Beyond Checkout: A Guide to Agentic Commerce

1. The premise, stated plainly
2. Three layers, not one trend
3. How an agent transaction actually runs
4. The payment problem and how it got solved
5. The standards race and what each side is optimising for
6. Which categories move first, and which do not
7. What changes for a Shopify merchant, concretely
8. Discovery: how a model decides what to put in front of a buyer
9. The data that decides whether you are included
10. Pricing and promotions when the buyer is software
11. Rebuilding measurement
12. Operations: fulfilment, returns and support
13. The risk register
14. The economics, and the take rate nobody has set yet
15. A ninety day plan
16. Signals worth watching
17. What is genuinely unresolved

1. The premise, stated plainly

For twenty years the checkout page was the centre of gravity in ecommerce. Everything upstream existed to push a shopper toward it and everything downstream depended on it firing correctly. Conversion optimisation, cart abandonment flows, express wallets, address autofill, one page checkout: all of it was engineering work aimed at a single screen.

That screen is now optional.

A growing share of purchases will be completed by software acting on a shopper’s behalf, inside a surface the merchant does not own. A shopper asks an assistant for a black leather weekender under a set budget, delivered before a set date, from a brand with a returns window longer than fourteen days. The assistant assembles a shortlist, verifies stock and price against live merchant data, presents two options, takes approval, and completes the purchase. The merchant fulfils the order, keeps the margin, and never renders a product page or a checkout page to that customer.

The important consequence is not that a page disappears. It is that the merchant loses control of the moment of persuasion and gains a new dependency: being legible to a machine. Legibility is now a commercial asset in the way page speed became one in 2015 and mobile layout became one in 2013.

2. Three layers, not one trend

Most confusion in this conversation comes from treating one label as one thing. There are three distinct layers moving at different speeds, and a merchant’s exposure to each is different.

Layer one: discovery agents. An assistant answers a shopping question and recommends products, but the shopper still clicks through and buys on the merchant’s own site. This is already at meaningful volume. The merchant keeps the full experience and loses only the top of funnel. The risk here is exclusion, not disintermediation.

Layer two: transacting agents. The assistant completes the purchase inside its own surface using the merchant’s commerce backend. The merchant keeps the order, the margin, the customer data and the merchant of record position, but loses the product page and the checkout. This is the layer that shipped in 2025 and 2026 and is where the protocol work sits.

Layer three: autonomous buying agents. Software that buys without a human in the loop for each purchase, operating inside a budget and a set of rules. Replenishment, procurement, and machine to machine purchases of compute, data and API access. This is early, and the payment rails for it look different from consumer rails because there is no cardholder to authenticate at the moment of purchase.

A store selling consumables is exposed to all three. A store selling engagement rings is realistically exposed to the first for years.

3. How an agent transaction actually runs

The mechanics matter, because most of the operational readiness work follows directly from them.

Step one, the shopper states intent with constraints. Budget, timing, size, colour, material, delivery window, sometimes a brand preference or exclusion. Constraints are the raw material of the whole flow. Anything the merchant does not publish as data cannot be matched against a constraint.

Step two, the agent assembles candidates. From a product feed, from indexed pages, from a retail search index, or from all three. This is where inclusion or exclusion is decided, and it happens before any commercial logic runs.

Step three, the agent verifies. It calls the merchant to confirm the item is genuinely in stock, at the stated price, deliverable to the stated address inside the stated window, with the stated return terms. A merchant who cannot answer this call in real time is dropped in favour of one who can, because the agent is optimising to avoid a failed order.

Step four, the agent presents and the shopper approves. Approval is the consent event. Everything about liability and dispute handling later depends on how cleanly this moment was captured.

Step five, payment is authorised. A payment credential scoped to this merchant and this purchase is issued and passed to the merchant.

Step six, the merchant accepts the order. The merchant runs its own fraud checks, its own tax calculation, its own inventory decrement, and either confirms or rejects. This is the point that preserves merchant control. The agent proposes and the merchant disposes.

Step seven, fulfilment and post purchase. Confirmation, tracking, delivery, returns. Some of this flows back through the agent surface and some goes directly to the shopper. How it is split is still being worked out and differs by implementation.

The design intent throughout is that the merchant remains the merchant of record. Their pricing, their fraud rules, their inventory truth, their customer relationship, their payment processing. The agent is a new front door, not a new landlord.

4. The payment problem and how it got solved

Agentic payments had two problems that had to be solved together.

The first is credential exposure. Nobody wants an assistant holding raw card details, and no merchant wants to accept a card presented by software with no cardholder present. The second is intent. An issuer approving a transaction wants evidence that a human actually asked for this, because a purchase made by a model with no evidence of consent is a chargeback waiting to happen.

The answer that emerged is a payment token scoped tightly to a single merchant and a single purchase, issued only after the shopper approves, and worthless to anyone else who obtains it. The merchant receives something that behaves like a normal payment, processed through their existing processor with their existing rules. The network receives a trail back to a specific act of consent.

The parallel answer from the card networks is registration plus rules. The agent is registered, bound to a cardholder, and constrained at the network level by spending limits, merchant categories and transaction counts. The shopper’s instructions travel with the transaction as signed evidence, so a dispute can be resolved by inspecting what the shopper actually authorised.

Both approaches converge on the same principle: consent has to become data that travels with the payment. A merchant does not need to build this, but does need to understand it, because it determines what evidence they will have when an order is disputed.

5. The standards race and what each side is optimising for

Several overlapping specifications are live at once. It is more useful to understand what each party wants than to memorise the acronyms.

The assistant platforms want the shortest possible path from question to completed purchase, with no redirect. Every redirect loses users. They optimise for a protocol that merchants can adopt quickly with minimal integration.

The payment processors want to remain in the flow and to keep the merchant of record with the merchant. Their proposals preserve existing processing relationships and add a token layer on top.

The card networks want the transaction to stay on card rails and want to own the trust and identity layer for agents. Their proposals put registration, mandates and rules at the network level.

The platforms with their own commerce stack want agent traffic to arrive at their own checkout rather than at a competitor’s, because checkout is where their economics live.

The crypto and stablecoin protocols are solving a different problem: payments between software with no human present, in small amounts, at high frequency, across borders. This is not consumer retail and should not be evaluated as if it were.

None of these has won. Supporting more than one at a time is the correct hedge, and it is cheap if the underlying product data is clean, because most of the integration cost is data quality rather than code.

6. Which categories move first, and which do not

Agent mediated buying works best when the purchase can be fully described in words and numbers, and worst when the shopper needs to look, touch or feel confident about taste.

Moves early: consumables and replenishment, printer ink, filters, supplements, pet food, cables and adapters, batteries, standardised components, office supplies, anything with a part number, anything the shopper has bought before, anything where the decision is price and delivery date.

Moves in the middle: small appliances, tools, tyres, electronics accessories, fitness equipment, baby and childcare basics, standardised apparel like plain tees, socks and base layers where fit is predictable and the brand is known to the shopper.

Moves late or partially: fashion, jewellery, furniture, beauty where shade matching matters, anything bought visually or emotionally, anything where the shopper wants to browse rather than acquire. In these categories the agent’s realistic role is discovery and shortlisting, with the purchase still happening on the merchant’s own site.

The practical read for a merchant in a late moving category is that layer two adoption is not the near term threat. Layer one is. Being absent from the shortlist an assistant produces costs real revenue today, regardless of where the transaction eventually completes.

7. What changes for a Shopify merchant, concretely

For a store on Shopify, most of the readiness work is not exotic.

Metafields become commercially load bearing. Material composition, care instructions, dimensions with units, weight, country of origin, warranty length, compatibility, certification. Attributes that currently live inside a description paragraph or a design section need to exist as discrete fields, because that is what gets matched against a constraint.

Variant naming has to be disciplined. Colour names that are poetic are a liability. If a variant is called Midnight and no field anywhere says black, a constraint asking for black may not match it. Keep the poetic name for display and add the plain attribute as data.

Anything rendered only after JavaScript runs is at risk. Price, stock status, variant selection, reviews and delivery estimates that appear client side after load may not be seen. Server render everything that describes the product or its availability.

Apps that inject content are a hidden dependency. Review apps, size chart apps, bundle apps and upsell apps often render into the page late, or into an iframe, or into a shadow DOM. The content exists for a human and does not exist for a machine. Audit each one for whether its data is also available in markup or via a feed.

Bundles and kits must exist as sellable data. If a bundle only exists as a section on a template, it is invisible. If it exists as a product with its own identifier, price and stock, it can be recommended and bought.

The theme is no longer the whole storefront. Design work still matters enormously for the traffic that arrives on site, and that will remain the majority for years in most categories. But the feed is now a second storefront with its own quality standards, and it needs a maintenance owner.

Robots and bot management need a commercial decision, not just a security one. Aggressive blocking of automated traffic now has a revenue consequence. This is a decision for the business, not a default set by a security app.

8. Discovery: how a model decides what to put in front of a buyer

There is no ranking algorithm in the familiar sense to reverse engineer. What is observable is that models assemble answers from sources they can retrieve, parse and attribute, and that they favour sources where a specific claim can be lifted cleanly and stated with confidence.

That has a few practical consequences.

Specificity beats persuasion. A page that states dimensions, materials, capacity, compatibility and care in plain declarative sentences is more extractable than one built on atmosphere. Keep the atmosphere for humans and add the specifics as clearly stated facts.

Answer the question in the format the question is asked. Comparison questions, best for questions, compatibility questions and sizing questions are the shapes shopping queries take. Content organised around those shapes gets cited. Content organised around brand storytelling does not, unless the brand is the question.

Third party corroboration matters more than it did. A claim that appears only on the brand’s own site is weaker evidence than one that appears in reviews, retailer listings, publications and comparison sites. Distribution of accurate product information beyond your own domain is now part of the discovery job.

Consistency across sources is a ranking factor in practice. If the specification on your site, your marketplace listing and your feed disagree, the model has no reliable answer and is likely to prefer a competitor whose sources agree with each other.

Freshness of availability is decisive at the final step. Discovery may include you, but a stale stock signal removes you at verification.

9. The data that decides whether you are included

This is the least glamorous section and the highest return one. The fields that repeatedly decide inclusion or exclusion:

• Product identifier that is stable and unique, plus GTIN or MPN where one exists
• Title that describes the product rather than the marketing angle
• Brand, stated explicitly, not implied by the domain
• Category, mapped to a recognised taxonomy rather than only your own navigation
• Price with currency, and the actual price including any active discount
• Availability, accurate at the time of the call, with a fallback of a realistic restock date
• Variant attributes as structured values: size with a unit system, colour with a plain colour word, capacity, length, material
• Dimensions and weight with units
• Delivery estimate expressed as a window, ideally by destination
• Return window, return cost, and who pays for return shipping
• Warranty period
• Images with clear primary and alternate roles
• Reviews as structured data with count and average, not only as rendered stars
• Compatibility or fitment data where the category needs it

Two rules govern all of this. Accuracy beats completeness, because an incorrect field is worse than a missing one and produces failed orders and disputes. And consistency across every surface beats optimisation on any single one.

10. Pricing and promotions when the buyer is software

Agent mediated shopping is more price transparent than any channel that came before it, because comparison is free and instant and happens on every query.

Complex promotional mechanics degrade badly. Spend thresholds, tiered discounts, gift with purchase, code stacking and cart level logic are difficult to represent in a feed and difficult for an agent to evaluate. Simple, machine readable pricing wins inclusion. Complexity survives on the owned storefront where a human can be walked through it.

Landed cost is the real comparison unit. Item price plus shipping plus tax plus the cost of a return. A merchant with a slightly higher item price, free returns and a reliable two day window can beat a cheaper competitor on the comparison an agent actually makes. This is an opportunity for brands that compete on service rather than price.

Discount dependency becomes visible. If your conversion depends on a first order code delivered through a popup, that mechanic does not exist in this channel. The offer has to be in the price or in the terms.

Price consistency across channels stops being optional. Disagreement between your site, your feed and your marketplace listing reads as unreliability and gets penalised at verification.

11. Rebuilding measurement

An order arriving through an agent has no session, no referrer chain, no landing page and no click path. Reporting built on those signals degrades quietly rather than obviously, which is the dangerous kind of degradation.

Tag order source from day one, even when volume is negligible. Without a baseline you cannot later distinguish growth from cannibalisation, and cannibalisation is the likeliest early pattern: existing customers buying the same things through a new door.

Move toward incrementality. Holdout tests and geographic tests answer the question that attribution modelling can no longer answer.

Treat first party order data as the source of truth. Platform reported conversions will increasingly describe a shrinking slice of reality.

Watch verification failure rates. If agents are calling and getting stale stock or price mismatches, that is a measurable operational defect with direct revenue consequence, and it will not appear in any standard report unless you build it.

Expect customer data to be thinner. Email may be shared, marketing consent often is not, and post purchase flows that assume a full customer record need a degraded path.

12. Operations: fulfilment, returns and support

Delivery promises become contractual in effect. A shopper who chose you because the agent said it would arrive Thursday has a specific expectation created on your behalf. Delivery window accuracy moves from a marketing nicety to a retention factor.

Returns get more frequent and more mechanical in agent driven categories, because the shopper never saw the product. The returns process needs to be cheap to run and clearly stated, because return terms are now a selection criterion rather than an afterthought.

Support gets a new class of ticket: the order the customer does not fully recognise, placed on their behalf, possibly with a variant they did not consciously choose. Agents should be trained on this scenario and given the ability to look up an order without the usual session context.

Cancellation windows matter more. The gap between agent approval and merchant acceptance is where the shopper changes their mind. A clean cancellation path before fulfilment reduces both cost and dispute volume.

13. The risk register

Margin compression through price transparency. The most likely medium term effect in commodity categories.

Cannibalisation of owned traffic. Existing customers who would have bought directly now buy through an intermediary, and the merchant loses the upsell and the data while keeping the cost.

Disputes on consent. The mandate and token designs exist to answer this, but the case law and the scheme rules are immature. Expect a messy period.

Fraud shifts shape. Automated purchasing at scale makes some existing fraud patterns cheaper to run. Fraud rules tuned for human behaviour need review.

Brand safety in the comparison. Your product will be summarised by software in a sentence you did not write, next to competitors you did not choose. The only lever is the accuracy and clarity of the information available to be summarised.

Data leakage through the feed. A public feed exposes pricing, stock levels and assortment to competitors as well as to agents.

Platform dependency. Building deeply against one assistant’s implementation creates the same concentration risk that heavy dependence on one advertising platform created. Adopt the protocol, avoid the lock in.

14. The economics, and the take rate nobody has set yet

The current arrangement, where the merchant keeps essentially the full margin and the assistant takes little or nothing, is a customer acquisition phase. It is not a permanent state and should not be modelled as one.

The historical pattern for any intermediary that reliably delivers demand is: free, then paid placement, then a commission, then both. Marketplaces, comparison engines, delivery platforms and app stores all followed some version of this arc. There is no obvious reason assistants will be different once volume is proven.

The practical implication is that a merchant should treat agent driven revenue at today’s economics as a temporary advantage rather than a permanent structure, and should keep building the direct relationship. First order through an agent, second order direct, is a good target pattern and needs a deliberate post purchase plan to achieve.

15. A ninety day plan

Weeks one to three: audit. Feed accuracy against the live storefront. Structured data validation across templates. Identify every element that renders only after JavaScript. Check price, stock, delivery and returns consistency across every surface you publish to. Document current bot and crawler policy.

Weeks four to seven: fix the data. Move key attributes out of description prose and into structured fields. Normalise variant naming. Add plain colour and size values alongside display names. Publish return and shipping terms as clear structured content on stable URLs. Convert bundles into sellable products. Fix the top failures found in the audit.

Weeks eight to ten: instrument. Tag order source. Build a simple report on verification calls and their outcomes. Set a baseline for direct versus assistant referred revenue. Establish a holdout for later incrementality work.

Weeks eleven to thirteen: position. Decide the crawler and agent access policy as a commercial decision. Publish comparison, compatibility and sizing content in the shapes shopping questions actually take. Distribute accurate product information to third party surfaces so corroboration exists. Review pricing and promotion mechanics for machine legibility.

None of this is wasted if agentic volume stays small, which is the main argument for doing it now. Every item on the list also improves conventional search performance, marketplace performance, paid shopping performance and internal site search.

16. Signals worth watching

• Whether the assistant platforms begin charging merchants for placement or completion
• Whether a single protocol consolidates or the fragmentation persists
• What share of agent orders come from existing customers versus new ones
• Whether return rates in agent driven orders settle materially above direct orders
• How disputes over agent authorised purchases are ruled on
• Whether category coverage expands from consumables into considered purchases
• Whether merchants begin to see measurable cannibalisation rather than incremental growth

17. What is genuinely unresolved

Nobody knows what share of purchases will end up agent mediated. Credible estimates span a wide range, and most of the confident numbers in circulation are extrapolations from very small bases.

Nobody knows how the value split settles between merchant, processor, network and assistant.

Nobody knows whether shoppers will delegate purchases beyond replenishment at scale. The behavioural question is larger than the technical one, and the technical one is now mostly solved.

What is not in doubt is the direction of the requirement. Whatever share of commerce runs through agents, the stores that win it will be the ones with an unambiguous product truth: accurate data, honest specifications, real reviews, reliable delivery, and policies a machine can parse and a human can trust. That is closer to running a well kept catalogue than to running a funnel, and it is a different discipline from designing a persuasive product page.

The checkout page is not disappearing this year. It is going from being the destination to being one of several ways an order can arrive.