Agentic Commerce: AI Is Already Buying for Your Customers and Your Store Is Not Ready

For twenty years we optimized for a human being: someone distracted, thumb hovering over a screen, who had to be won over with a nice photo, a punchy headline and a brightly colored button. That human is leaving. And not because they stopped buying. They are leaving because they sent someone else in their place.

That someone is an AI agent. And it could not care less about your bright orange button.

The number nobody puts in the headline

Tiendanube México reported that traffic arriving at online stores from ChatGPT and other AI assistants grew 396% year over year. Adobe Analytics measured almost exactly the same on the other side of the border: 393% for US retail in the first quarter of 2026. Numbers like that make any ecommerce owner open the dashboard smiling.

Now the part that never makes the headline: only 3% to 5% of those visits end in a purchase.

Translated into cash register language: AI is sending you crowds and your store is losing them. This is not a traffic problem. It is that the traffic no longer behaves like the visitor your site was built to serve. ChatGPT sits at roughly 900 million weekly active users and handles around 50 million shopping-related queries a day. That is not an emerging channel. It is a shopping mall that opened across the street, and you never put up a sign its doorman can read.

What agentic commerce actually is, minus the hype

Agentic commerce is what happens when the customer stops browsing and starts delegating. Instead of opening five tabs to compare, they tell an assistant: find me trail running shoes in a size 9, delivered before Friday, under a hundred dollars, with free returns. The agent goes out, reads, compares, filters and comes back with an answer. Increasingly, with the payment already lined up.

This is not science fiction or a conference demo. Amazon shipped Buy for Me. OpenAI released Instant Checkout and its Agentic Commerce Protocol. In January 2026 Google introduced the Universal Commerce Protocol alongside Shopify, Walmart and around twenty partners. Shopify turned on agent-readable storefronts by default. Gartner projects that by 2028, 90% of B2B purchasing will flow through AI agent-managed exchanges, worth more than fifteen trillion dollars in transactions.

Fully autonomous checkout, where the agent buys without you confirming anything, is expected to land around 2027. Which is to say: next year.

Your website stopped being a storefront and became an API

Here is the shift almost nobody has digested. An agent does not see your website. It is not moved by your typography, it does not admire the founder video, it does not grasp your value proposition written across a parallax banner. It reads data: product schema, price, real availability, delivery time, return policy, variants, weight, warranty.

The conversion decision happens in the data layer, not in the visual facade. If your product page says "fast shipping" instead of a concrete estimate, the agent cannot compare you. If your inventory refreshes every four hours, the agent discounts you as risky. If your sizes are written as M/L depending on the model, you simply do not exist to it.

Think of it this way: you have spent years paying for design and copy to seduce a buyer who now sends an accountant instead. And the accountant only looks at the table.

The protocol war already started and nobody asked your opinion

OpenAI is pushing the Agentic Commerce Protocol. Google, Shopify and Walmart are pushing the Universal Commerce Protocol. Payment processors are sprinting to wedge themselves into the middle of the transaction, because whoever controls agentic checkout controls the margin.

For a mid-sized merchant that means something uncomfortable: you will have to speak several doormen's languages at once, and none of them will guarantee you visibility. We are replaying the SEO story with one brutal difference. On Google you could compete with better content. Against an agent you compete with better data, and data does not reward creativity. It is either complete or it is not there.

What this breaks in marketing as we know it

First, your brand loses part of its shield. When an agent compares eight stores on price, delivery window and return policy, your brand premium has to be justified in a field it can actually read. If your differentiator is that you answer WhatsApp beautifully, the agent never sees it.

Second, the funnel compresses until it disappears. There is no abandoned cart to win back by email if there was never a visible cart. There is no remarketing for someone who never loaded your pixel.

Third, and this one stings: a large share of the ad budget that buys clicks today will end up buying presence inside answers. The question is no longer what you pay per visit, but whether your product makes the shortlist the agent returns.

Fourth, the market splits in two. Brands with clean data will absorb agentic traffic wholesale. Brands running hand-built catalogs out of a spreadsheet will watch their traffic climb while sales flatline. Exactly the 396% with a 3% conversion rate we opened with.

Latin America is not late, just comfortable

There is an optimistic read worth stating. Of Tiendanube's 180,000 Latin American stores, roughly 95,000 already use AI features actively. The region is not watching the train leave; it is on board, even if in the last car. Tool adoption is not the bottleneck.

The bottleneck is duller: catalog quality. A merchant using AI to write product descriptions while stock levels stay stale has improved nothing in an agent's eyes. They automated the facade and left the plumbing untouched. That is this industry's classic trap: we adopt what is visible and postpone what holds everything up.

Three objections you will hear, and how to answer them

"My customers will never buy that way." The same was said about mobile payments and grocery delivery. The question is not whether your current customer changes, but what the customer arriving in 2027 does, having never known another way to search.

"I sell an emotional product, AI does not get it." Half true. The agent will not pick your perfume, but it does build the list of five perfumes your customer chooses from. If you are not on the list, your emotion never reaches the table.

"This is a big-brand problem." The opposite. A small, well-structured catalog can be cleaned up in a week. Twenty thousand badly tagged SKUs is a year-long project. For once, being small is an advantage.

What to do this week, concretely

• Audit your Schema.org Product markup: name, brand, GTIN or SKU, price, currency, availability, condition and return policy. Complete, not half filled.
• Turn every vague promise into a data point. "Fast shipping" becomes "estimated delivery in 48 business hours." "Great returns" becomes "30 days, free, prepaid label."
• Sync inventory in real time, or as close as you can get. An agent penalizes uncertainty faster than it penalizes a high price.
• Normalize variants and sizing with one standard across the whole catalog.
• Check that your robots.txt and bot rules are not blocking AI assistant crawlers. Plenty of stores are locking their own door.
• Track referral traffic from assistants as its own channel. If it keeps landing in "direct," you are flying blind.
• Write product pages that answer real buying questions instead of padding paragraphs with a repeated keyword. Agents extract answers, not adjectives.

The uncomfortable opinion

The internet is about to fill up with gurus selling an "AI agent optimization" course with a new name and an expensive price tag. Ignore them. This is not a new discipline. It is data hygiene you have been postponing for a decade because nobody punished you for it. Now there is a punishment, and it is automatic, silent and never notifies you. The agent simply leaves you out, and you never learn about the sale that did not happen.

It is also worth cooling the hype. That 3% to 5% conversion rate proves the final decision is still human in most cases. There is still a window. But the window is exactly as wide as the time it takes you to get your catalog in order, and autonomous checkout is peeking around the corner in 2027.

Agentic commerce is not the end of ecommerce. It is the end of ecommerce built on a pretty facade over a messy backend.

The question is no longer whether your customers will use agents to buy. It is whether your store will be readable when they do. And you write that answer this week, in a data file nobody will ever look at and on which everything will depend.