Agentic commerce: what changes when an AI assistant, not a human, is the one buying from you

For twenty years every ecommerce decision assumed a human eye at the end of the funnel: hero imagery, urgency badges, review carousels, a checkout tuned to reduce hesitation. Increasingly the thing evaluating your product page has no eye, no hesitation and no patience for a cookie banner. It reads structured data, compares against four competitors, and reports one recommendation back to its owner.

Consumers are not abandoning shopping - they are delegating the tedious parts of it. "Find me a mid-range espresso machine under £400 with a stainless boiler, in stock, delivered by Friday" is a task people now hand to an assistant rather than perform across nine tabs. The assistant returns a shortlist of two or three, with reasons.

That single behaviour change moves the competitive battleground. You are no longer optimising to win attention on a results page; you are optimising to survive a machine-executed filter. If your stock status is wrong, your specification incomplete or your delivery promise unparseable, you are eliminated before a human ever sees your brand.

Strip away the marketing and the flow is consistent across assistants: interpret the request into constraints, gather candidates from feeds, indexes and retail APIs, verify the hard constraints - price, availability, delivery window, returns - then either present a shortlist or, where the user has authorised it, complete the purchase.

Three of those four steps consume structured data rather than rendered pages. The verification step is where most merchants lose: an agent that cannot confirm stock or delivery from a machine-readable source will discard the candidate rather than gamble on it. Silence reads as a "no".

Practically, this means investing in the least glamorous part of your operation. What we audit first:

• Complete Product and Offer schema on every PDP: price, currency, availability, condition, GTIN/MPN, shipping details and return policy - not just name and image.
• Real attributes, not prose. "Stainless steel boiler, 1.4 L, 15 bar" as typed fields beats a beautiful paragraph an agent has to guess at.
• Freshness. Price and stock accurate within minutes, not overnight. Stale feeds get merchants quietly downranked.
• One source of truth. Site, feed and API must agree. Contradiction is treated as unreliability.
• Machine-readable policies. Returns, warranty and delivery windows expressed as data, because they are frequently the deciding constraint.

This is the same discipline behind generative engine optimization - clarity, structure and verifiable claims - applied to a catalogue instead of a content library.

Agents justify recommendations. They favour products where the evidence for a claim is easy to cite: specifications from the merchant, corroboration from independent reviews, consistent pricing across sources. Vague superlatives do nothing; a specification table does a great deal.

Three moves that consistently help. Answer the comparison question directly on the page - the honest "who this is not for" paragraph gets quoted more than any feature list. Keep third-party listings and marketplaces consistent with your own data, because contradiction across sources is a downranking signal. And make sure your crawl and API surfaces are actually reachable: aggressive bot blocking now removes you from assistant-mediated demand, so decide deliberately which agents you allow.

Delegated purchase is where readiness gets tested. The emerging pattern involves scoped, single-use payment credentials, an explicit mandate describing what the agent may spend and on what, and a merchant endpoint that can accept an order programmatically with the same tax, shipping and fraud logic as the web checkout.

1. 01
   Expose a clean order API

   Cart, quote, tax, shipping options and order creation as first-class endpoints - not a headless wrapper around a form your storefront happens to submit.
2. 02
   Accept scoped credentials

   Work with your PSP on agent and delegated-payment support, including per-transaction limits and merchant-category scoping.
3. 03
   Return machine-readable confirmations

   Order ID, itemised totals, delivery window and cancellation terms in a structured response the agent can hand back to its user.
4. 04
   Keep a human off-ramp

   Any exception - partial stock, price change, address problem - must degrade to a human-resolvable link rather than a silent failure.

Under the hood this is ordinary integration work, and the interoperability layer is converging on the same tool-calling standards described in our piece on Model Context Protocol in the enterprise.

Your fraud stack was trained on human signals: mouse movement, typing cadence, session history. A legitimate agent looks exactly like a bot under those rules, and a malicious agent looks exactly like a legitimate one. Rebuilding this is the genuinely hard part.

What works today: treat agent traffic as its own channel with its own risk model rather than forcing it through human heuristics; require verifiable mandates so you can prove the shopper authorised the spend; log the full decision trail for disputes; and set conservative per-mandate limits until you have volume data. Retain the evidence - chargeback processes are still catching up, and the merchant with a clean audit trail wins.

There is a privacy dimension too. Agents transmit user preferences and history to complete a task; what you retain from that is regulated data. Fold it into the same consent and retention policy that governs the rest of your security and compliance posture.

Agent-mediated demand often arrives with no referrer, no campaign parameter and no session to stitch. Last-click attribution quietly reclassifies it as direct traffic, and the channel that is growing fastest looks like it does not exist.

Fix the measurement before you judge the channel. Identify agent traffic explicitly in analytics, watch assisted signals - branded search volume, repeat direct purchase of catalogue items you never promote - and run holdout tests on feed quality improvements rather than trying to attribute individual orders. Report the channel as a share of revenue, not as a click-based funnel, and revisit the definition quarterly.

We run agentic-commerce readiness as a scoped engagement for retailers and B2B sellers across the US, UK, EU and APAC: a data and schema audit, feed and API remediation, agent-ready checkout integration with your PSP, and the measurement model to prove it moved revenue. It builds on our digital marketing and AI agent engineering practices, so strategy and implementation sit with one team.

If you want a blunt assessment of how your catalogue currently reads to an assistant, email info@ranwebs.com or use the contact page. First call is free and goes to a senior consultant.