TechmadeazyAgentic commerce is the phrase of the year. Every conference has a panel on it, every platform has an announcement, and most brand owners are somewhere between anxious and baffled about what they are supposed to do on Monday morning.
Here is the calm version.
Three things are getting mashed together
Almost every article on this subject blurs three separate stages. They sit on very different timelines and they need very different responses.
Discovery is an assistant finding a product and putting it in front of a buyer. This is live, at scale, in every major assistant, today.
Recommendation is the assistant picking one product over three others. Also live today, and this is where nearly all of the value currently sits.
Autonomous checkout is an agent completing the purchase without a human at the keyboard. This is the part that gets the headlines and it is the least settled. OpenAI deprecated its Instant Checkout feature in March 2026, and the current model across most assistants is discovery and recommendation followed by a redirect to the merchant's own site. The buyer still lands on the brand's page and still checks out there.
The short version: the money in agentic commerce right now is in stages one and two, which are the same stages AEO has always been about. Stage three is worth preparing for and not worth rebuilding a business around this quarter.
The buyers are not fully sold either
It is easy to read the platform announcements and assume the shift is done. Consumer research suggests otherwise. Checkout.com's 2026 study found 27% of consumers said they trust no organisation at all to operate an AI shopping agent, and 24% said they would never delegate a purchase to one.
At the same time, a third expected at least a tenth of their purchases to be AI-driven within a year. Both things are true. Demand is forming faster than trust, which means the near-term winner is the brand an assistant confidently recommends to a human, not the brand with the slickest agent integration.
Five things that make a product selectable
The useful part of the agentic shift is that it rewards work most brands should be doing anyway. An agent evaluating options does, quickly and joylessly, what a careful human researcher does slowly. It reads facts and compares them. If the facts are missing, it moves on. There is no persuasion layer to fall back on.
1. Product data completeness. Every blank attribute is a comparison the product cannot enter. If the shopper's constraint is life stage, or hair type, or room size, and that field is empty on the listing, the product is not in the shortlist regardless of how good it is. Attribute fill is the least glamorous work in this field and consistently the highest return.
2. Price and stock accuracy in real time. An agent that surfaces a product at a price the brand no longer charges, or in a size it no longer stocks, produces a bad experience that gets attributed to the brand. Feeds that update daily were fine for shopping ads. They are marginal here.
3. Delivery terms stated as text, not as a widget. Delivery speed is one of the top constraints buyers put into a prompt. "Delivered by Friday" is a filter. A shipping timeline that only appears at checkout, or only after a script runs, does not exist for the purposes of that filter.
4. Returns and warranty in plain language. These are trust signals and assistants use them heavily, particularly for a first purchase from an unfamiliar brand. A clear thirty day policy stated in a sentence beats a well-designed page that requires interpretation.
5. One version of every fact, everywhere. Site, marketplace listings, feed and profiles should agree. Where they disagree, an assistant either hedges or picks the least flattering version. This turns up in almost every audit, and it is usually the fastest thing to fix.
The protocols, briefly
Three acronyms are in circulation. ACP came from OpenAI and Stripe and focuses on the checkout handoff. UCP came from Google and Shopify and covers more of the journey. MCP, from Anthropic, is a general way for models to reach live data rather than a commerce standard as such.
On Shopify, a good deal of this is being abstracted away at platform level, and the sensible response is to watch the release notes rather than hire an integration team. On custom infrastructure, the question to put to a developer is not "which protocol should we adopt" but "can our catalogue, price and stock be read reliably by something that is not a browser".
What does not work
Rebuilding a storefront for autonomous checkout while the product data is still half filled in. Brands have spent a quarter on integration work and stayed invisible in answers, because the agent never got far enough to want to buy anything.
The subtler failure is treating agentic commerce as a reason to stop investing in the pages buyers read. A human still reads the answer. A human still lands on the product page. Optimising purely for machine consumption loses the part of the transaction that is still doing most of the work.
Where to start
- Export the product feed and count the blank fields. That number is the first target.
- Take the ten best sellers and read their pages as an agent would, with no images. Is the size, the material, the delivery time and the return window all there in words?
- Check that price and stock update at least as often as they change.
- Track whether assistants currently recommend the brand for the questions that matter in the category, before touching any infrastructure.
