Instant Checkout, and completing a purchase inside an answer engine remains the least-adopted use case even among regular users. 

Trust is the constraint: 54% of US online adults say they’re not comfortable giving personal information to generative AI tools. Buyers will let an agent do the homework. They won’t let it hold the wallet yet.

Bad product data helped break agentic checkout

Here’s the detail most coverage skipped. Walmart made about 200,000 products available through Instant Checkout, but according to Pfeiffer, the product data scraped from retailer websites was often inaccurate on stock availability, delivery timing, and shipping costs. 

Forrester has also raised a question no one in the ecosystem has answered: when an agent places an incorrect order due to inaccurate product content, who is liable for the return? 

Agentic checkout didn’t stall only because buyers hesitated. It stalled because the underlying data couldn’t be trusted.

Which protocols power agentic commerce?

For an agent to shop, merchants and AI platforms need a shared language for product data, checkout, and payment. Through 2025 and 2026, that language went from missing to contested: there are now competing open standards, and which ones win will shape where your products can be sold. Three matter most right now.

1. OpenAI’s ACP set the first standard

OpenAI launched the Agentic Commerce Protocol in September 2025, developed with Stripe and released as open source. ACP defines how merchants publish structured product feeds and support agent-driven checkout. And it outlived the Instant Checkout pullback: OpenAI’s revised approach still runs on ACP, with retailers like Walmart, Shopify, and Etsy building their own apps inside ChatGPT on top of it.

2. Google’s UCP turned protocols into a standards race

At NRF in January 2026, Google announced the Universal Commerce Protocol, an open-source standard covering the full journey: discovery, buying, and post-purchase support. 

Google co-developed it with Shopify, Etsy, Wayfair, Target, and Walmart, and it has endorsements from more than 20 partners, including American Express, Mastercard, Stripe, and Visa. It’s built to work with existing retail infrastructure and is compatible with the Agent Payments Protocol (AP2) for secure agentic payments. Forrester’s read is blunt: this is a race for platform advantage, with Google second to market after OpenAI. 

3. MCP connects agents to the systems underneath

The Model Context Protocol, introduced by Anthropic, isn’t a commerce protocol. It’s the general standard for connecting AI agents to data sources and tools, and it sits underneath the commerce layer: McKinsey names MCP among the integration enablers merchants need to master, and Google lists MCP as one of the ways businesses can integrate with UCP. 

If ACP and UCP define how agents buy, MCP defines how agents reach your systems and data at all.

Not every platform is picking a shared standard

Amazon has joined neither protocol. It’s building its own agents (Rufus, Buy for Me) within its ecosystem, and in March 2026 it opened merchant feeds in its Shop Direct experience, just weeks after announcing a strategic partnership with OpenAI. The practical consequence for you is that no single integration will cover every agentic surface. Plan for several.

What do AI agents need from your product data?

Everything above converges here. Agents don’t browse your site the way a person does; they read your data. Whether your products get discovered, recommended, and eventually bought by agents comes down to three requirements the platforms have now put in writing.

1. Agents read structured feeds, not your website

OpenAI’s product feed specification is explicit: ChatGPT indexes a structured feed the merchant pushes directly, not data crawled from your site. The spec accepts refreshes as often as every 15 minutes to keep prices and stock current, and it includes required fields specifically to ensure accurate pricing and availability. 

Google moved in the same direction, adding new Merchant Center data attributes to help retailers get discovered in conversational commerce. Read those requirements as an operational test: can your systems produce a complete, accurate, correctly formatted feed for every product, for every platform, on a cycle measured in minutes? For most merchants, honestly, the answer today is no.

2. A large share of retail content is still invisible to AI

Adobe, alongside its traffic data, published a benchmark of AI visibility across US retail websites using its AI Content Visibility Checker, which scores what large language models can and cannot read on a page. 

Its finding: large portions of retail site content are not readable by the same AI models driving the traffic growth. The buyers are arriving through AI. The product data they need usually exists; AI just can’t read it.

3. Completeness drives ranking, accuracy drives trust

OpenAI’s documentation states that recommended feed attributes, such as rich media, reviews, and performance signals, improve ranking, relevance, and user trust within ChatGPT. So an incomplete product record not only looks unfinished; it also ranks lower in the channel that’s growing fastest, and inaccuracies cost even more. AI product data enrichment closes those gaps at the scale the channel demands.

Remember why agentic checkout stalled: stock, delivery, and shipping data the agents couldn’t trust. In agentic commerce, your product data is your storefront, your salesperson, and your reputation, all at once. 

How do you prepare for agentic commerce?

The evidence points to a clear sequence. You don’t need to predict which platform wins or when agentic checkout returns. You need to be ready for the parts that are already true.

1. Compete for discovery first

AI-referred buyers are the fastest-growing traffic source in retail, and they arrive closer to a decision than visitors from any other channel. The discovery layer is where agents are already filtering products in and out, so that’s where your effort goes first: product content that answers the questions agents ask: fit, compatibility, constraints, differences from alternatives, rather than content written only to inspire a human browser.

2. Fix the fields that broke checkout the first time

Stock availability, delivery timing, shipping costs, and pricing are the exact fields that undermined the market’s biggest agentic checkout experiment. Audit them across every channel where an agent might read them. If those fields disagree between your site, your feeds, and your retail partners, an agent has no way to know which version to trust, and neither does the buyer behind it.

3. Build a feed operation, not a feed file

The platforms have defined what participation costs: structured feeds per protocol, required fields, refresh cycles as short as 15 minutes. That’s not a one-time export. It takes a retail digital catalog run as an ongoing operation that must produce complete, correctly formatted, up-to-date data for every product across ACP, UCP, and whatever comes next, because no single integration covers every agentic surface. Evaluate honestly whether your current systems can run that operation or were built for a slower era.

4. Verify everything AI reads and generates 

The liability question Forrester raised remains open: when an agent orders incorrectly due to poor product content, someone bears the cost. Until the ecosystem answers it, the only protection you control is governance on your side; validation that checks AI-generated and AI-consumed product data against trusted records before it reaches any channel. Speed without verification is how you end up as the cautionary tale in someone else’s article.

5. Match delegation to the purchase, not the hype

McKinsey’s framework is worth keeping in mind as you prioritize: delegation is a curve, and different product categories sit at different points on it. Routine, low-risk purchases go agentic first, which is why agentic commerce in B2B, with its reorders and repeat procurement, is moving early. Considered purchases stay human longer.

Build product data agents can trust 

Every step above lands on the same foundation. Discovery runs on your product content; the fields that broke checkout were product data fields, and the feeds and verification the platforms require are product information work. 

That’s what Inriver does. Inriver’s flexible data model ingests product data from any source, as-is, with no upfront cleanup or ETL project, so your data becomes AI-ready fast. Inriver then orchestrates AI enrichment across PIM and LLM workflows, with built-in verification and validation that checks output against trusted product data before it reaches a channel. Agentic speed, with control you can demonstrate.

If an agent can’t read your product data, it will recommend someone else’s.

Learn how you can make your product data AI-ready; schedule a personalized demo today.