- ACP came from OpenAI and Stripe as a checkout and delegated-payment rail, public since September 29, 2025, still marked beta. [1][7]
- In March 2026 OpenAI said the first Instant Checkout lacked flexibility, let merchants keep their own checkout, and expanded ACP to discovery. [8]
- UCP came from Google and Shopify at NRF on January 11, 2026, scoped as the whole journey and starting with native checkout. [3][4]
- Both are Apache 2.0. Implementing is open. Getting a feed into ChatGPT is for approved partners, by application. [4][7][9][10]
- Shopify merchants are UCP-enabled by default, and their Catalog data is already in ChatGPT with no extra work. [6][8]
- OpenAI’s feed spec recommends full snapshots at least daily, with updates through the day via API. [2][9]
- The 15-minute figure comes from third-party mirrors and describes how often updates are accepted, not required. [11]
- What ranks a product is feed data: required fields for price and availability, then media, reviews, and performance signals. [1]
The standards-war story writes itself, which is usually the warning sign. Two of the largest companies in the world, two protocols, two camps, pick a side before the window closes.
Read both specifications, and the announcement OpenAI made in March, and the drama thins out considerably. These are two documents that started at different points in the same journey and have spent 2026 growing toward each other. The thing they both demand from a merchant is neither exotic nor new.
Where each one started, and where each one moved
The Agentic Commerce Protocol arrived from OpenAI and Stripe as a way for an agent to complete a purchase. Its repository was created on September 29, 2025, and still marks the specification beta. [7] OpenAI’s documentation describes what supporting Instant Checkout in ChatGPT asks of a merchant: three flows. [1]
Share a product feed
A regularly refreshed catalog carrying identifiers, descriptions, pricing, inventory, media, and fulfillment options, validated from an initial sample. The live file spec prefers parquet, accepts gzipped CSV or JSONL, and is pushed over SFTP. [1][2]
Handle orders and checkout
ChatGPT collects buyer, fulfillment, and payment details, then calls your endpoints to create or update a checkout session. You validate, determine fulfillment, calculate tax, run risk, charge through your existing processor, and return accept or decline. [1]
Take delegated payment
OpenAI prepares a one-time delegated payment request with a maximum chargeable amount and an expiry. The payment provider answers with a token that OpenAI passes to you. What is scoped down is what the token can be used for, not who holds the details. Merchant of record, the party that legally sells and carries the tax, refund, and chargeback liability, stays with you. [1]
Then, on March 24, 2026, OpenAI changed the emphasis in public. It said the initial version of Instant Checkout did not offer the level of flexibility it aspired to provide, so it would let merchants use their own checkout experiences while it focused its efforts on product discovery, and it was expanding ACP to support discovery so that more complete and current product information reached ChatGPT directly. [8] The checkout spec did not disappear. The universal ask did narrow to flow one.
The Universal Commerce Protocol started at the other end. Sundar Pichai introduced it at NRF on January 11, 2026, describing it as open, agnostic, built with Shopify, Etsy, Wayfair, Target, and Walmart, endorsed by more than 20 others, and available that day. He was also clear about sequencing: there is a lot UCP will enable, and Google was starting with native checkout. [3] Its own site calls it the common language for platforms, agents, and businesses, spanning the entire commerce lifecycle from discovery and search to final sale and post-purchase support, with catalog search and lookup, cart building, identity linking, checkout, and order management as its named capabilities, built-in support for the Agent Payments Protocol, Agent2Agent, and Model Context Protocol, and an Apache 2.0 license. [4]
| ACP | UCP | |
|---|---|---|
| Came from | OpenAI and Stripe | Google and Shopify |
| Public since | September 29, 2025 [7] | January 11, 2026 [3] |
| Started as | Agentic checkout and delegated payment | The full lifecycle, shipped checkout first |
| Where it moved | Discovery, with merchants keeping their own checkout, plus cart, feed, and orders in its April 2026 spec release [8][7] | Catalog search and cart building alongside checkout, with Shopify merchants enabled by default [4][6] |
| Published components | Product feed, agentic checkout spec, delegated payment spec | Catalog search and lookup, cart building, identity linking, checkout, order management |
| Verticals named | Shopping | Shopping, with lodging and food listed as coming |
| License | Apache 2.0, marked beta [7] | Apache 2.0 [4] |
| Where it shows up | ChatGPT shopping surfaces | Google surfaces and Shopify merchants by default |
Read the two middle rows together and the layer story dies. By summer both rails describe a feed at one end and a checkout at the other, and the same companies sit on both. Target and Wayfair were named as UCP co-builders at NRF and appear on OpenAI’s list of retailers integrated into ACP for discovery. [3][8] Stripe maintains ACP and is named on UCP’s site. [7][4] Shopify merchants are UCP-enabled by default, and OpenAI says their product data is already integrated into ChatGPT through Shopify Catalog with no additional work from the merchant. [6][8] They are still not substitutes, because supporting one does not imply support for the other and each surface runs its own admission. They are converging.
The cadence everyone is quoting is not in the spec
Search for advice on this topic and you will be told, repeatedly and confidently, that agentic commerce demands a product feed refreshed every 15 minutes. The line is traceable. Third-party mirrors of the feed specification carry a refresh frequency section that reads: our system accepts updates every 15 minutes. [11] Accepts. That sentence describes a ceiling on how often you may send, and it has been repeated as a floor.
What OpenAI’s specification says: the file-upload documentation describes a full snapshot feed treated as the source of truth, with a recommended cadence of at least daily, and tells merchants to publish full snapshots on a predictable cadence. [2]
What its onboarding guide adds: provide the entire feed once a day via file upload, then send updates throughout the day via the API. The products and promotions endpoints upsert partial changes for when a price or a stock level needs to move sooner. [9][5]
What is not there: any 15-minute requirement, in either path.
This matters because the two claims imply completely different projects. A daily full snapshot plus an upsert endpoint for exceptions is a job your existing feed pipeline can probably already do. A hard 15-minute full-catalog refresh is an infrastructure program with a budget attached. One of those is in the documentation and the other is in blog posts about a mirror of the documentation.
The accuracy point underneath the myth is still real, and it is worth separating from the number. A feed that says a thing is in stock when it is not produces a canceled order inside an assistant that will carry that outcome into how it treats you next time. The cost of being wrong went up. The required refresh rate did not.
The part nobody wants to hear
Here is what both protocols ask a merchant to be good at, stated plainly by OpenAI: required fields ensure correct display of price and availability, while recommended attributes such as rich media, reviews, and performance signals improve ranking, relevance, and user trust. [1]
Read that sentence looking for the things a marketing team normally controls. There is no ad unit. No landing page. No headline, no hero image of your own choosing, no brand moment, no retargeting. The levers are identifiers, accuracy, completeness, media quality, and review data.
Which means agentic commerce is not a new channel to learn. It is the same shift that already happened in Shopping and Performance Max, arriving one layer higher, and the discipline it rewards is the one where the feed is the targeting rather than a data-entry chore. The teams that will do well here are the ones who already treat catalog quality as a revenue system with an owner, a monitor, and a rollback.
Access is the second thing to be clear-eyed about, and it splits cleanly. Implementing ACP is open to any business, under Apache 2.0. [10] Getting a feed into ChatGPT is not: OpenAI’s onboarding guide says onboarding product feeds in ChatGPT is currently available to approved partners, by application, and the protocol site says each AI platform manages its own process for participation. [9][10] On the other side, a Shopify merchant is on UCP by default and in ChatGPT through Catalog without lifting a finger. [6][8] So the practical question is not which protocol wins. It is whether your platform has done the work for you, and if not, whether you have applied.
What to actually do this quarter
Stop treating this as a bet on a standard, because the standards are converging and the bet does not resolve. Audit the catalog against the required fields in both specifications. The data is one asset: identifiers, titles, descriptions, media, price, and availability. The delivery is two integrations, because ACP takes a pushed feed while UCP’s catalog capability answers search and lookup calls, and only one of those is a file you export. Fix accuracy before you fix coverage, because a wrong price is worse than a missing product once an agent is transacting on it. Establish who owns feed correctness by name, with a monitor that pages someone, in the same way the number you feed an automated system stopped being a reporting preference and became the thing that decides the outcome.
And measure the surfaces, not the protocol. Whether ACP or UCP is winning is a question for people who write roundups. Whether your products are appearing, and being bought, inside the assistants your customers actually use is a question your catalog answers on its own.
Sources
- OpenAI · Agentic Commerce Protocol: key conceptsThe three flows behind Instant Checkout, the delegated payment mechanics, and what the feed fields decide.
- OpenAI · Product feed via file upload: overviewFull snapshot as source of truth, recommended cadence at least daily, formats and SFTP delivery.
- Google · Sundar Pichai’s remarks at NRF 2026The UCP announcement, January 11, 2026, with named partners and the line that Google was starting with native checkout.
- Universal Commerce Protocol · ucp.devScope, the five named capabilities, related protocols, verticals, and the Apache 2.0 license.
- OpenAI · Product feed API: overviewFeeds, products, and promotions endpoints, with partial upserts.
- Shopify · Spring ’26 Edition for merchantsShopify merchants are UCP-enabled by default, and eligible products are in Catalog by default.
- GitHub · agentic-commerce-protocolThe ACP specification itself: Apache 2.0 in the LICENSE file, maintained by OpenAI and Stripe, status beta. Repository created September 29, 2025.
- OpenAI · Powering product discovery in ChatGPTMarch 24, 2026. The Instant Checkout flexibility statement, the expansion of ACP to discovery, and Shopify Catalog integration with no merchant work. Verified in a browser, the host declines automated requests.
- OpenAI · Agentic Commerce Protocol: get startedFeed onboarding available to approved partners by application, and the once-a-day file plus API updates recommendation.
- Agentic Commerce Protocol · agenticcommerce.devOpen source under Apache 2.0, any business can implement, and each AI platform manages its own participation.
- Agentic Commerce Protocol mirror · Product feed specification (third-party mirror)The origin of the 15-minute figure: a refresh section stating the system accepts updates every 15 minutes. Not an OpenAI page.
Frequently asked questions
What is the Agentic Commerce Protocol?
ACP is the open standard maintained by OpenAI and Stripe, public since September 29, 2025, for connecting buyers, their AI agents, and businesses. OpenAI’s documentation describes three flows behind Instant Checkout in ChatGPT: sharing a product feed, handling orders and checkout through the Agentic Checkout Spec, and delegated payment. In March 2026 OpenAI expanded the protocol to product discovery.
What is the Universal Commerce Protocol?
UCP is an open standard co-developed by Google and Shopify, announced by Sundar Pichai at NRF on January 11, 2026, and available the same day. It describes itself as the common language for platforms, agents, and businesses, covering the commerce lifecycle from discovery and search through to sale and post-purchase support. Pichai said Google was starting with native checkout.
Are ACP and UCP competitors?
They overlap more every quarter, which is the point. ACP launched as checkout and moved toward discovery. UCP was scoped as the full journey and shipped checkout first. Target and Wayfair sit on both: Google named them as UCP co-builders and OpenAI names them among the retailers integrated into ACP for discovery. Stripe maintains ACP and is named on UCP’s site, and Shopify co-developed UCP while its Catalog already feeds ChatGPT. Supporting one does not imply supporting the other, but the merchant asset both read is the same catalog.
What happened to Instant Checkout?
OpenAI said in March 2026 that the initial version of Instant Checkout did not offer the level of flexibility it aspired to, so merchants may use their own checkout experiences while OpenAI focuses its efforts on product discovery. The checkout specification still exists. The universal ask that remains is the feed.
Can anyone implement ACP today?
The specification is open source under Apache 2.0 and any business can implement it. Appearing in ChatGPT is a separate step: OpenAI’s onboarding guide says onboarding product feeds in ChatGPT is currently available to approved partners, by application, and the protocol site says each AI platform manages its own participation process.
How often does OpenAI want the product feed refreshed?
At least daily. OpenAI’s file-upload specification describes a full snapshot feed treated as the source of truth with a recommended cadence of at least daily, and its onboarding guide recommends the entire feed once a day via file upload with updates through the day via the API. No 15-minute requirement appears anywhere in OpenAI’s documentation.
Where does the 15-minute figure come from?
From third-party mirrors of the feed specification, where the line reads that the system accepts updates every 15 minutes. That is how often updates are accepted, not how often they are required. The distinction is the difference between a job your feed pipeline already does and an infrastructure program.
What does UCP cover beyond checkout?
Its site names five capabilities: catalog search and lookup, cart building, identity linking, checkout, and order management, organized into verticals for shopping, with lodging and food specifications listed as coming. It also declares built-in support for the Agent Payments Protocol, Agent2Agent, and Model Context Protocol.
Do I need to pick one?
Not at the level where the work happens. Both need a structured, accurate, machine-readable catalog with correct price and availability, which is the same asset. Pick where you spend integration effort based on which surfaces your buyers use, and treat the catalog as shared infrastructure rather than as a per-protocol project.
What actually decides whether an agent picks my product?
Data quality, not creative. OpenAI’s documentation says required fields ensure correct display of price and availability, while recommended attributes such as rich media, reviews, and performance signals improve ranking, relevance, and user trust. There is no ad unit, no landing page, and no brand impression in that sentence.
Mikołaj Salecki
Writes about media, tech, and AI business for people who actually run digital. Former agency lead. Skeptic of frameworks that read better than they perform.
