For twenty years, e-commerce optimisation meant persuading a person: better photography, sharper copy, trust badges, urgency. That work still matters. But a meaningful slice of your traffic is no longer a person — it is software shortlisting products on someone's behalf, and increasingly completing the purchase too.
Agentic checkout is now live across major search and assistant platforms, with "buy for me" flows executing orders on merchant sites directly. Estimates put AI-influenced orders at roughly a fifth of global volume through the last peak season. Whatever the exact number, the direction is settled: some of your customers are delegating the shopping.
An agent does what a diligent shopper does, only faster and without patience. It gathers candidate products, compares price, availability, delivery date and return terms, and then either hands the shopper a recommendation or completes the transaction through a commerce protocol. Standardised rails — from the agentic commerce protocols the large AI platforms have published — mean this is no longer scraping. It is a structured conversation with your store.
The consequence is blunt: an agent cannot be charmed. It cannot infer that "usually ships in a few days" means Tuesday. If your data is ambiguous or contradicts your page, it moves on to a merchant whose data is not.
The feed you set up years ago for shopping ads has quietly become a primary sales surface. It needs the same care as your product photography.
GTINs or MPNs, brand, precise variant attributes — size, colour, material, compatibility — plus weight and dimensions. Missing identifiers are the most common reason a product is invisible to comparison.
The price, stock status and shipping cost in your feed must match your product page and your checkout exactly. A discrepancy is not a rounding error to an agent; it is a reason to disqualify you. The same applies to your Product and Offer structured data — mark up price, currency, availability, shipping and return policy, and keep them synchronised with reality.
Publish the facts a comparison turns on: real delivery estimates rather than vague ranges, total landed cost including shipping and tax, return window and who pays for returns, and aggregated review data. Ambiguity reads as risk.
Agentic checkout struggles with exactly what human shoppers hate: forced account creation, multi-step wizards, surprise fees revealed at the last screen, and aggressive bot protection that blocks legitimate automated purchases. Clean guest checkout, standard payment rails and honest costs disclosed early make you completable — and every one of those changes also lifts conversion for the humans.
Optimising for AI agents is mostly just removing the friction you were already asking humans to tolerate.
Track it, too. Segment traffic and orders from assistant referrers so you can see agent-driven revenue as its own line rather than losing it inside direct.
Agents optimise within a consideration set they have been given. They rarely create desire. Brand preference, the reason someone asks for your product by name rather than "a good waterproof jacket under 200", still comes from content, social proof, community and design. The winning setup is a brand humans want and a data layer machines can act on — not one at the expense of the other.
Start with feed hygiene and schema accuracy this month, then remove friction from checkout, then instrument agent traffic so you can see it growing. None of it is exotic work. It is the same discipline good merchandising always required, applied to a buyer that reads your data instead of your headline.
We audit product feeds, structured data and checkout flows so your store converts both the people and the software doing the buying.
Talk to us