9 min read
In Checkout.com’s March 2026 survey, payment executives estimated that AI agents were involved in about 3% of transactions, while 89% of merchants said they were preparing for agentic commerce. The distance between those figures is the question: what would turn preparation into repeat purchases?
Checkout.com commissioned Censuswide to survey 12,005 consumers across the UK, US, Brazil, China, France and the UAE, plus 400 heads of payment at consumer-facing merchants, in early March 2026. The merchant sample covered the UK and US. “Involved” is broader than an agent completing an order: discovery can mean software helping a shopper choose a product, while delegated checkout lets it carry out the purchase.
Readiness is not adoption. It can mean a roadmap, a vendor conversation or a configuration change. Consumer responses reinforce the distinction, 24% said they would never delegate a purchase to AI, while 27% said there was no organisation they would trust to operate a shopping agent for them. These are stated attitudes rather than observed behaviour, but they identify a possible constraint without measuring why any individual abandoned delegation.
| Signal | What it supports | What it does not prove |
|---|---|---|
| Agents involved in roughly 3% of transactions | Agent participation is currently a small share of merchant activity | An independently verified share of completed orders; the figure is a merchant survey estimate |
| 89% of merchants preparing | Agentic commerce is on payment teams’ agendas | Deployment, customer use or commercial return |
| 24% would never delegate; 27% trust no operator | Trust is a material adoption constraint | How the same consumers will behave when products and protections improve |
| Adobe, AI referrals converted 54% better in May 2026 | Human visitors arriving from AI services can be commercially valuable | That an AI agent completed the checkout; this is vendor-reported aggregate referral data |
| Shopify, AI-referred orders nearly 13x year on year in Q1 2026 | The channel is growing quickly from its earlier base | Its total share of orders, profitability or the size of that base |
Adobe’s data shows how quickly the discovery side changed. AI-referred U.S. retail visitors converted 38% worse than non-AI traffic in March 2025, 42% better by March 2026 and 54% better by May 2026. Shopify separately reported AI-referred orders growing close to 13 times year on year in Q1 2026, with more than half of those sessions landing directly on a product page.
Both are vendor-reported aggregates, and neither tells us how much of total retail revenue came from AI. They do support a narrower conclusion, shoppers who use AI to narrow a decision may arrive with higher intent. The evidence supports improving product visibility and measurement. It is not yet a reason to assume autonomous checkout will replace the retailer’s site.
The immediate bottleneck is mundane. Adobe estimates that roughly a third of content on the U.S. product pages it assessed was unreadable to models. The estimate uses Adobe’s checker and sample, rather than a universal test of every model. Specifications inside images, variants loaded only after a click and shipping terms hidden in client-side widgets weaken discovery before checkout technology becomes relevant.
Checkout infrastructure still needs permission and product data
In September 2025, OpenAI launched Instant Checkout with US Etsy sellers and named Shopify merchants as the next group. The payment plumbing was the Agentic Commerce Protocol, developed with Stripe. Perplexity later widened its PayPal-powered checkout, Microsoft introduced Copilot Checkout, and Google announced the Universal Commerce Protocol.
The first model also narrowed quickly. On 5 March 2026, Digital Commerce 360 reported that OpenAI was moving Instant Checkout toward purchases through retailer apps inside ChatGPT. Reporting put the live merchant count at about a dozen and the future emphasis on a smaller group of integrated retailers.
Five days later, a federal judge granted Amazon a preliminary injunction blocking Perplexity’s Comet browser agent from password-protected areas of Amazon. The dispute turned on a distinction that extends beyond this case, a customer’s permission does not necessarily authorise an agent to operate on a merchant’s systems. A preliminary ruling is not a final resolution, but it makes access a negotiated commercial and technical question.
| Layer | Examples | Role |
|---|---|---|
| Commerce handoff | ACP, UCP | Catalogue, cart and checkout interactions while the merchant retains fulfilment responsibilities |
| Authorisation | AP2 | Evidence that a user authorised an agent and a purchase |
| Network identity and payment | Mastercard Agent Pay, Visa Trusted Agent Protocol | Agent identity, tokenisation and payment-network controls |
These standards do not all solve the same problem, so treating them as a single winner-takes-all race is misleading. They do, however, depend on a common input, structured, current, machine-readable product information.
OpenAI asks merchants for a refreshed structured feed covering identifiers, descriptions, pricing, inventory, media and fulfilment. Google and Microsoft require similar catalogue infrastructure. The checkout buttons differ, but the underlying shelf is shared.
A retailer framework for the current market
| Priority | Action | Measure |
|---|---|---|
| 1. Make products readable | Test 20 product pages with JavaScript disabled. Check price, availability, specifications, variants, reviews, shipping, returns and product schema. | Percentage of required fields available in raw HTML and structured data |
| 2. Improve the feed | Complete identifiers, descriptions, price, live inventory, fulfilment, compatibility, dimensions, materials and disqualifying use cases. | Feed completeness, freshness and price or stock error rate |
| 3. Separate the channels | Report AI-referred human sessions separately from orders completed through an agent. Treat unattributed AI traffic as a known measurement gap. | Referral conversion, agent-completed order share, refund rate and contribution margin |
| 4. Integrate in proportion to evidence | Use configuration-level or existing payment-provider integrations before bespoke builds. Confirm merchant of record, total take rate, returned customer data and the exit path. | Engineering cost, total fee, repeat use and kept-order margin |
This sequence preserves optionality. The first three priorities improve performance across search, AI referrals and any future checkout protocol. The fourth becomes more valuable only as actual order volume rises.
- Agent-completed order share. Not mentions, referrals or assisted sessions, but purchases completed through an agent.
- Kept-order contribution margin. Revenue after platform fees, payment costs, returns, support and fraud.
- Product-data accuracy. The rate at which price, stock, variants and fulfilment shown to an agent match the merchant’s systems.
- Repeat delegated use. Whether customers use an agent again after the first purchase, rather than merely trying it once.
- Post-purchase failure rate. Refunds, false confirmations, service contacts and disputes attributable to the agent journey.
Microsoft has reported that shopping journeys involving Copilot are 33% shorter, produce 53% more purchases within 30 minutes and are 194% more likely to end in a purchase when intent is present. These are comparisons among shopping journeys, not by themselves causal estimates of delegated-checkout lift, but they come from one provider and do not replace merchant-level economics. If the five measures above improve across a wider base, the case for deeper integration changes quickly.
- The 3%, 89% and consumer-trust figures come from a commissioned survey, not a transaction-level audit.
- Adobe, Shopify and Microsoft figures are vendor-reported aggregates with different definitions and observation windows.
- Rapid percentage growth can begin from a small base. Public sources do not consistently disclose AI commerce as a share of total orders or revenue.
- The Amazon and Perplexity ruling discussed above was preliminary and should not be read as a final rule governing all agent access.
The evidence describes a gap between preparation and reported participation, rather than proving how long it will last. Today, the strongest evidence supports investing upstream, readable products, accurate feeds and honest measurement. Retailers should let observed agent orders, not conference announcements, decide when checkout deserves a bespoke build.
TL;DRTap to expand the short version
- AI-assisted product discovery is growing faster than delegated checkout. Merchants surveyed by Checkout.com estimated agents were involved in roughly 3% of transactions, even as 89% said they were preparing for agentic commerce.
- Referral performance is improving, but it is not proof that shoppers want an agent to complete the purchase. Adobe reported AI-referred retail visitors converting 54% better than non-AI traffic in May 2026.
- Retailers should first make product data readable and measurable. Bespoke checkout work can wait until agent-completed orders, contribution margin and repeat use justify it.
Radisson’s hotel-discovery handoff leaves the transaction on the company’s website. Atlas24h’s chatbot goes further by connecting the conversation to a contract confirmed in its operating system.
FAQ
Agentic checkout lets a shopper complete a purchase through an AI assistant rather than navigating the merchant’s normal checkout. The exact division of payment, fulfilment and customer-service responsibilities depends on the protocol and integration.
Not automatically, but the merchant can lose the direct on-site session and some opportunities for consent capture, merchandising and post-purchase upsells. Check who is merchant of record, what customer data returns with the order and who handles refunds before enabling an integration.
Not entirely. ACP and UCP cover commerce interactions, while AP2 focuses on proving user authorisation. Card-network initiatives add agent identity and tokenisation. They may overlap, but they also operate at different layers.
OpenAI’s approach narrowed in March 2026 toward retailer apps and a smaller group of integrated merchants. The change represents a shift in implementation rather than evidence that AI-assisted shopping has stopped growing.
The legal position is unsettled. In March 2026, a federal judge granted Amazon a preliminary injunction against Perplexity’s Comet browser over access to password-protected areas. The practical lesson is that user permission and merchant authorisation may be separate requirements.
Spotted an error? Contact the desk. Our corrections policy explains how we handle corrections and updates.