When AI Helps You Shop, How Do You Confirm an Order and Follow Up?
Google's September 2026 announcement and public UCP specifications: product visibility, buyer consent, order records and support after purchase.
Agentic commerce uses AI to help find products and carry out purchase steps. Google’s September announcement and the public UCP specifications distinguish visibility, buyer consent, order placement and post-purchase support. What should buyers and sellers check?
TL;DR
- Google’s September announcement treats AI visibility measurement and checkout integration as separate features
- Checkout integration alone guarantees neither ranking, sales nor buyer satisfaction
- Check reviewed terms, processing status, order placement and delivery or refund handling separately
The thumbnail is an AI-generated illustration, not a photograph documenting a real service or purchase.
This article explains public specifications and business guidance, not legal advice for an individual case. Consult a qualified legal professional or the responsible authority about particular transaction terms and personal-data obligations.
What Is Happening
Read the September announcement and regional conditions, separating AI visibility from orders.
Imagine asking AI to find cups for a gathering next week and receiving suggestions that fit your preferences. Even if that screen lets you buy, a product appearing, a person deciding to purchase and a seller receiving an order are different events. If you feel the shopping is finished before checking the total including shipping and the expected arrival, it may be difficult to know what to verify afterward.
Agentic commerce describes systems in which AI helps find products or advance purchase steps on a person’s behalf. On September 16, 2026, announced related features for the holiday season. This article uses Google and UCP materials checked on November 4 and rechecked on November 6 (UTC) to examine what buyers and sellers can confirm. It also adds discussion of recommendation-related interests from Consumer Commission expert materials checked on November 6.
The announcement covers insights into how products are found on AI surfaces, a U.S. beta for answering questions within YouTube ads, and updates to the UCP checkout integration hub. Their availability differs. The new integration hub capabilities are rolling out gradually in the U.S., with Australia and Canada planned for early 2027. The announcement cannot be presented as making every feature available in Japan.
More Visibility Does Not Establish More Orders
is a feature in Merchant Center, Google’s merchant management service, for examining appearances on AI search surfaces. It covers accounts in Australia, Canada, India, New Zealand and the United States, and English-language queries. It measures organic AI traffic and excludes paid advertising.
Its visibility share is the merchant’s share of total appearances for the defined comparison brands and the merchant, rather than sales across the whole market. Changes to the comparison group can raise both the merchant’s share and competitors’ average; insufficient competitor data can also produce a high reported share. Sellers should align the scope and period, then check orders and returns in separate records.
The official help also distinguishes discovery, evaluation and queries close to a purchase. A query classified as close to a transaction does not establish that the person accepted purchase terms or actually placed an order. Sellers should distinguish a classification of search intent from a record of what a person did. The report can inform understanding of interest before purchase without establishing what happened afterward.
Check visibility
AI performance insights
- Traffic covered
- Organic
- Comparison scope
- Defined peers
- Order placement
- Check apart
Integrate checkout
UCP integration hub
- Preparation
- Implemented
- Participation
- Selection
- Native ranking
- No change
Author’s synthesis of Google’s September 16 announcement, the insights and integration hub help, and the UCP FAQ. Check countries, languages and individual features separately.
Increasing visibility and reducing failures during checkout are different tasks for a seller to assess. More detailed product descriptions might increase appearances while some people still stop at shipping charges or address restrictions. Visibility alone cannot establish which event occurred. Comparing purchase stages over the same period would help investigate this.
Background & Context
Examine the shared protocol and the inventory and address checks sellers still perform.
Introduced in Google’s January announcement（外部サイト、新しいタブで開きます）, Universal Commerce Protocol (UCP) is an open standard for AI surfaces and sellers’ systems to exchange product and checkout information in a shared format. It aims to reduce the work of building a different connection for each AI service.
In March, Google（外部サイト、新しいタブで開きます） added a cart, retrieval of product details such as prices and inventory, and account linking. Adopters choose capabilities, so a claim of UCP support does not establish that member pricing or multi-item purchases work. A feature defined in the standard and one available at a particular store must be distinguished.
This distinction also affects sellers’ preparation. A local shop might have consistent product descriptions while still needing to check whether it can return inventory and shipping charges at purchase time. Where the same product appears through several channels, one possible approach is to assign responsibility for description updates and checkout conditions. Actual cost and time depend on the store’s systems; a common connection does not establish that the workload disappears.
Sellers Still Check Inventory and Delivery Conditions
Google’s describes completing technical implementation, registering interest, being selected and then accessing a validation environment. Implementing an open standard and qualifying to offer checkout on Google are different processes. A declaration of compatibility cannot by itself establish a feature’s actual availability.
The seller remains the Merchant of Record, the seller responsible for the purchase transaction. Google’s assigns inventory checks and deliverable-address validation to the seller and explains that the seller’s terms govern the purchase. Even when a buyer is speaking to AI, the seller still handles stock shortages and shipping restrictions.
The FAQ also says that choosing Native integration, which directly integrates checkout, does not affect ranking in product listings. This is Google’s account of its feature, not an independent assessment of fairness across every AI recommendation. Even so, it does not support interpreting checkout integration as a guarantee of a better ranking. Sellers should check listing conditions and purchase conditions separately.
Keeping the customer relationship does not mean customer information may be used for any purpose. The FAQ states that standard privacy and consent practices apply to re-marketing. Receiving information required for a purchase and obtaining permission to send later communications are separate matters for a seller to check. An integration choice alone cannot determine how customer information may be used.
Recommendation Interests and Purchase Authorization
In materials submitted to the Consumer Commission’s expert panel on November 5, 2026, raises the possibility that AI outputs could favor providers or sponsors when their interests differ from users’ interests. The presentation also discusses the difficulty of relying solely on market competition, courts or administrative authorities. These are an expert’s issues for discussion, rather than verified findings of biased recommendations by a particular service or new rules adopted by the committee.
Applied to shopping assistance, this raises a question alongside what the buyer authorized: which options were considered, and on what basis were they recommended? How much users can learn about recommendation reasons, the range of candidates and relationships among providers, sellers and sponsors is separate from a purchase signature. This is the article’s proposed check; it does not mean that the presentation establishes a particular disclosure obligation.
Reading the Structure
Check buyer review, processing, order placement and delivery or refunds as distinct records.
Preserve the Terms the Buyer Reviewed
The UCP requires the user to finalize a standard checkout through a trusted interface. It separates AI assistance with choosing products and entering information from reviewing the order and deciding to purchase. Support for the AP2 authorization extension introduces different conditions, so this should not be interpreted as a protocol that prohibits every automated purchase.
The uses verifiable signatures to bind the seller’s terms, such as items and prices, to the purchase the user explicitly authorized. This design records authorization; it is different from establishing that the person understood those terms. Users should also examine readable presentation and a route back to a person when they are unsure.
For the cups in our hypothetical example, it would be useful to compare quantity, color, the total including shipping, the destination and expected arrival across the purchase review and order record. This is a proposed check, not a finding from testing an individual service. A record of how “something that arrives before the gathering” became purchase conditions could help investigate differences between the request and the order.
Suppose an item sells out and a different color is offered. The initial request alone cannot establish authorization for the changed order. The AP2 specification requires checking that the terms in the authorization match the current checkout. Showing the buyer what changed and recording the new choice would be one approach. Binding the terms through signatures and explaining the change clearly to the person should be considered together.
Distinguish Processing from Order Placement
The checkout specification separates ready_for_complete, when required information is present, from complete_in_progress, when completion has been accepted and the seller is processing it, and completed, when the order has been placed. The seller omits order information from a processing response and includes it in the response after placement. AI saying it advanced the purchase steps is therefore insufficient evidence of placement; check the seller’s reported state and the order record.
The specification also provides for handoff to the seller’s interface when buyer input or review is needed. Rather than treating every screen change as a failure, the system needs to explain what the person must decide. Measuring convenience only through fewer operations overlooks the steps needed to check the current state when a response is delayed. Operations staff implementing the specification check the current state before repeating completion.
Sellers and operations staff could test out-of-stock situations, address changes and delayed completion responses alongside successful purchases. They could record what the buyer saw, which state the seller returned and what ultimately happened to the order. This is a proposed validation approach. The existence of a public specification does not establish that an actual service has performed these checks.
Before purchase
The buyer reviews terms
- Items and quantity
- Match request
- Total and destination
- Review terms
- Completion response
- Check state
After placement
Consult seller records
- Placed order
- Order record
- Arrival
- Delivery log
- Refunds and questions
- Track support
Author’s synthesis of the public UCP Checkout, AP2 Mandates and Order specifications. This is neither a completed test of an individual service nor a delivery or refund guarantee.
Check Support After the Order
The distinguishes delivery expectations from actual shipment events and changes such as refunds. Checkout’s completed state does not establish that goods arrived or that a refund was handled. Fields for conveying this information also cannot establish that a seller answered an inquiry.
A community organization considering a new sales channel could test whether an order number connects to shipment information and inquiries, alongside checking the integration itself. Rather than measuring success only through sales, it could also record orders requiring renewed confirmation, instances where buyers could not find a contact and the progress of refunds over the same period. Actual records would make it possible to examine how purchase assistance changes the experience of buyers and staff.
Reference Books
選択の科学 コロンビア大学ビジネススクール特別講義 (文春文庫 S 13-1)（外部サイト、新しいタブで開きます） (editorial translation: The Science of Choice: A Special Lecture at Columbia Business School; Japanese edition) is related reading on choice and delegation. The publisher introduction, contents and bibliography were checked; the book itself is unread and is not evidence for current AI-shopping outcomes or UCP requirements. The link is an Amazon Associates advertising link.
Related Guide
- Risk Management in the Age of AI Agents（このサイトの記事）. Consider usage policies, review steps and permissions.
Remaining Questions
Consider how to assess readable consent, participation costs and support after an order.
Verifiable authorization signatures leave a separate question: could the buyer read the price and delivery conditions? Text size, language, explanations when terms change and a contact for uncertainty may affect different users. Alongside recording that a review screen existed, how can actual use establish whether people understood the conditions and could make a decision? Convenience should be assessed through both fewer operations and the person’s ability to judge the purchase.
Sellers still update product information, maintain integrations and answer inventory or delivery exceptions. An open standard alone cannot establish whether small shops and community organizations can sustain that preparation. Costs, staff time, regions and selection criteria should be examined, and visibility opportunities for participating and non-participating stores compared under the same conditions. Which stores reach buyers after adding a sales channel, and who takes on the resulting work, are questions for further investigation.
Buyers should check whether they can return to an order record and contact the seller after closing the conversation. Comparing AI replies, seller records and actual handling can help explain a problem when delivery is late or a refund is needed. Distinguishing a product being found and ordered from support progressing when something goes wrong enables a more concrete judgment of how much shopping can be entrusted to AI.
References
Boost your holiday sales with these agentic commerce updates — Ashish Gupta / Google (2026). Google
About AI performance insights — Google (2026). Google Merchant Center Help
How to onboard to the UCP integration hub in Merchant Center — Google (2026). Google Merchant Center Help
New tech and tools for retailers to succeed in an agentic shopping era — Vidhya Srinivasan / Google (2026). Google
AI shopping gets simpler with Universal Commerce Protocol updates — Ashish Gupta / Google (2026). Google
Frequently Asked Questions — Google (2026). Google for Developers
Checkout Capability — UCP Authors (2026). Universal Commerce Protocol
AP2 Mandates Extension — UCP Authors (2026). Universal Commerce Protocol
Order Capability — UCP Authors (2026). Universal Commerce Protocol
「AI vs AI」時代の消費者保護法制 — 得津晶 (2026). Cabinet Office, Consumer Commission, 14th expert panel meeting, Material 1
Questions to Reflect On
- Which countries, languages, surfaces and comparison brands does the metric cover?
- Can you compare the original request with the terms the buyer actually reviewed?
- Whom can the buyer contact about delivery or a refund after the order is placed?
Related Content
Data Literacy for Social Impact: A Practical Guide to Japan's e-Stat and RESAS
A foundational guide to statistical literacy for NPOs and civic organizations, covering the basics of reading data critically and practical use of Japan's free government data portals: e-Stat, RESAS, and jSTAT MAP.
Risk Management in the Age of AI Agents — OWASP Agentic Top 10 and Practical Countermeasures
As AI agents act autonomously, ten new risk categories have emerged that traditional security measures cannot fully address. Based on the international standard "OWASP Agentic Top 10," this guide provides actionable checklists for nonprofit staff, small business owners, and public sector employees.
Series: AI Governance & Risk Management
Can Vocational Training Be Measured? — EBPM and the Evaluation Design of Workforce Development Policy
Japan invests billions in vocational training yet lacks rigorous impact measurement. An EBPM-based evaluation design guide with international comparisons.
Series: Evaluation & Measurement Frameworks
Introduction to EBPM — What Evidence-Based Policymaking Asks of Nonprofits
As governments demand evidence from nonprofits, what competencies are expected? This guide explains EBPM basics and the path forward for organizations.
Series: Evaluation & Measurement Frameworks
Previous
After Fill Removal, How Can the Public Verify Cost Recovery?
Next
What Does Japan's 74.4% School Reuse Rate Count?
Get new columns by email
1-2 social structure analysis columns per week. Free to subscribe.
Join ISVD's activities?
Sign up to receive the latest research and activity reports. Feel free to reach out about collaboration or project participation.
Support ISVD’s research and activities
Support ISVD’s research and publications with JPY5,000 per year. There is no automatic renewal.