A year ago, "AI shopper" meant a person asking ChatGPT which running shoes to buy and then opening a browser tab. That still happens, and it happens far more often. What is new in 2026 is the machinery being built so the agent can finish the job: find the product, check the price and stock, and pay, without the person ever seeing the store.
That machinery is arriving in pieces, some of it faster than expected and some of it slower. Here is where it actually stands, and what it means if you run a store.
The traffic is already here
Adobe's numbers are the cleanest signal. AI-referred traffic to US retail sites grew 393% year over year in the first quarter of 2026, and has grown more than fourteen-fold since Adobe started tracking it in October 2024. Those visitors spend 48% longer on the page and convert 42% better than visitors from other channels. Salesforce, from a different dataset, reports that agentic search as the first step in a purchase grew 200% year over year.
The assistant is doing the discovery and the comparison. The store is still doing the sale. That split turns out to matter.
The protocols, in plain terms
Three families of standards shipped in the last year, and they solve different problems.
The Agentic Commerce Protocol, from OpenAI and Stripe, is a way for a merchant to expose a product feed and a checkout so an assistant can complete a purchase. It is open, and works with payment providers other than Stripe. It launched with Etsy and Shopify merchants inside ChatGPT's Instant Checkout in late 2025.
The Agent Payments Protocol, AP2, started at Google and is now governed by the FIDO Alliance. It solves the question underneath every agent purchase: how does a merchant know a person actually authorised this? AP2 uses signed mandates, built on verifiable credentials, that record what the shopper asked for and what they agreed to pay. More than sixty organisations, including Mastercard, Visa, PayPal and Adyen, contribute to it.
The card networks' own schemes, Visa Intelligent Commerce and Mastercard Agent Pay, handle the payment leg. They let an agent hold a tokenised card credential with limits, so the merchant sees a recognisable, bounded payment, and never a raw card number held by a bot. Both networks joined AP2 in September 2025, so the layers are converging.
If you sell on Shopify, most of this is done for you. Shopify's Agentic Storefronts went live by default for eligible merchants in March 2026, packaging catalogue, policies and checkout so ChatGPT, Copilot, Gemini and Perplexity can present the store natively. The platform also exposes a Storefront MCP server with tools to search the catalogue, manage a cart and answer policy questions, which any agent can call.
Why in-chat checkout stalled
The most useful lesson of the year is one the vendors did not plan. In March 2026 OpenAI deprioritised Instant Checkout inside ChatGPT in favour of merchant apps that run inside the assistant and hand off to the merchant's own checkout. Shopify's president listed the reasons any operator would recognise: subscriptions, inventory, shipping, taxes, merchandising rules. Checkout is where a retailer's complexity lives, and merchants were not willing to hand that to a chat window.
The pattern the industry has settled on for now is discover in the assistant, buy on the store. That is good news for merchants. It means the near-term work is making the store legible to agents, not rebuilding checkout.
Being readable to a machine
This is where most stores are behind. Adobe ran its content visibility checker across retail categories and found that assistants could read only about half of what is on a typical retail product page: cosmetics sites scored 63%, electronics 56%, apparel 51%, grocery 48%. The rest is rendered in ways a model cannot parse, or simply missing.
What an agent needs is unglamorous. Structured product data with the attributes a shopper would ask about: size, material, compatibility, what is in the box. Accurate stock and price, updated when they change, not once a night. Policies in plain text: returns, shipping times, warranty. A checkout an agent can reach through a standard interface, with no pop-ups in the way.
McKinsey's framing is that retailers who make their inventory, pricing and fulfilment options machine readable and policy safe will show up in agent recommendations more often. The agent is a buyer that reads everything and remembers nothing you did not write down.
What changes in the economics
Discovery gets flatter. Shopify's president argues agents are a more neutral channel than search, because they optimise for the shopper's preferences, with no commission pulling the ranking. A small brand with a precise product and clean data can win a recommendation it would never have won on a results page.
Price comparison becomes instant and total. When 62% of consumers already use AI to compare price and brand, a store cannot rely on a shopper not noticing that a competitor is cheaper. Differentiation moves to things an agent can verify: delivery speed, return terms, stock availability, reviews.
Loyalty has to be restated in terms an agent can use. A points balance and a member price only matter if the agent shopping for the customer can see them and apply them. Programmes that live only inside a mobile app become invisible.
The size of it
ICSC and McKinsey estimate that US retail revenue orchestrated through agents could reach $1 trillion by 2030, with a global figure of three to five trillion. Forecasts like that are always wrong in the details. The direction is not in doubt, because the traffic is already measurable and the protocols are already shipping.
The practical advice is smaller than the headline. Make the store readable. Keep stock and price honest in real time. Put policies in text. Expose checkout through the standard your platform supports. Then watch the AI-referred line in analytics, because it is the one that is growing.