AI agents are shopping for your customers now — and most catalogs aren't built to sell to one

A growing share of shopping no longer starts with a person typing into a search box and clicking through blue links — it starts with a person telling an AI assistant what they want and the assistant doing the browsing, the comparing, and increasingly the buying. ChatGPT, Google's Gemini-powered AI Mode, and Salesforce's Agentforce have all shipped a way for an agent to complete a transaction with little more than a spending limit and a nod from the human it's shopping for. That's a different problem from the one this site covered a few weeks ago — being cited in an AI answer is about your business getting mentioned. This is about whether an agent can actually buy from you, and it's a far more mechanical question.

The numbers say this is not a rounding error. AI-assistant-referred shopping traffic grew 119% year over year in the first half of 2025, and by the 2025 holiday season AI and agents were driving roughly a fifth of global online orders — about $262 billion in real, counted sales, not a forecast. Nearly half the people already using AI to shop say they're open to letting the agent finish the purchase itself, not just recommend it.

None of that means the plumbing is settled — a few months later, the company leading the push quietly scaled its own in-chat checkout back down. What follows is what agentic commerce actually requires from a business underneath the hype cycle, the two standards competing to define it, and — more usefully — what a small or mid-size operator should build now versus what can wait.

Start with what 'agentic commerce' actually means, because it's easy to confuse with the AI-answer-box territory search behavior went through this past year. Being cited in an AI Overview or a ChatGPT answer is about being mentioned — a different, more mechanical step comes after: an AI agent building a cart, checking your price and stock, and completing a transaction with little more than a spending limit and a nod from the human it's shopping for. McKinsey's research puts the scale of that shift at up to $1 trillion in orchestrated US retail revenue and $3 trillion to $5 trillion globally by 2030 ^[1] — treated by the firm not as a speculative product category but as a structural change in how a purchase gets executed, on top of how it gets decided.

That number is not purely a 2030 forecast either — it's already showing up in receipts. Salesforce's pre-holiday research tracked AI-assistant-referred shopping traffic growing 119% year over year in the first half of 2025 alone, and projected AI and agents would drive 21% of all global holiday orders that season — about $263 billion. The same research found 48% of shoppers who already use AI to shop said they were open to letting the agent finish the purchase itself rather than just point them toward it ^[2]. The actual season landed close to that projection: roughly $262 billion in AI-driven online sales, close to a fifth of global online holiday revenue, inside a record $1.29 trillion global total ^[3]. This moved real, countable money in one quarter — it isn't a hypothetical channel anymore.

Two competing technical standards emerged in 2025 to make an agent-completed purchase actually work, and the mechanics matter because they define what a business has to expose. OpenAI and Stripe co-developed the Agentic Commerce Protocol (ACP): a merchant publishes a structured product catalog, an agent queries it, builds a cart, and Stripe handles the payment step through a standardized API — by February 2026, OpenAI had extended 'Buy it in ChatGPT' to more than a million connected Shopify merchants plus PayPal as a second payment processor ^[4]. Google took a different route with its Agent Payments Protocol (AP2), developed with more than 60 partners including Mastercard, American Express, PayPal, and Salesforce: it chains three cryptographically signed authorizations — an Intent Mandate, a Cart Mandate the merchant itself signs to guarantee the price, and a Payment Mandate the payment network verifies — so a merchant can prove a real, specific purchase was actually authorized by a real person ^[5]. The mechanics differ; the underlying requirement doesn't. Both assume your product data already exists somewhere a machine can query directly, not just a page a person can read.

That requirement is the actual work, and it's unglamorous: a structured product feed — schema.org Product and Offer markup, or an equivalent structured catalog, carrying accurate price, stock level, and a stable product identifier (SKU or GTIN) — kept in sync with whatever a human sees on the page. Some of this may already be half-done without a business realizing it. Shopify alone now syndicates catalog data from its 5.6 million connected merchants into these systems by default through what it calls Agentic Storefronts, meaning a store on the right platform can already be technically visible to an agent without anyone having built anything new ^[6]. But visible isn't the same as trustworthy, and a feed with the ordinary small-business gaps — a missing SKU, a stock count that hasn't synced in two days, a price the page shows that doesn't match what checkout actually charges — is exactly the kind of source an agent-vetting layer learns to route around rather than trust with real money.

None of this should read as a settled, mature channel — the two companies furthest ahead are still actively renegotiating what it should even look like. In March 2026, OpenAI scaled its own in-chat checkout back down after finding that people used ChatGPT to research products but rarely completed a purchase there; Walmart's own numbers, cited in the reporting, showed a checkout completed inside ChatGPT converted at roughly a third the rate of sending the same shopper on to walmart.com to finish it. OpenAI's stated shift was toward discovery and evaluation inside the chat, with the actual purchase handed off to the retailer's own site ^[7]. The honest read is that agentic checkout specifically is still an open bet — but the structured-data foundation underneath it isn't a bet at all. That same clean, accurate product data is what earns a rich result in ordinary Google search and what an AI Overview quotes from when it answers a shopping question, independent of whether in-chat checkout itself survives in its current form.

For most small and mid-size businesses, then, the actual project looks less like 'connect to ChatGPT commerce' and more like a data-integrity exercise a business should be doing anyway: one place where price, stock, and product identity live, kept accurate, and served consistently everywhere it needs to show up — your own website, your platform's product feed, any protocol connection you eventually add. The alternative is the pattern most small catalogs already carry without noticing it: a website that says one price, a spreadsheet that says another, and a shelf that says a third — tolerable when a person double-checks at the register, and exactly the gap an autonomous purchase doesn't pause to notice.