At some point in the next year, a good customer will buy from you, or from your competitor, without ever seeing either website. Their assistant will take the brief, shortlist, compare, and check out. The forecasts vary in their enthusiasm: roughly one in five shoppers used an AI agent somewhere in the journey last year, one industry forecast has that near half by the end of this one, and Morgan Stanley expects agents to be touching a quarter of online spending by 2030. The precise numbers will be wrong. The direction won’t.
eCommerce has been here before, and that should be more comforting than it sounds.
The moat always moves upstream
Before Shopify, having a working online shop was a competitive advantage. Heritage retailers, catalogue businesses, anyone with the budget for a proper build held a structural edge over everyone who didn’t. By 2015 that moat had cracks. By 2020 it was a punchline. The advantage that survived sat in the layer above the shop: brand, product, taste, the customer relationship. When the cost of a task collapses, the advantage doesn’t disappear. It moves up a layer, to whatever the cost wall was protecting.
Agents collapse a different task: the shopping itself. Browsing, comparing, remembering, returning, checking out. The things conversion optimisation has spent two decades polishing are exactly the things an agent doesn’t do. It doesn’t linger on your homepage film. It can’t be retargeted. It won’t join the newsletter for ten per cent off. It reads what is machine-readable, your product data, availability, price, reviews, and what everyone else says about you, and it makes a recommendation.
Clarity becomes machine-readable
Here is the part I find genuinely interesting: what an agent can and cannot resolve. A brand that says the same thing on its product pages, in its reviews, in its press and in its returns policy is easy for a machine to represent accurately. A brand that says six different things in six places is noise, and machines don’t recommend noise. Consistency used to be a brand discipline, the sort of thing a marketing director defended in an annual workshop. It is becoming a distribution channel.
An agent can’t be charmed. It can only be informed. Which is, if you think about it, the customer brands always claimed they wanted.
The sensible objection is that the plumbing is a moving target, and it is. Checkout protocols, product feeds, whichever standard wins the current turf war: most technical advice written today will be stale within a year, and this piece is deliberately not a feed specification. But underneath the protocols, the question the agent asks is the question customers have always asked. Who is this for, and why this one? A brand that can answer in one sentence will translate cleanly to whichever standard wins. A brand that can’t will simply be replatforming its confusion.
So the preparation for agentic commerce looks surprisingly unlike a technology project. Tidy product data matters, but the tidy data is downstream of a tidier answer: what this brand is for, stated once, held everywhere. The brands that get chosen by machines will be the ones that were already clear enough to be chosen by people. The agents just remove everything that used to hide the difference.