The most consequential change in AI shopping may not be a better recommendation.
It may be what happens immediately after one.
A buyer asks an AI assistant which product to choose. The assistant presents options. The buyer selects one. Payment is authorized. An order reaches the merchant.
The conversation does not need to end with a link.
It can end with a purchase.
Stripe and OpenAI made this architecture concrete with Instant Checkout in ChatGPT and the Agentic Commerce Protocol. Google introduced the Universal Commerce Protocol for agentic shopping. Microsoft announced Copilot Checkout. Perplexity and PayPal connected conversational discovery with payment. Amazon’s Buy for Me introduced another route: an agent completing a purchase on a brand’s website on the customer’s behalf. stripe
These initiatives are not identical. Their availability, integration requirements, and levels of delegated action differ.
But together, they expose a question that recommendation measurement alone cannot answer:
What happens between an AI-generated recommendation and an executable commercial order?
This article calls that missing architectural dimension the execution layer.
Not because recommendations have become transactions.
Because the infrastructure connecting them is becoming a subject worth measuring.
1. The Boundary We Must Preserve
The previous article in this series introduced an exploratory benchmark for AI Recommendation Share.
ARS measures qualified recommendation incidence across a declared set of buyer-intent opportunities. It distinguishes a recommendation from a neutral mention and allows multiple brands to receive qualified recommendations in the same response. dev
That boundary remains essential.
A recommendation is not a selection.
A selection is not a payment.
A payment is not successful fulfilment.
And none of these events, viewed alone, establishes incremental revenue caused by AI.
Connecting the stages does not make them interchangeable.
Instead, it creates a new research task: observing transitions between distinct events.
The right question is not whether ARS can replace transaction data.
It is whether recommendation evidence and transaction evidence can be connected without losing the meaning of either.
2. The Checkout Is Moving Toward the Answer
In a familiar online shopping journey, the search interface and the merchant’s checkout are visibly separated.
A user discovers a product, follows a link, evaluates the merchant’s offer, and completes a purchase elsewhere.
Conversational checkout changes that arrangement.
In Stripe’s September 29, 2025 announcement, ChatGPT users could discover products and complete eligible purchases within the conversation. The initial availability covered users in the United States buying from US-based Etsy businesses; Shopify merchant support was described as forthcoming. stripe
The interface could host more of the journey.
But the underlying responsibilities remained distributed.
The merchant received the order through ACP, retained the ability to accept or decline it, and remained responsible for charging the payment method, taxes, fulfilment, and returns. stripe
This is a crucial distinction:
The conversational platform can become the place where the purchase happens without becoming the merchant that fulfils it.
From the buyer’s perspective, the experience may feel unified.
From the infrastructure perspective, it remains a coordinated system.
3. This Is Not a Single-Platform Experiment
A framework built around one checkout announcement would be too narrow.
Several architectures now deserve comparison.
Google: connecting discovery with commerce infrastructure
Google introduced UCP in January 2026 as an open standard covering the shopping journey from discovery and purchase to post-purchase support.
Its announcement connected UCP with checkout experiences in AI Mode and Gemini, while preserving the retailer’s role as seller of record. business.google
Google’s developer documentation also presents UCP as a route for enabling commerce actions in those AI environments. developers.google
This complicates a simple division between “traditional search infrastructure” and “new AI decision interfaces.”
The same company can participate in discovery, conversational assistance, merchant connectivity, and checkout.
The strategic unit is not necessarily a standalone model.
It may be a stack of connected capabilities.
Microsoft: discovery and transaction eligibility are different
Microsoft announced Copilot Checkout as a way for shoppers to complete purchases discovered in Copilot without being redirected to external sites. Its announcement named PayPal, Shopify, and Stripe among its partners. news.microsoft
Microsoft’s commerce materials also distinguish product feeds used for discovery from UCP-based capabilities used for transactions and post-purchase actions. about.ads.microsoft
That distinction deserves attention.
A product can be discoverable without being purchasable through a particular interface.
Integration into a recommendation environment and integration into its transaction environment are separate achievements.
Perplexity and PayPal: a conversational payment path
PayPal’s March 2026 guide describes a Perplexity shopping flow in which users receive product cards and, for eligible offers, can select Instant Buy or Buy with PayPal.
The guide describes availability for US users on supported web and desktop experiences, with participation depending on eligible merchants and products. paypal
This is another example of an answer-oriented interface connecting discovery with payment through a partner’s infrastructure.
It is not evidence that every product mentioned in an answer can be purchased there.
Amazon: the agent can visit the checkout for you
Amazon’s Buy for Me follows a different route.
In its April 2025 beta announcement, Amazon described an agent purchasing selected products on brand websites on behalf of customers, following customer confirmation within the Amazon experience. The brand remained responsible for delivery, returns, and customer service. aboutamazon
Here, the merchant’s existing website can remain the execution environment.
The agent acts across that environment rather than relying exclusively on a shared commerce API.
This difference matters.
An integrated checkout and an agent-operated website purchase may look similar to the buyer.
They are not the same architecture.
4. Defining the Execution Layer
For this framework, the execution layer is:
The interfaces, permissions, identity mechanisms, commercial data, and transaction services that allow a selected offer to become an order a merchant can evaluate and fulfil.
This is a proposed analytical definition, not terminology attributed to a particular platform.
It includes more than payment.
Before an order can proceed, several conditions may need to be satisfied:
- The correct product and variant must be identified.
- The offer must be available to the buyer.
- Price and delivery information must be usable.
- The buyer’s authorization must be established.
- The agent’s request must be accepted.
- The merchant must receive a valid order.
- Payment and fulfilment must proceed.
A brand can perform well at one stage and fail at another.
That is why “AI visibility” is too broad a label for the entire process.
Visibility describes neither operational readiness nor completed commerce.
5. Four States That Should Never Be Collapsed
Consider a hypothetical buyer asking for a product below a fixed price, delivered to a specific location before a deadline.
Brand A receives the strongest recommendation.
But its relevant variant is unavailable.
Brand B also receives a qualified recommendation, has stock, and supports delivery within the required period.
The buyer selects Brand B.
The payment then fails.
Four distinct observations exist:
| State | Question |
|---|---|
| Recommended | Did the response endorse the product for the buyer’s requirements? |
| Executable | Was a valid purchase path available under those requirements? |
| Selected | Did the buyer choose the offer? |
| Completed | Did the transaction reach the defined successful outcome? |
Brand A may deserve recommendation credit.
Brand B may deserve both recommendation and selection credit.
Neither event establishes a completed order.
A measurement system that compresses these states into one score will struggle to explain what actually happened.
It may also prescribe the wrong intervention.
More visibility will not fix unavailable stock.
A smoother checkout will not fix poor product fit.
6. Trust Is Part of Execution
An agent arriving at a merchant’s website introduces another problem:
How does the merchant distinguish a legitimate buyer-authorized agent from an unwanted bot?
Visa’s Trusted Agent Protocol addresses this boundary.
Its technical documentation describes mechanisms for recognizing an agent, carrying consumer-recognition information, and transmitting payment-related information through signed structures. The stated purpose includes helping merchants distinguish trusted commercial interactions from other automated traffic. developer.visa
This is not recommendation ranking.
It is permission to participate in the commercial interaction.
Visa also describes TAP as usable across different channels and protocols, including web-based flows and ACP. developer.visa
The implication is architectural:
Commerce protocols and trust mechanisms can be complementary.
One may define how an order is communicated.
Another may help establish whether the actor communicating it should be trusted.
A useful execution framework must account for both.
7. There Is Another Step Beyond “Buy Now”
Not all agent-mediated purchases happen immediately after a recommendation.
Amazon’s November 2025 description of Rufus included a capability allowing customers to set a target price and have a purchase made when the condition was met, using default payment and delivery settings. aboutamazon
This introduces a different timeline:
- The user expresses a preference.
- The user delegates a conditional instruction.
- The condition is monitored.
- Execution occurs later.
The recommendation, authorization, and purchase may happen at different times.
For research, that creates difficult but important questions:
- Which event defines the start of the journey?
- How long does authorization remain relevant?
- Was the purchased product the one originally evaluated?
- Did price or availability change?
- Which intervention deserves attribution?
A session-based view may be insufficient.
The unit of analysis may need to become an authorized task or a purchase journey.
8. Open Commerce Is Not Automatically Open Recommendation
ACP and UCP are presented as open standards.
Stripe also states that merchants do not have to use Stripe as their payment processor to participate in ACP-based commerce. stripe
These are meaningful interoperability properties.
But interoperability does not, by itself, explain recommendation governance.
A protocol can describe how an agent interacts with a merchant without explaining:
- Which products enter the candidate set.
- Which merchants receive exposure.
- How suitability is assessed.
- How transaction eligibility affects ranking.
- How commercial incentives influence presentation.
These are separate questions.
An ecosystem can have an open transaction interface and still require independent scrutiny of how recommendations are produced.
Equally, an unexplained ranking difference should not automatically be treated as evidence of commercial bias.
Availability, geography, product suitability, and integration coverage may also explain it.
The research task is to separate these possibilities.
9. Advertising Makes Classification More Important
Google’s UCP announcement also introduced a Direct Offers pilot: a way for advertisers to present special offers, such as discounts, to shoppers in AI Mode. The example was explicitly presented as a sponsored deal. business.google
This adds another classification problem.
An observer may encounter:
- An organic recommendation.
- A sponsored offer.
- A product card with checkout eligibility.
- A response from a brand-specific assistant.
Those are not equivalent recommendation environments.
An ARS study should declare how each is handled.
If sponsored placements are included, they should be identifiable.
If they are excluded, the exclusion should be explicit.
Otherwise, a rise in measured “recommendation share” might partly reflect a change in advertising exposure or interface composition.
The solution is not to assume every commercial result is compromised.
It is to preserve the evidence needed to distinguish its origin.
10. What Should an Observatory Record?
The earlier articles proposed an AI Recommendation Observatory: evidence-linked measurement across declared platforms, scenarios, and observation periods. dev
The execution layer extends its possible event model.
It should not require every researcher to observe every event.
A proposed record could include:
| Event | Evidence to preserve where observable |
|---|---|
| Recommendation | Response text, brand, product, classification |
| Offer presentation | Merchant, variant, price, timestamp |
| Execution eligibility | Geography, stock, delivery, supported purchase path |
| Authorization | Observable confirmation and scope |
| Checkout initiation | Session or task reference |
| Order acceptance | Merchant-confirmed status |
| Payment outcome | Success, failure, or unresolved state |
| Post-purchase outcome | Cancellation, fulfilment, return, or refund |
These are proposed research fields.
They are not a claim that any named protocol exposes all of them.
A public benchmark may see the response but not the order.
A merchant may see an order but not the alternatives previously shown.
A payment provider may see payment activity without the complete recommendation context.
Those visibility limits must survive into the analysis.
Unknown is not zero.
Unobserved is not failed.
11. Connect the Metrics—Do Not Merge Their Denominators
ARS should retain its declared recommendation denominator. dev
Execution analysis introduces different denominators.
For example:
[
P(\text{checkout initiated}\mid\text{qualified recommendation observed})
]
and:
[
P(\text{order accepted}\mid\text{checkout initiated})
]
These measure different transitions.
They also require linked observations, compatible units, and clear treatment of missing data.
One response can recommend several brands.
One order can contain several products.
One user can return across multiple sessions.
A study must therefore specify whether it measures conversations, recommendation events, users, tasks, orders, or order lines.
Without that discipline, a conversion figure can look precise while remaining conceptually unstable.
Even correctly linked data does not automatically establish causation.
People receiving different recommendations may differ in intent, location, budget, or product requirements.
Descriptive association is useful.
Incremental commercial impact requires a stronger research design.
12. Who Wins? First Specify the Layer
The emerging architecture does not support a simple winner-takes-all conclusion.
Different actors may hold different advantages:
| Position | Potential source of advantage |
|---|---|
| Buyer interface | Access to intent and repeated interaction |
| Recommendation system | Influence over the consideration set |
| Commerce protocol | Lower integration friction |
| Trust infrastructure | Recognition and authorization of agents |
| Payment infrastructure | Reliable transaction processing |
| Merchant operations | Product availability and fulfilment |
These are analytical categories, not claims of equal economic value.
Their importance must be tested.
But they help explain why Google, OpenAI, Stripe, PayPal, Microsoft, Amazon, Visa, and merchants can participate in overlapping roles rather than forming two clean opposing camps. Their announced initiatives span different parts of this architecture. stripe
The company measuring recommendation share occupies another position again.
Designing the metric does not automatically confer control over the recommendation system, checkout protocol, or transaction economics.
The observatory and the marketplace should remain analytically distinct.
13. The Next Empirical Study
The previous benchmark examined software recommendations. It should not be presented as evidence about retail checkout conversion. dev
A credible extension needs a new scope.
Start with one product category, one geography, one defined observation period, and a limited set of commerce environments.
Then ask:
- Which products receive qualified recommendations?
- Which have executable offers under the same buyer constraints?
- How often do those sets overlap?
- What happens after selection, where authorized outcome data exists?
- Where does the journey become unobservable?
Record price, stock, delivery conditions, platform configuration, and timestamps.
Repeat comparable scenarios.
Separate sponsored presentation from non-sponsored recommendation.
Link merchant outcomes only where lawful access and appropriate authorization exist.
The first objective should be a reliable map of transitions.
Not a dramatic claim that recommendation share has become market share.
Conclusion: The Market Between Answers and Orders
The earlier question was:
Which brands does AI recommend?
The next question is:
Which recommendations can become valid commercial actions—and under whose rules?
Instant Checkout, UCP, Copilot Checkout, Perplexity–PayPal, and Amazon Buy for Me demonstrate different ways to connect conversational discovery with purchasing. Visa’s trust framework shows that execution also depends on recognizing and authorizing the actors involved. stripe
These developments do not erase the distinction between influence and transaction.
They make the relationship measurable in new ways.
For brands, the challenge is no longer only whether an AI system mentions them.
It is whether the right product is recommended, a valid offer is available, the buyer authorizes the action, and the merchant can deliver.
For researchers, the task is to connect those events without confusing them.
The answer box is becoming a commercial interface.
The next measurement frontier is the distance between what it recommends and what the market can actually execute.
Author: Seyed Alireza Alhosseini Almodarresieh
Website: https://almodarresieh.github.io/
Top comments (0)