Quick answer: agent-to-agent commerce — where your customer's AI agent transacts with your business's systems — went from announcement to production to partial retreat in about twelve months. The checkout layer sprinted and stumbled: OpenAI launched Instant Checkout in ChatGPT in September 2025, expanded it to all US users in February 2026, then shut it down in March 2026 after roughly 30 Shopify merchants (of a million promised) had actually integrated. The discovery layer, meanwhile, quietly boomed: Shopify's own telemetry shows AI-referred traffic up ~8x and AI-driven orders up ~13x year-over-year in Q1 2026, and Adobe measured AI-sourced retail traffic growing 393% and converting 42% better than organic by March 2026. The lesson of year one is the whole strategy: checkout mechanics keep churning, but being findable and machine-readable — the discovery layer — is where the volume, the conversions, and the durable work all live.
This is the honest chronicle of agentic commerce's first year, with every claim graded the way we grade everything: established, emerging, or hype. It updates our August piece on agentic shopping, written before the checkout layer had been tested at scale. This space changed materially three times in twelve months; we re-verify the protocol status lines before every republish.
What is agent-to-agent commerce, precisely?
The full vision: your customer tells their AI agent what they need ("running shoes for trail use, under $150, delivered by Friday"), and the agent interprets the intent, queries merchant catalogs, compares options, and completes the purchase — talking to your systems (your product feed, your policies, your checkout endpoint) rather than browsing your website like a human. When the merchant side is also automated — structured feeds answering structured queries, policies expressed in machine-readable form, an agent-compatible checkout — you have machines transacting with machines. The human sets parameters; software does the shopping.
How much of that exists today? The discovery half is real and scaled. The transaction half is real but wobbly. The negotiation half — your agent haggling with a merchant's agent — is still mostly a conference-slide vision. Let's take them in order of what actually happened.
The twelve-month timeline (established facts, dated)
September 2025: OpenAI and Stripe launch the Agentic Commerce Protocol (ACP) — an open standard, Apache 2.0 licensed — alongside Instant Checkout in ChatGPT: US users buying from Etsy merchants in-chat, with over a million Shopify merchants announced as "coming soon." Merchants remain the merchant of record; OpenAI takes a transaction fee (reported at 4%) while stating that product recommendations are based on relevance, not enrollment.
November 2025 – January 2026: the land rush. Perplexity launches Instant Buy with PayPal; Salesforce announces ACP support; holiday data lands — Salesforce ties AI and agents to 20% of retail sales and $262 billion during the 2025 holiday season, and Adobe records 805% year-over-year growth in AI-driven retail traffic on Black Friday, with AI-referred visitors completing purchases at a 38% higher rate. (Grade these vendor figures as directional: "AI-influenced" is a broad, unaudited definition — but the direction is corroborated across independent platforms.) In January, Microsoft launches Copilot Checkout, and Google unveils the Universal Commerce Protocol (UCP) at NRF with co-development from Shopify, Etsy, Wayfair, and Target, and endorsements from 20+ partners including Stripe, Visa, Mastercard, and American Express.
February 2026: OpenAI expands "Buy it in ChatGPT" to all US users, including the free tier. The card networks go live in earnest through the spring — Mastercard completes its first live agentic transactions in Asia-Pacific markets, Visa's Trusted Agent Protocol goes commercial after piloting with 100+ partners, and American Express ships an agentic developer kit with purchase protection for registered AI-agent purchases. Shopify flips on Agentic Storefronts by default, syndicating eligible merchants' catalogs to ChatGPT, Google AI Mode, Microsoft Copilot, and Perplexity simultaneously.
March 2026 — the retreat: OpenAI sunsets Instant Checkout. The stated reason: the initial version didn't offer the flexibility they wanted, so merchants now use their own checkout while OpenAI focuses on product discovery. The revealing numbers around that decision: Forrester's Emily Pfeiffer put actual Shopify merchant integrations at roughly 30 — against the million promised — and Walmart, which had made about 200,000 products available in ChatGPT, found conversion rates three times lower for in-chat purchases than for shoppers redirected to Walmart's own site. ACP itself survives — it changed jobs from checkout to discovery infrastructure — and the ecosystem's center of gravity shifted to "discover in the chat, transact on the merchant's site."
Meanwhile, one giant sat the whole dance out: Amazon joined neither ACP nor UCP, building its own walled garden instead — Rufus, its shopping assistant, reportedly serves 300 million users and drove an estimated $12 billion in incremental sales in 2025.
So is agentic commerce failing? No — the funnel just sorted itself
Read year one carefully and a clean pattern emerges: every layer of the agent funnel matured at a different speed, and the order is instructive.
Discovery: established, and compounding. This is where the verified platform telemetry lives — Shopify's ~8x traffic and ~13x order multiples are counted sessions on real infrastructure, not surveys. Adobe's Q1 2026 data (393% traffic growth, 42% conversion advantage) is the same class. People are letting AI find and shortlist things at enormous and accelerating scale.
Transaction: emerging, and messier than the press releases. Consumers use AI to decide, then prefer to buy in familiar places — a Semrush survey found only 22% had ever bought a product inside an AI tool, while about half had purchased somewhere after using AI in their research. The trust numbers explain the Walmart conversion gap: depending on the survey, only 4% to ~30% of consumers trust AI at the payment-authorization step, versus ~62–63% happily using it for comparison and discovery. And when people do trust an agent to buy, Bain found they trust retailer-owned agents three times more than third-party ones. The rails are being built ahead of the riders — normal for infrastructure, fatal for anyone who confuses rails with demand.
Negotiation (true agent-to-agent): speculation with scaffolding. Merchant-side selling agents, machine-readable offers, agent identity verification — the pieces are appearing (agentic feed products, bot-management vendors repositioning to let legitimate shopping agents through while blocking scrapers), but live agent-negotiates-with-agent commerce is not a 2026 reality for any normal business. The forecasts here are enormous — McKinsey sketches $3–5 trillion in redirected global retail spend by 2030, Gartner projects $15 trillion in agent-intermediated B2B purchases by 2028 — and they are forecasts, not measurements. Respect the direction; don't spend against the number.
How do agents pick which businesses to transact with? (The part that decides winners)
Strip away the protocol politics and this is the question that matters, because whichever checkout standard wins, the selection step comes first — and it works nothing like the ad auction merchants are used to.
Agents choose programmatically. When an agent evaluates candidates for "birthday gift, skincare, under $75," it queries structured catalogs and scores what it can verify: data completeness, price, live availability, shipping and return policies it can actually read, reviews, and merchant trust signals. Practitioner data points in one direction: merchants with near-complete structured attributes get surfaced dramatically more; a business with a beautiful website and a thin machine-readable feed gets skipped for one with a plain site and a clean feed. Policies matter in a new way too — an agent qualifying merchants against "must allow 30-day returns" needs your returns policy as structured markup, not as prose on a page it half-parses. And notably: you can't buy your way in — recommendations in the major surfaces are relevance-based, which means the quality of your machine-readable presence is the primary lever. For small businesses, that's the most level playing field commerce has offered in decades.
The readiness gap is the opportunity: early-2026 research found 40% of e-commerce businesses still standardizing their product data for agents and another 33% not started at all. Nearly three-quarters of your competitors haven't done the unglamorous work.
What should a small or medium business actually do?
The March retreat is your strategy guide: OpenAI itself pulled back from checkout to focus on discovery — which tells you which layer is durable. In order:
- Win the discovery layer first — it's the same GEO work. Before any agent transacts with you, an AI system has to find you, classify you, and trust you. Crawler access, a plainly stated identity, extractable content, presence in cited sources — the entire visibility stack is the top of the agent funnel. This pays today (the 42%-better-converting AI-referred humans) and pre-qualifies you for whatever agents do tomorrow.
- Make your commerce facts machine-readable. Product schema with real prices and live availability, structured shipping and returns policies, complete attributes on your top products. If you're on Shopify, much of this now ships by default via Agentic Storefronts — verify it's on and your data is complete rather than assuming (our platform study covers what each platform switches on for you).
- Let your platform carry the protocol risk. ACP vs. UCP vs. whatever's next is a war between giants; Shopify, BigCommerce, Wix, and the payment networks are integrating the winners on your behalf. Don't hand-build protocol integrations; do keep the toggle on and the feed clean.
- Skip the hype spend. No paid "agent optimization" service, no protocol consultancy, no panic replatforming. The two-question test still governs: cost if it does nothing, evidence it does something.
The one-sentence version of year one: the machines learned to shop before the humans agreed to let them pay — and the businesses winning both eras are the ones machines can read. That reading layer is what SeeGeo's free audit measures — crawler access, machine-readable identity, structured data, extractability — in about 20 seconds, free, no signup for your score.
Frequently asked questions
What is the difference between agentic commerce and agent-to-agent commerce? Agentic commerce is the umbrella: AI agents researching, comparing, and increasingly transacting on a shopper's behalf. Agent-to-agent commerce is the fuller vision where the merchant side is also automated — the buyer's agent interacting with a merchant's structured feeds, policies, and checkout endpoints rather than a human-oriented website. Discovery-stage agentic commerce is established at scale in 2026; true agent-to-agent negotiation remains early.
Did ChatGPT shopping fail? The in-chat checkout version was retired in March 2026 — after low merchant uptake (~30 Shopify integrations) and weak in-chat conversion (Walmart measured 3x lower than site checkout) — but ChatGPT shopping discovery is bigger than ever, and the underlying Agentic Commerce Protocol survives as infrastructure. The accurate summary: checkout retreated, discovery won.
What is the difference between ACP and UCP? ACP (Agentic Commerce Protocol) is the open standard from OpenAI and Stripe, launched September 2025, now oriented toward discovery and merchant-side checkout. UCP (Universal Commerce Protocol) is Google's coalition protocol, announced January 2026 with 20+ partners, aimed at Google's AI surfaces. They are not yet interoperable; most SMBs should let their commerce platform handle both rather than integrating directly.
How do AI agents decide which store to buy from? Programmatically: structured data completeness, price, live availability, machine-readable shipping and return policies, reviews, and trust signals — evaluated against the shopper's stated constraints. Ad spend doesn't enter it; recommendations on the major surfaces are relevance-based. The practical consequence: catalog and policy data quality is the primary competitive lever, which favors well-organized small merchants over big-but-sloppy ones.
Will consumers actually let agents buy things for them? Slowly. Trust collapses at the payment step — surveys through 2026 put willingness to let AI spend autonomously between 4% and ~30%, versus ~62–63% comfortable using AI for comparison and discovery — and consumers trust retailer-run agents about 3x more than third-party ones (Bain). Expect bounded, chore-like purchases first (reorders, bookings, gifts within a budget) while high-consideration buying stays human-approved.
Is any of this relevant if I'm a small local business, not an e-commerce store? Yes — the discovery layer is identical. Agents and assistants recommending local services run on the same machine-legibility: clear identity, consistent listings, readable policies and hours, structured data (our local guide walks through it). The transaction layer (bookings via agents) will reach services through reservation and scheduling platforms the same way checkout reached retail through Shopify — your job now is being findable and classifiable when the recommendation happens.