# Who Gets the Abandoned Cart When the Shopper Is an Agent?

Canonical page: https://sunlocke.com/articles/who-gets-the-abandoned-cart-when-the-shopper-is-an-agent

Page type: article

## Summary

AI shopping does not remove marketing. It breaks the assumption that every page view, cart, and abandoned checkout belongs to a person a merchant can reach. A strategy guide for marketing when an agent is browsing on behalf of a person, but the signal does not always come with a customer relationship. How marketing changes when agents create shopping signals without always exposing a human identity.

## Article details
A strategy guide for marketing when an agent is browsing on behalf of a person, but the signal does not always come with a customer relationship.

## Main point
How marketing changes when agents create shopping signals without always exposing a human identity.

Author: Chris Martinez

Section: Strategy

Post type: Diagnostic Lesson

Story format: feature

Reading time: 13 minutes

## Article content
The abandoned cart has always been a small story about desire. Someone wanted something enough to put it in a cart, but not enough to pay. That hesitation is where modern marketing found its opening. Send a reminder. Offer a reason to return. Give the person one more chance to finish the decision.

Now imagine the same cart assembled by an agent. It compares ten sellers, checks shipping, tests a budget, waits for approval, and then disappears. The merchant sees the cart. The merchant may even see a checkout event. But did a person hesitate, or did a process simply end?

The previous article in this discussion, [Who Gets to Be the Customer When AI Does the Shopping?](/articles/who-gets-to-be-the-customer-when-ai-does-the-shopping), asked who the customer might be when an agent becomes the buyer's interface. The marketing question is more uncomfortable. If the agent creates the signal, who gets to use it?

## The cart is about to stop meaning what we think it means

The traditional digital funnel depends on a quiet promise: the actions belong to one person. A person views an item, adds it to a cart, begins checkout, leaves an email address, and can be reached later. Google Analytics makes that sequence measurable through events for item views, cart changes, checkout starts, purchases, refunds, and promotions. The numbers are useful because they appear to describe a continuous human journey.

The recovery machinery depends on the same assumption. Shopify defines an abandoned checkout around an incomplete checkout after an email has been provided. Its browse, cart, and checkout automations can send email or SMS because the merchant has both an event and a person to contact. Even the platform's bot filtering reveals the old boundary. Some automated activity is removed because it is not considered real customer behavior, but Shopify also notes that not every automated event is filtered out.

An agent makes the old vocabulary unstable. A page view may be an agent reading the facts. A cart may be a scratchpad for comparing options. A checkout may be a temporary calculation that expires when the price changes or the person refuses to approve the total. The event is real in the database. Its meaning is no longer obvious.

That is why the coming change is not simply that AI will shop faster. It is that the marketing system may lose the ability to tell the difference between interest, evaluation, authorization, and intention. For years, the funnel turned behavior into a story about a customer. Agents can produce the behavior without producing the story.

## The infrastructure is arriving before the marketing rules

The companies building agentic commerce are solving the transaction problem first. OpenAI's [Agentic Commerce Protocol documentation](https://developers.openai.com/commerce/guides/key-concepts) describes separate flows for product feeds, checkout, and payment. The merchant still validates the order, calculates fulfillment and tax, evaluates risk, charges its existing payment provider, and decides whether to accept the order. The checkout can appear inside an AI interface while the merchant receives a very different record of how the decision was made.

Stripe is working on the layer below that. In its [WebMCP account of agent friendly Checkout](https://stripe.dev/blog/how-stripe-is-designing-checkout-for-ai-agents), Stripe reports 42 percent fewer tokens and 38 percent fewer tool calls against its own baseline after enabling WebMCP on Stripe Checkout interfaces. That may make an agent faster and cheaper to operate. It does not tell the merchant whether the agent was browsing for a person, testing a rule, or comparing every competitor in the category.

The payment plumbing is evolving at the same time. Stripe's [agentic commerce documentation](https://docs.stripe.com/agentic-commerce) separates selling through agents with UCP or ACP from machine payments with MPP or x402. Wallets, delegated payment tokens, and single use virtual cards can let an agent pay without exposing the underlying payment details. These systems are important because they make a transaction possible. They do not settle who owns the marketing relationship around the transaction.

That gap is where the speculation begins. The infrastructure being built can answer how an agent reads a catalog, how it opens a checkout, and how it spends with permission. It does not yet provide a shared answer to what a page view means, what consent follows an agent session, or who can retarget a person after an agent has moved on. The payment can be standardized before the meaning of the customer is.

## Who gets the moment between wanting and buying?

There are several possible answers, and each one produces a different market. In the first version, the merchant keeps the relationship. The agent brings a customer to a direct or account linked checkout, the merchant receives the information needed to fulfill the order, and the business earns the right to continue the relationship through receipts, service, loyalty, and future permission. This is the most familiar outcome, but it depends on the agent passing more than a payment token.

In the second version, the agent platform becomes the storefront. It owns the conversation, the comparison, the ranking, the approval prompt, and the moment when the person sees the result. The merchant may still own fulfillment and support, but the platform owns discovery. Marketing moves toward eligibility inside someone else's interface. The brand is not trying to win a click from a person. It is trying to be included in an answer.

The third version belongs to the payment layer. Wallets and networks may become the trusted intermediary because they can authorize a purchase, protect the underlying credentials, and pass a limited token to the merchant. Stripe says its Link agent wallet can complete purchases at more than one million Link accepting businesses and can issue a purchase scoped virtual card for other businesses. That is a strong payment proposition. It is not automatically a customer relationship proposition.

The fourth version is the one marketers will dislike most. Nobody gets the intent. An agent reads a feed, tests a product page, assembles a cart, and leaves no durable identity behind. The merchant has a record that the system was consulted, but no person to email and no clear permission to do anything with the event. The future could contain a large amount of demand that is visible to machines and dark to marketing systems.

The market is likely to settle somewhere between these models. A person may discover inside an agent platform, approve inside a wallet, transact through an existing payment provider, receive fulfillment from the merchant, and become loyal because of the service afterward. That is not one funnel. It is a chain of institutions, each holding a different piece of the relationship.

> The agent may have intent. The merchant may have an event. Neither proves that the merchant has a person it can contact.

## Retargeting meets the phantom customer

Retargeting works because an observed action usually carries a human explanation. A product view suggests interest. A cart suggests stronger interest. A checkout suggests that the person is close enough to purchase that a reminder might change the outcome. The merchant spends to close the gap between wanting and buying.

The agent introduces a phantom customer. The product view exists, but the person may never have seen the page. The cart exists, but it may have been created only to calculate shipping. The checkout exists, but the person may have rejected the price before the merchant ever received a usable identity. If every machine action triggers a human campaign, the merchant will spend money recovering decisions that were never lost.

The problem gets larger as agents become better at comparison. One person can browse a few stores. An agent can evaluate many stores quickly and repeat the task whenever price or inventory changes. A dashboard optimized around page views and cart events could end up rewarding machine activity rather than customer demand. The business may buy more media in response to a market that is only appearing in its analytics.

This does not mean remarketing disappears. It means the trigger has to become more selective. A known customer with a permitted account relationship is one thing. An unidentified machine session is another. The first may deserve a message. The second may deserve better product data, a clearer policy, or no intervention at all.

## The future may split into three storefronts

The first storefront will be the one people recognize. It will still have photographs, stories, reviews, service language, and the details that make a person feel that one business is more trustworthy than another. This storefront will matter after an agent has narrowed the field, because people will still approve purchases and decide whether the business deserves a second chance.

The second storefront will be machine facing. It will be made of product feeds, structured variants, inventory, pricing, shipping promises, return rules, seller links, and payment capabilities. OpenAI's [guidance for agent commerce](https://developers.openai.com/commerce/guides/best-practices) already treats factual descriptions, durable links, stable variant data, and feed attribution as operational requirements. In that environment, product information is not merely copy. It is the material from which a recommendation is assembled.

The third storefront may belong to the platform that decides which answers deserve attention. That platform could become a search engine, a retail marketplace, a wallet, a media company, or all four. It may not need to show a banner ad to change the market. It may only need to decide which products are eligible to be considered and which signals improve their position.

This is the speculative danger for traditional marketing. The old system rewarded the ability to interrupt a person. The new system may reward the ability to satisfy a machine's requirements before the person enters the scene. Brands that are emotionally memorable but structurally illegible may disappear from consideration. Brands that are easy for machines to compare but empty after purchase may win the clickless transaction and lose the customer.

Marketing therefore splits in two directions. It moves upstream toward eligibility, accurate data, and machine readable proof. It moves downstream toward service, memory, trust, and the moments after the payment. The middle of the funnel, where a person could once be followed from page to cart to reminder, becomes the least reliable place to build a relationship.

_A measurement warning_

### A sale is not proof of new demand

If an agent platform reports a sale, the merchant still needs to ask whether the platform created a new relationship or captured a decision that was already made elsewhere. Attribution can identify the final handoff without proving that the handoff caused the purchase.

## What businesses should prepare before the pattern hardens

The first preparation is not a proprietary agent application. It is a clean description of what the business sells, how much it costs, when it is available, how it is delivered, and what happens when the customer changes their mind. A small business should make those facts reliable before spending money to enter another interface. It should also keep unknown machine sessions out of email and SMS recovery audiences unless a legitimate relationship and permission exist.

A medium sized business should make the distinction visible in its data. Discovery, agent session, human approval, merchant checkout, payment authorization, fulfillment, and repeat purchase should not collapse into one conversion event. The business should be able to compare an agent assisted order with a conventional order and ask whether the new channel created demand, shortened a journey, or simply claimed the final step.

A large business will have to negotiate the relationship instead of merely integrating a feed. It may need agreements about product freshness, ranking inputs, customer identity, consent, returns, referral reporting, and fraud responsibility. It will also need to decide which part of the customer relationship it is willing to let an agent platform mediate and which part must remain direct.

Every business should watch the same tension: attributed conversion is not necessarily incremental conversion. If a platform reports a sale, the merchant still needs to ask whether the platform created a new relationship or captured a decision that was already made elsewhere. The more the agent becomes an intermediary, the more important that question becomes.

The useful measures will not be limited to page views and cart recovery. Businesses will need to understand agent sourced discovery, human approval, order completion, payment failure, margin after channel fees, returns, repeat purchase, and the share of orders attached to a durable customer identity. When the platform does not expose a signal, the business should record that limitation rather than manufacture precision.

## What if the cart belongs to no one?

The most important possibility is not that AI gets all the marketing. It is that marketing loses its old claim over the middle of the journey. Discovery may belong to an agent platform. Payment may pass through a wallet or a delegated token. Fulfillment may belong to the merchant. Loyalty may be earned by the business that keeps its promise after the transaction.

In that world, the abandoned cart is no longer automatically a missed opportunity for a recovery email. It may be a decision that the system made correctly. It may be evidence that the offer failed a constraint. It may be a person refusing to approve the total. Or it may be nothing more than a machine moving to the next result.

The businesses that prepare well will not be the ones that assume every consumer wants an autonomous shopper. The current systems are limited, partner based, and unevenly adopted. The prepared businesses will be the ones that can tell the difference between a machine looking, a person deciding, and a customer returning.

The first article asked who gets to be the customer. The next question is harder: when the cart no longer proves that a person is there, who gets to decide whether the moment belongs to marketing at all?[ Return to the opening discussion.](/articles/who-gets-to-be-the-customer-when-ai-does-the-shopping)

Related story: [Who Gets to Be the Customer When AI Does the Shopping?](/articles/who-gets-to-be-the-customer-when-ai-does-the-shopping)

***

## Sources

Source: Google Analytics, Measure ecommerce

Google describes ecommerce events as a way to measure shopping behavior, including item views, cart changes, checkout starts, purchases, refunds, and promotions.

Reference: https://developers.google.com/analytics/devguides/collection/ga4/ecommerce

Source: Shopify, Recovering abandoned checkouts

Shopify defines an abandoned checkout around an incomplete checkout after an email has been provided, limits some recovery flows by channel, and notes that not all automated activity is filtered out.

Reference: https://help.shopify.com/en/manual/promoting-marketing/create-marketing/abandoned-checkouts

Source: OpenAI Developers, Key concepts for Agentic Commerce

The documentation describes separate product feed, checkout, and payment flows. The merchant remains responsible for order validation, fulfillment, tax, risk, payment processing, and acceptance decisions.

Reference: https://developers.openai.com/commerce/guides/key-concepts

Source: Stripe Dev Blog, How Stripe is designing Checkout for AI agents

Stripe reports its own WebMCP baseline results of 42 percent fewer tokens and 38 percent fewer tool calls. The figures measure agent interaction efficiency, not consumer demand.

Reference: https://stripe.dev/blog/how-stripe-is-designing-checkout-for-ai-agents

Source: Stripe, Agentic Commerce documentation

Stripe distinguishes selling through agents with UCP or ACP from machine payments with MPP or x402 and lists different capabilities for product feeds, embedded checkout, payment credentials, and agent wallets.

Reference: https://docs.stripe.com/agentic-commerce

Source: Stripe, Stripe helps Muse shop across the internet with Link

Stripe describes Meta Muse using Link for agent purchases. It reports more than one million Link accepting businesses, one time virtual cards for other businesses, and consumer approval of each purchase total.

Reference: https://stripe.com/gb/newsroom/news/stripe-helps-meta-muse-shop-with-link

Source: Shopify, Creating and managing marketing automations

Shopify lists separate browse, cart, and checkout automations across email and SMS. The flows are designed around customer actions and marketing permissions, which is precisely the assumption agent activity puts under pressure.

Reference: https://help.shopify.com/en/manual/promoting-marketing/create-marketing/shopify-messaging/marketing-automations/create