AI e-commerce's true business is hidden before the user makes a payment

Over the past thirty years, e-commerce infrastructure has almost always been optimized around the same move: making it smoother for users to hit the buy button. One-click checkout, tokenized payment credentials, facial recognition, and fingerprint verification—all aim to reduce friction at the moment of payment.

But once AI Agents enter the shopping flow, change begins before payment. What the Agent must solve is how to understand user intent, filter products, build the shopping cart, and perform the next step within clearly defined authorization boundaries.

Therefore, Agentic Commerce cannot be understood as just a payment problem. Payments still matter, but they are only the last link in the shopping journey. The first part rewritten by AI is the process that happens before transaction approval.

1. User intent becomes executable constraints

In traditional e-commerce, human purchasing decisions are usually a prolonged browsing process—stretched out and inherently random. Users search, click into product pages, compare reviews, switch platforms, and are influenced step by step by prices, page design, promotions, and recommendations.

But when an Agent acts on the user’s behalf, the decision process before checkout is reorganized. Users may not provide a specific platform or a specific product; instead, they provide a natural-language description that includes multi-dimensional constraint conditions. These conditions typically include: a clear budget cap, specific fulfillment timeframes, a brand blacklist to exclude, and personalized parameter preferences.

In the past, these conditions were just preferences in the user’s head; now they become filtering rules the Agent uses when executing tasks. The Agent’s job is to break down a natural-language need into constraints that machines can understand, compare, and carry out.

This changes how value is allocated across the business chain. Previously, merchants competed for clicks and conversions on product pages; in the future, many products will first pass through the Agent’s filtering. Pages will still matter, but product data, real-time prices, inventory accuracy, delivery commitments, return rules, and parameter structures will come first—deciding whether a product can even enter the candidate list.

2. AI-generated answers become the new product display interface

These parameter structures determine first whether a product can enter the candidate list. And after user intent is broken down into constraint conditions, those constraints are fed first into AI search engines and models—giving rise to an entirely new product display interface.

If the entry point of traditional e-commerce is the search results page, then the new shelf in AI commerce is the answer itself, generated directly by the model.

In the past, when users searched for a consumption need, what they saw was a set of web pages, ads, reviews, and e-commerce links. Brands competed for rankings, clicks, and conversions; users had to open pages, compare parameters, and judge whether the information was genuine.

Now, AI generative search shortens this process. Products like AI Overview, ChatGPT, and Copilot compress multiple sources into a single answer, directly providing candidate products, applicable scenarios, and purchase recommendations. Users may not click through dozens of links, and they may not know which content sources were referenced behind the answer. What brands now need to compete for is inclusion in the candidate list within the AI answer.

This is also why GEO is more sensitive than SEO: it affects not only exposure, but also the model’s reasoning process. In traditional search, brands compete for higher positions among links; in generative search, brands compete for being included in the candidate list inside the answer. When AI compares different products in a seemingly neutral tone, it has already performed an initial screening for the user.

The issue is that AI recommendations are not generated from nothing. They reference review rankings, forum discussions, short videos, e-commerce reviews, industry reports, and more, and then compress that information into an answer that appears objective. Brands may not need to buy AI ad placements directly; they can also influence what the model sees when searching and summarizing through content distribution. A single piece of content may look like ordinary word-of-mouth, but when similar viewpoints show up repeatedly across multiple channels, AI may treat them as stronger evidence for recommendation. In this way, commercial placements may not appear in the form of ads, yet can still enter the model’s decision-making process.

This explains why Google is more cautious about GEO: its core asset is search trustworthiness. For a long time, users trust Google to rank relatively reliable information higher—so advertisers are willing to pay around this entry point. In traditional search, Google mainly displays links, and users can judge sources themselves; AI Overview directly provides an answer. If this answer is influenced by fake reviews, content farms, or biased content, Google is not just showing a lower-quality page—it may also be generating advice that misleads users.

Of course, different platforms’ stances on GEO are also shaped by their respective business models. Google’s focus is to preserve search trustworthiness, so it emphasizes anti-toxicity and content quality; Microsoft treats GEO as a way for advertisers to reach Copilot, Bing, Edge, and future Agents. In other words, in the future, GEO won’t have only a single unified set of rules—it will evolve different governance boundaries and commercial entry points across search, AI assistants, and model platforms.

However, for brands to be accepted and trusted in generative search, besides laying out opinion content across the web to influence the model’s summarization, there is a more fundamental technical prerequisite: the product itself must have a high degree of machine readability.

3. The split of the e-commerce storefront: data belongs to machines, taste belongs to humans

To be accepted and recommended by AI agents, products first must be highly machine-readable.

Early web commerce interfaces were tailored for human visual perception. Designs such as product images, copy descriptions, and add-to-cart buttons were aimed at extending users’ time on the page. But AI agents cannot perceive these visual designs; the machine’s basis for evaluating products completely returns to underlying structured data: SKU specifications, real-time inventory, net prices, Service Level Agreement (SLA), and structured return policies.

This shift makes machine readability the foundational competitive advantage in the AI shopping era. Schema.org markup, llms.txt files, real-time inventory and price APIs, and the structure of return policies all affect whether an AI Agent can accurately understand a product. Large language models can scrape unstructured information from web pages, but such information is often incomplete, updates slowly, and is prone to noise. By contrast, a standardized structured catalog can more directly tell an AI Agent: product specifications, real-time prices, inventory status, delivery capabilities, and return rules. This is the prerequisite for products to enter the Agent’s filtering and recommendation pipeline.

But this transformation will not happen at the same pace for all products; it can be categorized by consumption type into two groups.

One group is efficiency-driven consumption. For example, buying toilet paper, data cables, office supplies, or comparing flight and hotel prices. These decisions often have clear, hard standards: price, specifications, delivery time. Users don’t need to enjoy the selection process; they just want the answer as quickly as possible and as reasonable as possible. In these areas, AI agents can move extremely fast, directly handling price comparisons and placing orders on the user’s behalf.

The other group is consumption driven by taste and self-expression—for instance choosing a coat, a vintage lamp, or a painting. This kind of consumption carries people’s emotions, personality, and aesthetics. The process of selecting, hesitating, and comparing is itself a form of enjoyment. In these scenarios, AI’s value is more likely to occur before payment: helping users organize inspiration, understand styles, and aggregate scattered information to make exploration smoother.

This is precisely the entry point for the fashion shopping app The Mall. Today, the discovery entry points in online shopping are broken into extremely fragmented pieces: brand official websites, Instagram, TikTok, email newsletters, discount sites, friends’ recommendations, and creator outfit content each occupy their own corners, forcing consumers to keep jumping back and forth between these scattered spots. The Mall chooses to bring these scattered points together again, within the space of a virtual mall, to respond to everyday needs around how people discover, track, compare, save, and share brands and products.

Within this space, user behavior is reorganized. People can follow brands and track new releases and discounts, save products and observe what friends or creators are doing, and even use AI to understand style—jumping from one product directly to similar items from other brands, thereby stumbling upon niche brands that would otherwise not be pushed into view by ads.

This means the new commercial scene for AI intelligence does not have to be squeezed only into the final transaction closed loop; the front-end of consumption decision-making also contains enormous potential.

While the industry generally discusses how to have Agents help humans place orders with a single click, the pre-checkout processes—hesitation, comparison, and building taste preferences—namely browsing, tracking, discovery, comparison, and following that take place before checkout, themselves can be built into a high-value business.

And helping users manage these fragmented aesthetic intentions can not only build deeper trust, but the intent data accumulated over time can also deliver significant commercial returns. By organizing and recording users’ cross-brand preferences and comparison behaviors before payment occurs, products in this space have the potential to become a System of Record for consumers’ taste and intent. A discovery layer close to the source of decisions—its accumulated data asset value—may not be any less than the commission extracted in the final link.

So in the future, e-commerce storefronts will have two layers. One layer is for machines: it is responsible for data granularity, structuring, and verifiability, enabling Agents to efficiently perform price comparisons, procurement, and ordering. The other layer is for humans: it is responsible for brand expression, the transmission of aesthetics, building an experience, and creating Serendipity—so users are willing to stay, explore, and form a distinct style. In the past, merchants mainly operated the visual experience of the front-end web pages; in the future, they will need to operate both machine-readable structured product catalogs and that more hidden, more imaginative intent space before payment.

4. The promise chain moves upstream: letting Agents act within authorization boundaries

Once an Agent completes product discovery, filters options, and builds the shopping cart, the transaction truly enters the payment system.

From the operational perspective, modern card network mechanisms are essentially a promise chain with delayed fulfillment. During authorization, merchants verify credential validity by routing through acquirers and card organizations to the issuing bank; after the issuing bank returns an approval instruction, the merchant delivers the goods, while actual clearing and settlement are completed asynchronously in later windows. This system is built on a simple premise: transactions are initiated by humans and ultimately borne by humans.

The involvement of AI Agents breaks this assumption. The user authorization is no longer merely a single, definite payment; instead, it becomes a stream of decisions autonomously advanced by software components. If the Agent is attacked by malicious prompt injections, misreads context, or misconfigures parameters—leading to over-privileged or out-of-bounds transactions—then the legal boundaries for responsibility and liability become extremely blurred.

Although some retail giants try to push risk back to the user side by modifying Terms of Service (ToS) and legally defining transactions initiated by third-party Agents as being authorized by the user, this does not solve the engineering-level risk. Agentic Commerce needs to implement constraints before the transaction occurs.

In traditional financial networks, such constraints are mainly enforced through centralized gateway authorization. Visa, Mastercard, and other card schemes are developing standards around Agent identity, tokenized credentials, and verifiable intent, with the goal of narrowing the scope of what machines can do before a payment is approved. In practical scenarios, this usually means turning payment credentials into a programmable boundary: generating a one-time virtual card or token at the time of purchase and limiting it to a specific merchant, budget, category, time window, or specific task. Once the Agent exceeds these rules, the network can intercept the transaction at the authorization stage.

However, filtering only at the fund-disbursal endpoint or gateway layer is still a relatively downstream defense. The perspective of settlement giants is moving upstream toward the production source. Recently, Visa made an undisclosed strategic investment in AI software development platform Replit. Although the collaboration is still in an early exploration phase and has not yet released an official joint product, the industry signals it sends are very clear: global payment networks are trying to directly connect to the initial source generated by AI applications. By bringing Visa Intelligent Commerce and Trusted Agent Protocol into developer platforms, Visa aims to incorporate Agent identity, intent, and customer context into the payment system starting from the application development and deployment stages, rather than handling it only at checkout.

This is important because many future Agentic Transactions may not be initiated from a retailer’s own APP. They are more likely to come from software built by developers, distributed across different tools, and automatically executed by an Agent on behalf of users or enterprises. In that case, Replit would not just be a coding environment—it would become an application-layer entry point for Agent Commerce. For Visa, the bet on the future is that card network capabilities must become machine-native infrastructure: callable via APIs, capable of recognizing identity, and able to understand intent.

This is also the logic behind contract-based Agent wallets such as Cobo CAW Pact. It avoids giving Agents direct access to the full wallet balance; instead, it generates a temporary contract around a specific task and embeds the transaction path, the spending limit, and the validity period into the constraints. If a request falls outside the scope of the contract, the MPC nodes refuse to generate a signature. Before signing, the underlying calldata can also be translated into a human-readable transaction intent so the user can confirm what the transaction will actually do.

In the long run, the promise chain is shifting from trusting Agents to constraining Agents. Card networks place constraints at the gateway layer and are moving them upstream to the developer layer; on-chain systems must place constraints at the signature layer. Future payment systems will need to verify not only the payment identity, but also whether machine behavior stays within allowed boundaries.

Conclusion: Agentic Commerce needs a new promise chain

Technology often changes the medium of commerce, but it rarely eliminates responsibility itself.

E-commerce changed the transaction environment; mobile wallets changed payment credentials; API-based card issuance made authorization programmable; stablecoins began to affect parts of settlement. Each round of technological evolution adds a new layer of capability on top of the financial system. But what payment networks preserve—always—is authorization, clearing, settlement, and dispute handling.

The reason is simple: once a transaction enters the commercial system, someone must confirm the transaction can occur, someone must promise payment, and someone must bear responsibility when something goes wrong.

AI Agents make this chain longer. In the past, search, comparison, add-to-cart, and checkout were mostly done by users themselves; in the future, these actions may be handed to Agents and automatically executed across multiple systems. The experience may be faster, but it will become harder to judge what exactly the user authorized, how far the Agent can go, what obligations merchants bear, and how payment responsibility is divided.

That is the foundation infrastructure that intelligent commerce needs to rebuild. It requires an entirely new promise chain: at the very moment the transaction happens, bind the user’s initial intent, the Agent’s permissions, the payment commitment, and dispute responsibility together—and ensure the whole process is technically verifiable and traceable.

The future transformation of AI Commerce may look, on the surface, like a payment and automation issue, but at its core it is fundamentally a responsibility issue.