From Storefronts to Agents: The Next Era of Commerce - Part 1

Most ecommerce conversations about AI right now are stuck on the wrong question. They are arguing about which large language model to plug into. The harder question, and the one that will decide the next decade of online retail, is who owns the agent that decides what every shopper sees, what they are recommended, and how they are sold to.

The dominant story about AI in retail is a story about discovery. Shoppers are asking ChatGPT what running shoes to buy. They are asking Gemini to plan a gift. They are asking Perplexity to compare a kitchen appliance against three alternatives. The numbers behind that shift are real, and they have moved faster than most retail roadmaps planned for. Generative AI traffic to retail sites grew dramatically through 2025. A meaningful share of consumers now prefer asking an AI assistant over typing into a search bar. The behaviour is no longer a forecast.

So retailers are responding. They are publishing apps inside ChatGPT and Claude. They are bolting chat widgets onto their product pages. They are commissioning vendors to build "AI shopping assistants" that float in the bottom right corner. Most of this activity is well-intentioned. Almost none of it is a commerce strategy.

The reason is a category error. The industry is treating agentic commerce as if it were a feature you ship, a widget you install, or a marketplace you list on. It is none of those things. It is a structural change to who interfaces with the customer at the moment of intent, and that change is now happening at three completely different layers of the stack at once. Retailers who do not separate those layers will keep investing in the wrong one.

A storefront is no longer a page someone reads. It is a surface that is built for them in the moment they arrive. The question is whether you are building it, or someone else is building it about you.

What Agentic Commerce Actually Is

Agentic commerce is not a chatbot. A chatbot answers a question and stops. Agentic commerce is a system in which an autonomous software agent stands between the shopper and the catalogue, and decides what the shopper sees, what they are asked, what they are recommended, and how they get from intent to fulfilled order. The shopper does not navigate a site that has been laid out in advance. They arrive at a surface the agent has built for them, drawing on whatever the retailer knows about that shopper, the live state of the catalogue, and the retailer's own merchandising rules and brand priorities.

The change is not cosmetic. Traditional ecommerce is reactive. The shopper does the work. They search, scroll, compare, filter, click into a product page, configure a variant, add to cart, fight through a checkout flow, enter shipping details, and confirm. The site is a passive surface, the same one for everyone. Every conversion is the result of a customer pushing themselves through a layout they did not ask for.

Agentic commerce inverts that load. The agent does the work. From the moment the shopper arrives, the agent is composing what they see: which products are surfaced, in what order, with what framing, alongside which recommendations, with which questions to ask. When the shopper does state intent explicitly, in words or voice or by uploading an image, the agent treats that as one signal among several. A request like "I need a humidifier under one hundred fifty dollars that can handle a six hundred square foot room and arrive before Saturday" is a structured query in disguise. A static storefront forces the shopper to translate that into a sequence of filter clicks and tab switches. An agentic storefront translates it directly into the surface the shopper is shown.

The technical preconditions for this have arrived. Language models are now reliable enough at structured extraction to turn fuzzy intent into clean criteria. They are also reliable enough to assemble layouts, write copy, sequence recommendations, and adapt tone without a human author for each variation. Open commerce protocols, payment authorisation standards built specifically for agent-driven transactions, and structured product feeds have started to standardise the channel between agent and merchant. Real-time inventory, shipping cost calculation, and tax and duty resolution can now be exposed through APIs that an agent can hit in the same session. None of this individually is new. What is new is that the pieces are now coordinated enough that a single agent can run the full loop, from generating the surface to closing the sale, in one visit.

Three Layers Under One Label

Most of the confusion in the market right now comes from the fact that "agentic commerce" is being used to describe three different things, sitting at three different layers of the stack, with three different ownership models. Treating them as one product is the source of most of the bad investment decisions retailers are making this year.

These layers are not mutually exclusive. A serious retailer will have a presence at all three. The strategic point is that they are not interchangeable, and the economics of each are completely different. Layer one and layer two have a long history in retail under different names. Layer three is the structural change. It is not a faster catalogue or a smarter widget. It is a storefront whose entire design is decided by the agent at the moment the shopper arrives.

Why Layer One Is a Channel, Not a Strategy

It is tempting to treat the AI giants as the new Google. That framing is half right, which makes it dangerous. They are like Google in that they are becoming the entry point to discovery for a growing share of shoppers. They are unlike Google in that they no longer hand the shopper off to the retailer's website at the point of intent. The conversation continues inside the AI surface. The shopper compares, evaluates, sometimes purchases, and forms a relationship with a product, all without ever loading the retailer's domain.

This produces a familiar problem in an unfamiliar form. Retailers spent the last fifteen years dependent on traffic from a small number of platforms they did not control. Search rankings, ad auctions, and social algorithms set the terms of access to their own customers. Agentic discovery on someone else's surface is the same dependency in a more extreme form. When the shopper is buying through a conversation that lives inside a third-party assistant, the merchant is competing on whatever criteria that assistant chooses to weight, with whatever brand voice that assistant chooses to render, alongside whatever competitors that assistant chooses to surface.

None of this means a retailer should ignore layer one. Discovery is moving there. Optimising for it, what some are now calling generative engine optimisation, matters. Structured product data, clear policies, fast and accurate fulfilment signals, and clean APIs all increase the probability that an AI assistant surfaces a retailer's catalogue when intent matches. But that work is the same logic as SEO. It is investing in being findable on a platform you do not own. It is not a substitute for owning the conversion.

Discovery on someone else's platform is not a commerce strategy. It is traffic dependency. Traffic dependency without ownership is exactly where retailers have been stuck for the last decade and a half.

Why Layer Two Solves the Wrong Problem

The chat widget is the most overdeployed answer in retail right now, and it is mostly solving the wrong problem. The pitch is reasonable on the surface: shoppers have questions, an LLM can answer them, so put an LLM in a bubble on the product page and conversion will improve. The reality is more constrained.

A widget sitting on a static page is still subordinate to that page. The shopper has already navigated through search, category, and product hierarchy to get there. The agent's job is reduced to answering follow-up questions. It cannot rearrange the storefront. It cannot reorder results. It cannot take the shopper from "I need something for a long flight" to a recommended pair of noise-cancelling headphones, with the right shipping speed, with a discount applied, with a checkout completed, in one conversation. It can only respond to what the shopper has already done on a layout that was designed for browsing, not for asking.

There is also a control problem. Most chat widgets are powered by third-party platforms. The retailer rents the agent rather than owning it. The agent's voice is generic. The recommendation logic is whatever the vendor's model produces. The conversational data, which is the most valuable byproduct of these interactions, often does not flow cleanly into the retailer's own data estate. Retailers end up paying for engagement metrics on a surface they cannot fully shape, while the most strategically useful artifact of the interaction, the structured intent log, lives somewhere else.

The diagnostic question is simple. If you removed the widget tomorrow, would the storefront's structure be any different? In most layer-two deployments, the answer is no. The site is still a catalogue. The widget is decoration on top of a catalogue. That is not agentic commerce. It is a customer service feature with a better interface.

The Architecture of Layer Three

A merchant-owned agentic storefront is a different architecture, not a different feature. The shopper does not arrive at a page. They arrive at a session. What they see, what they are asked, what is recommended, and how they are guided through to a purchase are all decisions made by the agent in the moment, against everything the retailer knows about that shopper and everything the retailer knows about its own catalogue and operations. Two shoppers landing at the same URL, one second apart, do not see the same surface. They see two different storefronts, each one personalised to what the agent infers each shopper is trying to do.

This is what people mean when they say the design itself is agentic. The retailer is not picking a layout, picking a hero product, picking a category order, picking a set of filters, and serving that arrangement to everyone. The retailer is supplying the agent with a brand voice, a catalogue, a set of merchandising rules, a set of constraints, and a set of business priorities. The agent assembles the storefront from those inputs, per shopper, per visit. The same shopper returning a week later sees something different again, because the agent has more context, and because the catalogue and the priorities have moved.

To work, this requires a stack that most retailers have only built in fragments. The agent needs live access to the catalogue, not a nightly snapshot. It needs real-time inventory visibility across whatever fulfilment network the retailer operates. It needs the customer's history, when they are signed in, and a clean way to operate without it when they are not. It needs the cross-border tax, duty, and shipping logic baked in so that a quoted price is the price the shopper actually pays. It needs the returns and exchanges flow on the same rails so the agent can answer "what if it doesn't fit" with a real policy, not a generic disclaimer. And it needs to be able to take a payment without bouncing the shopper to a separate checkout page.

A schematic of the working architecture looks like this:

shopper arrives (signed in, returning, or anonymous)
   → context resolution (history, location, prior intent, current session signal)
       → storefront generation (layout, hero, recommendations, prompts, all per-shopper)
       → agent voice + brand rules + merchandising priorities applied
   → conversation begins (text, voice, image, or implicit signal)
       → live catalogue query (in-stock, in-region, in-policy)
       → personalisation layer (history, preferences, prior sessions)
       → fulfilment + duty + tax resolution (real, not estimated)
       → recommendation set (ranked against stated and inferred constraints)
   → in-conversation configuration (size, variant, bundle, gift)
   → in-conversation checkout (auth, payment, address, confirmation)
       → fulfilment trigger
       → post-purchase agent (tracking, exchange, return, repurchase)
   → first-party intent log (every keystroke, every hesitation, every accepted recommendation)

The arrows are doing real work in that diagram. Each one is a system call that has to succeed in real time, in milliseconds, in the same session. That is why most retailers cannot stand this up by writing a prompt. The hard parts are not the language model. The hard parts are the integrations beneath it, and the discipline to let the agent generate the surface rather than overriding it with a fixed template.

The Data Asymmetry

One of the most underappreciated consequences of layer three is the data it produces. A traditional storefront generates click data: page views, time on page, add to cart, checkout completion. That is event data, and it is shallow. It tells you what the shopper did, not what they wanted. Most of the actual intent is lost to the gap between what the shopper had in mind and what the navigation made it possible for them to express.

An agentic storefront produces something different. Every visit is a structured signal. When the shopper speaks or types, they say, in their own words, what they are looking for, what matters to them, what they are worried about, what they have tried before, what they have ruled out. They ask follow-up questions that reveal which features they actually weight. They reject recommendations and explain why. They abandon halfway through and explain what is making them hesitate. And even when the shopper says nothing, the surface the agent generated for them is itself a record: the layout it chose, the products it surfaced, the prompts it offered, and which of those the shopper engaged with or ignored.

This is intent data, not engagement data, and it is qualitatively different. It is the closest thing online retail has ever had to the artefact a good in-store associate accumulates over a year of customer interactions: a working theory of what each customer segment is actually trying to solve, in their own language, with their own framing. No A/B test on a static site produces this. No survey produces it at scale. No third-party platform that intermediates the experience will hand it back to you in usable form.

Retailers who deploy layer three early will accumulate this asset for as long as they run the surface. Retailers who delay will not. The data is path-dependent. It is generated only when the agent runs the storefront on infrastructure the retailer owns, and it cannot be reconstructed retroactively from log files. This is the meaningful moat in agentic commerce, and it is the one that compounds.

Engagement data tells you what shoppers clicked. Intent data tells you what they meant. Only one of those is useful for training the next year of merchandising decisions.

The Conversion Mechanics That Change

Most retail funnels look the same. A large number of shoppers enter the top of the funnel through search, social, or direct traffic. The funnel narrows aggressively at category browsing, narrows again at product detail, narrows again at cart, and narrows hardest at checkout. Each transition is a place where shoppers leak. The standard playbook for the last fifteen years has been to optimise each transition incrementally: faster pages, clearer photography, better recommendations, fewer checkout fields, smarter abandoned cart emails.

An agentic storefront collapses several of those transitions into one. The shopper does not have to choose a category, then narrow on filters, then click into a product, then choose a variant, then navigate to cart, then fight through checkout. The agent has already done much of that work before the shopper types a word, by generating a surface that is biased toward what this shopper, in this context, is most likely trying to do. When the shopper does speak, the agent refines the surface instead of restarting from a generic catalogue. The configuration step is a follow-up question. The checkout step is a confirmation. The transitions where shoppers historically leak are no longer transitions in the same sense.

The other change is that the agent never gets tired or distracted. It does not give up on a hesitating shopper at minute four. It does not assume that someone who has not added to cart is uninterested. It can hold context across sessions, remember what was almost purchased last time, surface the same items at a lower price when they go on sale, and reopen the dialogue when something the shopper asked about comes back in stock. The closest human analogue is the personal shopper who knows you. Most ecommerce has never had that. Most agentic commerce will, by default.

Where This Applies and Where It Does Not

Agentic commerce is not equally valuable in every category. The categories where it changes the economics most are the ones where shoppers have questions a static storefront cannot answer well, or where the right product depends on context the static storefront has no way to know. Anywhere fit, sizing, suitability, or use case drives the purchase decision, an agent that can both generate a surface biased toward the right answer and ask one or two clarifying questions outperforms a filter-driven catalogue by a large margin. Anywhere the shopper is buying for a specific situation rather than a specific SKU, the same logic applies.

Consumer electronics is a clear fit. A shopper who needs a monitor for colour-accurate photo editing is poorly served by a grid of monitors sorted by price. They are well served by a storefront that already opens with the two or three options that match their workflow, with the tradeoffs explained, and an agent ready to refine further. The same logic applies to home appliances, where compatibility, capacity, and energy ratings matter, and where a wrong purchase is expensive to return.

Pet products, baby products, and health and wellness all have the same structural property: shoppers are anxious, the stakes feel high, and the decision is gated on questions the catalogue cannot answer. Food and beverage, especially in subscription replenishment, benefits from a different angle: the agent does not just configure the order, it remembers the order and re-runs it on cadence with adjustments.

Categories where the value is lower include true commodity purchases where the shopper already knows the exact SKU, has bought it before, and is optimising only on price. There, agentic generation adds friction rather than removing it, because there is nothing to personalise. Most retailers do not run pure commodity catalogues, but the segment exists, and it is reasonable to leave it on a standard checkout path.

The Operational Bar

The operational requirements of layer three are higher than most retailers initially expect. The agent only converts well when the underlying systems converge in real time, and the failure modes are visible to the customer in a way they were not on a static site. A nightly inventory snapshot is not enough; a shopper who orders something the agent has just promised is in stock, and then receives a backorder email two hours later, has had a worse experience than they would have had on a regular product page. A shipping estimate that turns out to be wrong, a tax line that surprises at checkout, a return policy that turns out to vary by region, all of these are minor flaws on a static site and are catastrophic on an agentic surface where the agent has just personally promised the shopper otherwise.

The agent is, in effect, a single accountable surface for the entire commerce operation. It cannot say one thing and the system do another. The implication is that retailers cannot deploy agentic commerce on top of fragmented infrastructure. The catalogue, the inventory, the tax engine, the shipping logic, the returns logic, and the payment authorisation all have to be reachable from one place, with consistent answers. Retailers who have spent the last decade stitching together best-of-breed point solutions often find that the integration tax is the actual project.

This is also why the rollout pattern that works for layer three is not the same as a typical site redesign. The retailers who are getting traction with this are not flipping the switch on their entire traffic at once. They are starting with a contained audience, often loyalty members or VIP cohorts, where the bar for the agent is high but the support model is manageable. They are routing a meaningful slice of paid traffic, often in the range of twenty to thirty percent, to the agentic surface specifically and measuring conversion against the standard site as a control. They are using the early data to refine the agent's behaviour, the underlying product information, and the recommendation logic before scaling. This is the right shape for the rollout because the agent improves with traffic, and the agent's mistakes are most expensive when the audience is least forgiving.

The Fourth Sales Channel

The most useful frame I have found for layer three is to treat it as a sales channel, not a feature. Most retailers think in terms of channels: wholesale, retail, direct ecommerce, marketplaces. Each channel has its own economics, its own operational requirements, its own merchandising logic, and its own measurement model. They are not interchangeable, and a serious retailer manages each one as a distinct line of business.

Agentic commerce is the fourth channel. Sell with AI. The infrastructure is different, the unit economics are different, the customer relationship is different, and the data flywheel is different. Treating it as a feature of the existing direct ecommerce channel underweights it. Retailers who set up dedicated ownership of the agentic channel, with its own targets, its own roadmap, and its own measurement framework, are the ones who get the most out of it. The retailers who treat it as an enhancement of the existing site tend to bolt on a chat widget and move on.

What Happens to the Static Site

The static site does not vanish overnight, but its role changes fundamentally. Some pages will continue to exist, primarily as machine-readable surfaces: structured product data for search engines to crawl, canonical URLs for direct linking, public pages for press, partners, and the AI giants in layer one to ingest. These are infrastructure, not the shopper experience. They exist so that the catalogue can be discovered, indexed, and surfaced in places the retailer does not control.

The shopper-facing surface is different. When a real customer arrives, the site they see is not the static catalogue. The agent intercepts the visit and assembles the storefront for that specific person. There is no canonical homepage to which all shoppers are sent. There is no fixed category page that everyone hitting the same URL sees identically. There is no shared product detail layout where every shopper, regardless of context, gets the same five photos in the same order with the same recommendations underneath. Each of those, when they appear at all, is generated for the shopper looking at it.

This is a real departure from twenty-five years of ecommerce design practice. The job of the merchandising team stops being "design the homepage." It becomes "define the rules and priorities the agent uses to design the homepage for each shopper." The job of the conversion optimisation team stops being "A/B test the product detail layout." It becomes "tune the agent's behaviour and the inputs it draws from." The artefact is no longer a layout. The artefact is a system that produces layouts.

Retailers who navigate this transition well will treat the remaining static pages as a publishing surface for crawlers and partners, and the agentic storefront as the surface for actual shoppers. The two will share a catalogue and share data, but they will not share a layout. Most current implementations conflate the two, which is one of the reasons layer-two chat widgets feel bolted on. They are bolted on, because the static site beneath them is still doing the work it was designed to do for an audience that no longer arrives in the same way.

What Retailers Should Be Doing Now

The window for first-mover advantage in layer three is open and will not stay that way for long. The retailers who treat this as a channel rather than a feature, who make the underlying integrations work rather than bolting an LLM onto a static site, and who start collecting structured intent data now will have an asset that is hard to replicate later.

The shape of the work, in roughly the order it should happen, is this:

1. Audit the underlying systems. The agent only works as well as the integrations beneath it. Inventory visibility, real-time pricing, fulfilment logic, returns policy, and payment authorisation all need to be reachable from a single layer with consistent answers.
2. Define the channel. Treat agentic commerce as its own line of business. Assign ownership. Set its own targets. Do not measure it as an enhancement of the existing site.
3. Start with a contained audience. Loyalty members, returning shoppers, or a paid traffic cohort are all reasonable starting points. Resist the urge to flip the agent on for the entire traffic before the behaviour is stable.
4. Invest in the data layer. The structured intent data the agent produces is more valuable than the immediate conversion lift. Capture it, store it, and route it into merchandising and product decisions.
5. Measure honestly. Conversion lift on the agent surface is a valid headline metric, but the deeper signals are time on site, repeat session rate, depth of intent captured per visit, and reduction in returns. These are the indicators that the agent is actually solving the problems shoppers had with the static site.
6. Optimise for layer one in parallel. The work to be discoverable inside ChatGPT, Gemini, and similar surfaces is real and worth doing. It is not a substitute for layer three. Run both.

What does not work is treating this as a procurement decision. The retailers who go shopping for an "AI shopping assistant," sign a vendor, install a widget, and declare the box ticked are the ones who will have to redo the project two years from now. The ones who treat it as a channel build, with the operational depth that implies, will not.

The Underlying Argument

The deeper claim under all of this is that the storefront is no longer a destination. It is something the retailer's agent constructs, in real time, for each shopper who arrives. For two and a half decades, ecommerce has been the digital version of a catalogue: a layout, a hierarchy, a set of product pages, and a checkout. The interaction model has been browsing. The retailer's job has been to organise the browse so that it converts, and to do that for the average shopper rather than the specific one.

The agentic shift does not improve that model. It replaces it. The retailer's job becomes operating an agent, in their own voice, on their own infrastructure, that decides what each shopper sees. The catalogue is still there, but it is no longer the surface. The surface is whatever the agent generates for the person looking at it. Two shoppers, two surfaces. The same shopper a week later, a different surface again. The static layout, the universal homepage, the one product detail page that everyone sees, all of these become artefacts of an older era, kept alive primarily for crawlers and partners.

Retailers who absorb this and build for it will be better positioned in five years than retailers who keep treating AI as a feature to be added. The technology is not the hard part. The technology is broadly available, improving fast, and increasingly commoditised. The hard parts are the integrations beneath the agent, the operating model around the channel, and the discipline to let the agent generate the surface rather than overriding it with the layouts the team is used to designing.

The next decade of ecommerce will not be won by the retailer with the best AI model. It will be won by the retailer who owns the surface their customer sees at the moment intent is formed, and who has built an agent that can construct that surface better than anyone else's.

That is the strategic frame; everything else is implementation. It does not end here. In Part 2, I will discuss the technical perspectives of agentic commerce.