When Your Customer Is an Agent: What Changes for Retailers When AI Does the Shopping

Most ecommerce sites are built to persuade a human eye: the photography, the layout, the well-placed urgency message, the checkout flow tuned over years of A/B tests. The visitors now arriving in growing numbers don’t have eyes. They read data, not design, and they decide based on whatever you’ve made machine-legible. If the only thing they can read is price, price is how you lose.

That first problem is unglamorous and very solvable; but there’s a second, more interesting problem behind it. The emerging agentic commerce standards all push stores towards becoming clean, structured, transactable endpoints, with the customer relationship owned by whoever runs the agent. There’s an alternative shape to that future, and what retailers build now will influence which one wins.

We’ve already looked at what’s real in agentic commerce in 2026 and the protocols moving it from lab to production. This piece is about the storefront itself: what an agent actually sees when it arrives, and what that means for who owns the customer.

It’s already happening

Adobe’s analytics, drawn from its customers’ sites, tell a consistent story.

UK: AI-sourced traffic to retail sites grew 329% year-on-year over the 2025 Christmas period (November and December), and AI referrals on Boxing Day were up 685%.

US: AI-sourced retail traffic is up 1,324% since Adobe started tracking in October 2024, with travel up 2,215%.

Conversion: In May 2026, AI-referred traffic to US retail sites converted 54% better than non-AI traffic, a reversal from a year earlier.

That last figure is worth taking note of, because these are higher-intent visits. The usual caveat applies: big percentages from a small base. The direction of travel is the point, not the multiple.

Getting found is the familiar half

The scramble to appear in AI answers — GEO, AEO, AIEO, take your pick — is substantially SEO in a new suit. For now, it mostly means ranking in whichever search index sits behind the assistant and being present in the directories and roundups it reads. It’s not a dark art, and it isn’t where most retailers are losing.

The actual visit is where most stores fail

Once an agent shortlists you, it comes to check the facts that decide the sale, and it arrives with a very limited browser. It won’t spin up Chrome, run your JavaScript or admire your photography.

We tried it: a pair of trail shoes, a specific size, next-day delivery. Several sites handed the agent hundreds of lines of navigation markup and no product data at all, and not one exposed the fact that actually decided the purchase: whether it could arrive the next day. That sat behind the checkout.

There’s a baseline every agent will want: size, price, availability and delivery. Beyond that, it may be weighing criteria you never built a field for, such as material, colour, provenance, shipping method or compatibility. If it only exists in the photograph, it doesn’t exist.

That creates two failure modes, both expensive:

• It moves on. A lost sale you never knew you were in the running for.
• It fills the gap itself. A hallucinated spec leads to a return, a chargeback or a bad review, and reviews are exactly what the next agent checks.

Rubbish in, rubbish out still applies. There’s already evidence that some retailers generate product spec data with AI and publish it unverified. Data quality is the foundation, and adding a PIM won’t fix it if the data going in is wrong.

None of this is hard. It’s incremental, it’s mostly hygiene, and it’s work the industry already knows how to do; it’s just pointed at a new reader – the smartest thing in the world, looking at your site with the dumbest possible tools. 

The harder part: who owns the vending machine?

The last twelve months have produced a raft of agentic commerce standards: UCP, ACP, AP2, A2A and Microsoft’s Copilot Checkout among them. We’ve unpacked how they fit together before, so we won’t repeat it here. What matters is that they differ in detail but all point in the same direction: give the agent everything it needs to transact.

Follow that to its end state and the store is reduced to a feed and an API. A vending machine. That’s fine, until you ask who owns the machine and who decides when it gets seen. In that model, the customer is talking to an agent operated by Google, OpenAI or Anthropic, and the retailer sees the transaction and nothing else. The relationship, and the data, sit with the agent’s owner.

That surface is already being monetised. Google’s Direct Offers, announced in January 2026 and currently piloting in the US, places paid promotions inside the AI shopping conversation. They’re drawn from the merchant’s own product feed, with Google’s AI deciding when they surface. Google’s own framing is “a salesperson negotiating a deal on your behalf.” It’s worth reading twice: whose salesperson, and on whose behalf?

It’s the same question we put to airlines in Why have an airline app in an agent-mediated world?: when someone else owns the front door, what’s left for you to own? You can safely assume the choices those agents make will serve their owners first, their users second and retailers a distant last.

The other path: personal agents

That future isn’t inevitable. Frontier model compute costs are rising sharply while small and locally run models get steadily more capable. Personal agents are already appearing, from OpenClaw and Hermes in the local space; Dots, Muse, Gemini Spark and Duet from the frontier labs. It’s a young space, but agents can now be truly personal and actually spend money – which is where it gets interesting.

An agent that genuinely belongs to the shopper behaves very differently. It carries persistent memory and explicit instructions from its owner: prefer this brand, never that one, pay up to 5% more to shop locally. Those aren’t preferences it optimises away; they’re constraints it’s bound to honour. We’ve looked at the buy side of agentic commerce separately, and the implication is the same: your sell-side strategy has to be built for those agents if you want to be more than a feed and an API.

For a retailer, that agent is both the best customer and the worst. It will ask for everything, check everything and remember everything, including what the customer decided to buy and what they said about it a week later. It can’t be bought with a sponsored slot or nudged with a dark pattern. For some retailers (Amazon for one), there are plenty of reasons to block them, but it’s also the only version where being genuinely good at the thing gets rewarded, and where an individual relationship with a customer is possible at all. 

Where to start

The real question for retailers is this: do you put the effort into being something an agent can genuinely interact with, or do you accept being a vending machine?

For anyone asking what to actually do, here’s the sequence:
1. Be readable. Server-rendered or properly structured data, clean schema markup, a current feed, and don’t block AI crawlers.
2. Answer the deciding questions. Variant-level stock, real deliverability, and the specific attributes that differentiate you in your category, structured rather than buried in prose.
3. Be honest. Expose constraints and limitations. A clean “no” costs a sale; a hopeful “yes” costs the relationship.
4. Let the protocols follow. They’ll largely arrive via your commerce platform, and they’re worthless over a catalogue that can’t answer the basics.

Almost all of this is engineering rather than marketing: legibility, structured data, feeds and entity clarity. It’s also a conversation most retailers haven’t had yet. Start with the data instead of the protocol.