Agentic traffic: when AI agents shop your store

For thirty years the unit of ecommerce has been a person with a browser. They arrive, they squint, they hesitate, they compare, and somewhere in the next few minutes they either buy or leave. Every tool you run, every funnel you have ever drawn, assumes that shape of visitor. Agentic traffic is the first thing in a long time that doesn't fit it: a visit made not by a person but by an AI agent, acting on someone's behalf, with a goal and very little patience.

An agent is software that takes an instruction in plain language and carries it out across the open web: "find me a waterproof shell under £150 with good reviews and add it to a cart." The agent opens pages, reads them, makes decisions, fills forms, and increasingly completes the purchase. The shopper never sees your storefront. The agent does.

That is the important part. Agentic commerce doesn't replace the human's intent. It replaces the human's attention. The decision still belongs to a person. The browsing, the comparing, and increasingly the checkout belong to a model. Your store is no longer being read by the buyer. It is being read by a reader the buyer trusts to read on their behalf.

You can already watch this happen at the edges. AI assistants that operate a browser, shopping features bolted onto chat products, research agents that fan out across a dozen tabs to assemble a recommendation. None of it is mainstream yet. All of it points the same direction.

The reason agentic traffic matters operationally, and not just philosophically, is that AI shopping agents leave a different behavioural signature than people do, and behaviour is what most modern storefront intelligence is built to read.

A person scrolls unevenly. They dwell on an image, drift, come back, hover over a price, abandon a tab, return an hour later. An agent moves with intent and almost no waste: straight to the spec, the price, the returns policy, the structured data, then a decision. It does not get distracted by your hero video because it cannot be distracted. It does not respond to urgency banners because it is not anxious.

That is a problem if your stack reasons about who a visitor is from clicks and cookies, and an opportunity if it reasons from behaviour. An agent's behaviour is legible precisely because it is so purposeful. It just isn't the behaviour your dashboards were tuned for.

We have written before about the four seconds before the click: the narrow window in which a human storefront visit is effectively decided. Agents compress that window toward zero. There is no deliberation to influence, no scroll to reward, no second session to win back. The agent reads what is machine-readable, scores it, and moves.

This is the uncomfortable inversion of agentic traffic: the work you do to charm a human, the motion, the copy that unfolds, the social proof that builds, is invisible to the visitor who increasingly decides. What an agent rewards is clarity that is present immediately and structured well enough to be parsed: clean specifications, honest returns terms, legible pricing, real answers to the question actually being asked.

Agentic traffic breaks the three load-bearing assumptions of conventional web analytics at once.

Cookies assume a returning browser you can recognise. Agents are frequently fresh, sandboxed, or shared, so the cookie either isn't there or doesn't mean what you think. Sessions assume a continuous human arc with a beginning and an end. An agent's visit is a burst, sometimes one request, sometimes a hundred in parallel. Identity assumes the visitor is the buyer. Under agentic traffic the visitor is a proxy, and the person it represents may never load a single page.

The storefronts that cope are the ones that already stopped depending on those three things. Building a per-store behavioural model that profiles shoppers from behaviour, not identity was, when we made the choice, an argument about privacy and cookie deprecation. Agentic traffic turns it into an argument about basic legibility: when the visitor has no stable identity by design, the only thing left to read is what it does.

There is a second-order issue the ecommerce conversation hasn't caught up to yet, and it is a security one. An agent acts with delegated authority. It holds the keys, or the session, or the payment method, of the person who sent it. The moment software acts on behalf of someone with more privilege than itself, you have walked straight into a forty-year-old class of bug: the confused deputy, a trusted intermediary tricked into misusing authority it was only ever lent.

Storefronts and the tools around them are about to become deputies at scale, parsing untrusted page content and acting on it with a buyer's credentials. The question of which agent is allowed to do what, on whose behalf, with what proof is not a 2027 problem. It is the design problem hiding underneath agentic commerce, and most of the stack is being built as if it doesn't exist.