Agentic commerce, read one post at a time

"Agentic commerce" is used for three different things, and the argument in most rooms comes from people defending different ones. Sorted by what they demand of a store, they are easy to keep apart.

The vocabulary that follows is worth ten minutes because every vendor conversation you have this year will assume it.

The protocols, without the war metaphor

Two protocols carry commerce. ACP, the Agentic Commerce Protocol, came from OpenAI and Stripe in September 2025 and powered Instant Checkout inside ChatGPT, starting with Etsy. Stripe's own president described the first mainstream implementation plainly when Instacart went live that December.

Stripe is partnering with @Instacart to enable direct checkout in ChatGPT. This new AI commerce experience is powered by the Agentic Commerce Protocol (ACP) that we launched with @OpenAI a few months ago.

Patrick Collison, co-founder, Stripe — X, 8 Dec 2025

UCP, the Universal Commerce Protocol, launched at NRF on 11 January 2026, co-developed by Google and Shopify with Etsy, Wayfair, Target and Walmart, and endorsed on day one by more than twenty companies including Visa, Mastercard, Stripe, PayPal and Adyen. Its structure is four layers — services, capabilities, extensions, transports — and the transports are the interesting part: REST, MCP and A2A.

The two are not symmetrical, and the comparison that circulated in January is still the clearest summary: ACP is agent-led commerce inside one surface; UCP is a shared rail meant for many surfaces.

Two more acronyms keep appearing and neither is a commerce protocol. MCP, the Model Context Protocol, is how an agent calls a tool and reads structured catalog, price and inventory data; it originated at Anthropic and now sits under the Linux Foundation. A2A is agent-to-agent messaging. Both are plumbing that the commerce protocols run on. AP2, Google's Agent Payments Protocol, handles rules-based agent payment and rides alongside UCP as an extension. One cheat sheet that travelled widely put the relationship in a sentence worth keeping.

MCP is how agents access your product catalog, pricing, inventory, and specs — verified and structured, not scraped from your website at 3am 🕐.

Joshua Hellewell — LinkedIn, 8 May 2026

Which protocol the market was talking about, by quarter

Posts in this corpus naming each protocol. UCP did not exist before January 2026; within one quarter it was named more than ACP had been in two.

UCPACPMCPAP2

Named but secondhandCounted across the 1,500 selected posts. This is share of conversation in a sampled corpus, not market adoption.

Show data table

	UCP	ACP	MCP	AP2
2025-Q3	0	4	0	0
2025-Q4	0	27	6	1
2026-Q1	132	40	19	9
2026-Q2	62	14	21	13
2026-Q3	66	23	28	9

What a merchant is actually asked to publish

Strip the keynotes away and the technical requirement is a short list. A JSON manifest at /.well-known/ucp declaring which capabilities you support. An MCP endpoint, on Shopify at /api/ucp/mcp. Product data in real taxonomy fields with attributes as attributes. Policies and reviews in JSON-LD rather than footer prose. Shopify pushed the manifest live on every store within two days of the UCP launch, which is why a merchant can look at a competitor's file today.

The instruction file beside it is written for machines and is worth reading once in full, because it is the part most merchants have never opened.

Businesses publish the services they support and corresponding capabilities in a standard JSON manifest located at /.well-known/ucp.

Juozas Kaziukėnas, ecommerce analyst — LinkedIn, 13 Jan 2026

Where the liability sits, which is the part that matters

Both protocols keep you as merchant of record. That is not a footnote; it is the whole risk position. Refunds, chargebacks, tax collection and the customer data relationship stay with the store. What changes is who is on the other side of the transaction.

Merchant of record is where chargeback liability attaches. That hasn't changed. What's changed is that the buyer is now a probabilistic system, and most merchant order records weren't designed to capture what the human actually authorized the agent to do.

Richard E., payments — LinkedIn, 9 Sep 2026

Multi-merchant carts are the clearest illustration that nobody wants that liability. A shopper buying from three stores in one conversation still triggers three separate checkouts under both protocols, deliberately, because centralising payment, shipping and returns would make the agent a marketplace.

The identity layer is consolidating faster than commerce

While the commerce protocols competed, the card networks built agent identity separately — Visa's Trusted Agent Protocol in October 2025, Mastercard's Verifiable Intent, Ant International's Agentic Mobile Protocol — and then merged the interfaces. On 10 September 2026 Visa, Mastercard and Ant announced an interoperable Know Your Agent framework: who operates an agent, whom it represents, whether it meets certification standards, and whether its behaviour is monitored.

The mechanism underneath is a token that binds an agent to the cardholder's issuer, so an agent can be recognised at any merchant rather than at one.

This is the layer to watch, because it is where the answer to "was this purchase authorised?" will be decided — and, as the third agentic-commerce chapter shows, that question has already arrived in the form of two people asking publicly for their money back.

What the word does not cover

Two things get filed under agentic commerce that belong elsewhere. The first is merchant-side automation — using Claude Code or Shopify's AI Toolkit to rewrite product data, audit a store or monitor stock. It is real, it works today, and it has nothing to do with these protocols; it is covered in the third agentic-commerce chapter. The second is machine-to-machine payment between agents with no human anywhere. In this corpus it remains demo-stage, and most of the posts pushing it hardest were promoting a token, which is why they are not in this report at all.

The clearest way to read the year is by what a merchant can plug into, not by who announced what. This table is the whole chapter in one screen; the rest is evidence for it.

Google and Shopify: the standard, and the distribution

UCP was announced by both companies on the same day, and the Shopify framing is the one worth keeping, because it tells you what the protocol is for.

Shopify is building the foundation for agentic commerce. Universal Commerce Protocol, which we co-developed with Google, is now live. UCP will make it faster for agents and retailers to integrate. It's open by default, so platforms and agents can use UCP to start transacting

Tobi Lütke, CEO, Shopify — X, 11 Jan 2026 · the most-engaged post in this corpus

Google shipped the surface at the same time: checkout inside AI Mode and Gemini, with business tooling behind it. By February 2026 it was live for US shoppers on Etsy and Wayfair, which is what a completed agentic purchase looks like end to end.

Then the council grew, and the membership list is the actual story of the year. Amazon, Meta, Microsoft, Salesforce and Stripe joined in April — competitors, a rival protocol's co-author, and a company that was suing an agent maker at the time.

Who the market actually talks about

Posts naming each company, out of 1,500. Shopify appears in nearly half — the platform is the subject, not just the venue.

Named but secondhandName matching across post text in the 1,500 selected posts; a post naming three companies counts three times.

Show data table

Shopify	687
OpenAI / ChatGPT	407
Google / Gemini	405
Anthropic / Claude	291
Amazon	174
Stripe	141
Visa	115
Walmart	113
Microsoft / Copilot	95
Mastercard	91

OpenAI: the retreat that taught everyone else

Instant Checkout was the first in-chat purchase most people saw, and it did work. It was switched off around March 2026, roughly five months in. Reporting put the number of Shopify merchants that ever went fully live in the low double digits, out of millions. The named causes were not AI problems.

It failed on sales tax collection, fraud controls, and keeping inventory accurate across millions of listings.

Raul Goycoolea Seoane — LinkedIn, 2 Sep 2026

The number that ended the argument came from the largest partner. Walmart tested 200,000 items and measured conversion at roughly a third of its own site's.

What replaced it is the model most merchants will actually live in: the assistant finds the product, and the sale closes on your site. One operator's description of the new arrangement is uncharitable and useful — affiliate marketing with a chat interface on top.

Amazon: walls and bridges at the same time

Amazon blocked third-party shopping agents in robots.txt, issued an Agent Policy requiring agents to identify themselves, and pursued Perplexity over its Comet browser. It also built "Buy For Me" to shop other people's stores, and hired for the opposite posture.

In August 2026 the Ninth Circuit reversed the order blocking Comet, on reasoning that matters to every merchant writing an agent policy: an agent acting inside a user's own logged-in account is the user, not the agent maker.

The court's reasoning: when Comet shops inside your Amazon account, it's you accessing Amazon, not Perplexity.

Grayson Kinsella — LinkedIn, 11 Aug 2026

Anthropic: a blueprint, not a destination

Claude for Commerce arrived on 2 September 2026 as open source under Apache 2.0 — a shopping agent for a retailer's own site and a merchant agent for the back office, with Shopify, Visa, Mastercard, Accenture and Priceline listed as partners rather than rails.

We're open-sourcing Claude Commerce Agents. This is a blueprint for building shopping and merchant agents, with reference implementations across retail, travel, telecom, and entertainment.

ClaudeDevs — X, 2 Sep 2026

The merchant-side half is the part operators reacted to: a working dashboard where the agent drafts and a human approves.

Anthropic's own launch numbers — carts up to 35% larger, shoppers 60% more likely to complete — are repeated everywhere in this corpus and carry no sample size, no named retailer and no methodology. At least four independent analysts here say so explicitly. Treat them as vendor claims, which is how this report grades them.

Meta, Microsoft, and the surfaces nobody planned for

Meta's Muse launched in the US in September 2026 with three checkout modes — browser automation, Shopify over UCP with Shop Pay, and Stripe Link — and an explicit Allow/Deny before it spends.

Microsoft took the quieter path: Copilot Checkout, a catalog enrichment agent that reads attributes out of product images, and a seat on the UCP council rather than a fourth standard. For a merchant this is the good outcome — one manifest, several surfaces.

Asia runs a different race

Alipay launched what it calls a full-stack agentic commerce platform in Hangzhou in August 2026, built on its own AHA protocol, connecting merchant pages and workflows into agent-readable skills for its Ah Bao agent. Its reported scale — a billion connected users, 300 million-plus cumulative agentic transactions by May 2026 — is self-reported at launch and unverified here, but the strategic point does not depend on the number: China is not adopting UCP or ACP.

The payment rails reached this region before the shopping surfaces did. Mastercard has run authenticated agent transactions in Singapore, Malaysia and Korea with a phased commercial rollout ahead, and its September 2026 Wallet Pay launch plugs directly into Asian super-app wallets — AlipayHK, GCash, TrueMoney, KakaoPay. For a merchant selling across Southeast Asia, that is the part to watch: the card networks are ready in market before consumer demand is.

I assumed the third was years out because the payment rails didn't exist. Wrong. Visa and OpenAI announced their integration in June. Mastercard has run authenticated agent transactions in Singapore, Malaysia and Korea, with phased commercial rollout ahead.

Richard E., payments — LinkedIn, 9 Sep 2026

India's case is different again and rests on rails that already exist, which is why the argument travelled.

India is probably the only market where agentic commerce could work at scale right now. The UPI infrastructure is already there. Our habits already favor delegation. 950M people who never adapted to app-driven commerce.

Shashank Kumar, co-founder, Razorpay — X, 12 Mar 2026

Pine Labs and OpenAI bought a full-page newspaper advertisement to launch what they called India's first agentic commerce stack, in February 2026. Brazil shows the same pattern from the seller side: a Mercado Livre seller agent reporting the week's revenue and a queue of listings that need a human.

For an APAC audience the useful read is that the rails arrive locally, not from a single Western standard — and that the merchant-side work, catalog and policy, is identical in every one of these markets.

Three postures, and how to pick one

Every merchant in this corpus is implicitly taking one of three positions. The mistake is taking one by accident.

The evidence for starting at "readable" is that the same data work keeps showing up as the thing that moved a number. One DTC brand had a live agentic feed for months and saw nothing change until it moved product-page content to server-side rendering, because OpenAI's crawler does not execute client-side JavaScript.

And the reason to do the work even if you are sceptical: agents rank on what your data says, not on how big your brand is. A CEO testing his own category found both Claude and ChatGPT ranking competitors ahead of Arc'teryx because the competitors' structured data answered the question that was asked.

Zero results. Arc'teryx makes that jacket. Their search box handed me a sizing guide and the returns policy. Then I asked Claude and ChatGPT the same question. Both came back with real jackets, ranked, with reasoning, across five brands.

Mike Micucci — LinkedIn, 5 Sep 2026

The 90-day version

Compiled from the readiness checklists in this corpus, in the order practitioners actually run them. None of it requires believing any forecast.

1. Read your own agent files

   Open /agents.md, /llms.txt and /robots.txt on your own domain. On Shopify these exist whether or not you wrote them — and the stock version recommends the Shop app over your storefront. An audit of the 100 biggest Shopify stores found most still serving it untouched.

2. Scan with two tools, not one

   Shopify's own commerce-readiness.shopify.io and one third party. Treat the gap between the scores as the interesting part.

3. Move attributes out of prose

   Material, fit, compatibility, fragrance-free, dimensions — into taxonomy fields and metafields. Prose is for humans; fields are what an agent intersects with a query.

4. Put policies and reviews in JSON-LD

   Scanners treat missing structured markup as blocking even when the same text is visible on the page.

5. Check price and stock accuracy in the feed itself

   Not in the admin. A recommendation for an out-of-stock item is worse than no recommendation.

6. Server-side render product pages

   If they are client-rendered today, this is the change that moved the number for the brand above.

7. Search for your own products monthly

   In ChatGPT, Gemini and Perplexity. Note whether the buy link points at you or at a reseller.

8. Give it an owner

   Five teams touch the catalog and usually none owns how agents read it.

9. Report agent sessions separately

   From search and social, with their own conversion and AOV lines. This is the only way you will know whether any of the above worked.

10. Write the policy

    The five decisions in the next section, before you need them.

Those two scores are the same storefront. Manage the underlying gaps — structured data, readability — and let the number be a by-product.

Decide the liability questions while they are still cheap

Two public incidents inside one week in September 2026 show what "autonomous" means in practice when nothing was decided in advance: a travel agent that cancelled a flight before showing the fee, and a booking agent that charged twice.

Neither case involved a merchant doing anything wrong, and both landed in a support queue somewhere. The five decisions worth writing down now, drawn from the governance material in this corpus:

• Who is liable for an unauthorised agent purchase on your store, and where does that sentence live? Target's terms already put it on the shopper. Silence is also a choice, just not one you made.
• Which agents may access your store — open, allow-list, or blocked — and does robots.txt match what you told the sales team? eBay, Amazon, Walmart and Shopify have each landed somewhere different.
• What evidence do you require for an agent-initiated order: agent identity, the scope the human authorised, a confirmation record. Without it, disputes are unwinnable.
• Will your fraud rules misclassify legitimate agents? "Bot equals fraud" stopped being true. Meanwhile Visa's VAMP dispute threshold tightened to 1.5% in April 2026, so the cost of getting this wrong went up.
• Who answers the customer when an agent orders the wrong thing to the wrong address? Write the runbook before the first ticket.

The use case that already works: the agent on your side of the counter

The most reliable agentic-commerce win in this corpus has nothing to do with shoppers. Shopify's AI Toolkit, launched 9 April 2026, connects Claude Code, Codex, Cursor and others to a live store's Admin API, and operators use it for catalog rewrites, dead-stock flags, margin monitoring and audits. Every practitioner account that reads as real insists on the same guardrail: the agent drafts, a human approves the write.

The reason for that guardrail is measurable. On CommerceAgentBench — 107 real ecommerce workflows — the best model finishes 61.7%.

What the loudest posts say, and what survives reading them

Ranked by engagement, this corpus is not led by protocol news. It is led by "Claude Code plus Shopify" threads, and the same author often appears three or four times with the same structure. They are worth naming because they set expectations you will be measured against.

Against that, the quiet posts hold up better. A store owner posting a dashboard. A consultant posting a before-and-after with the deploy date on it. An analyst posting the second chart from the same research deck — the one showing the same AI referrals at 0.8% to 1.4% of total traffic.

Both charts come from the same deck. Only the first one gets screenshotted, and that single editorial choice explains most of the distance between what merchants expect and what their analytics show.

The one-sentence posture

Do the catalog work now because it pays without agents; expose checkout when your own analytics show agent traffic arriving; decide the liability questions before an incident decides them for you; and treat every percentage in a launch post as a claim until someone names the retailer, the window and the sample.

The second pass: the agentic-commerce corpus (Chapters 12–13a)

Chapters 12, 13 and 13a were built later and from a separate sweep, run on 14 September 2026, because the first corpus covered agentic commerce as one surface among fifteen and the subject had outgrown that share. Three pulls of LinkedIn and X returned 11,908 raw items, 9,105 unique after de-duplication. Two buckets were defined: posts whose own text contains the phrase "agentic commerce", and posts that do not but report a named platform's move in AI shopping, agent checkout or merchant AI tooling. Only posts carrying at least one image were considered, which narrowed the field to 3,780 candidates.

Those candidates were read one at a time — 2,728 of them — and graded on two axes: is the post or its author promoting crypto, and is the post substantive about agentic commerce or merely brushing against it. 249 were dropped as crypto and 587 as incidental (hiring ads, event promos, hashtag stuffing, stock chatter, and a large family of templated "I built a store with Claude in an hour" posts). The remaining pool was ranked and cut to 1,500: 500 phrase posts and 250 platform-move posts from each network.

All 1,840 attached images were opened and classified individually; 1,032 carried enough legible information to be usable and 31 are printed in these three chapters. Quotes in these chapters were machine-checked against the source post text: 26 that turned out to be paraphrase rather than verbatim were sent back and rewritten, and one claim that carried a figure absent from its cited post was removed outright.

Two differences from the first pass are worth stating plainly. Crypto was excluded far more aggressively here, because on X roughly a third of everything using the phrase was promoting a token — a stricter rule also cost the corpus a handful of legitimate posts whose authors work at crypto-native payment companies. And the monthly post counts in this second corpus cannot be read as market volume: the X sweep took the top posts per month and the LinkedIn sweep ranked by relevance, so the shape reflects the sampling, not the discourse.

From 11,908 raw items to 1,500 posts

Each step is a filter someone can disagree with, so each one is stated.

First-party / researchCounts produced by the pull script and the grading files kept with the corpus.

Show data table

Raw items pulled	11,908
Unique posts	9,105
Candidates with an image	3,780
Read and graded	2,728
Selected	1,500

What this method cannot do

• It measures discourse, not the market. A corpus of posts over-represents whoever posts: vendors, consultants, agencies and people with something to sell. The merchant quietly shipping a catalog fix is absent.
• It cannot audit a first-party claim. When Anthropic says carts run 35% larger, this report records who said it, when, and what was not disclosed. It cannot check the number.
• Counts by month are sampling, not volume. The X sweep took top posts per month; LinkedIn ranked by relevance. Neither is a census.
• It is English-first. Non-English posts were kept when relevant — the corpus includes Japanese, Portuguese and Chinese material — but the queries were English, so Asia is thinner here than it is in reality.
• Six months is short. Several things in this report changed twice inside the window. A protocol version, a fee, a policy: check the date on anything before acting on it.

Tools

Posts and images were pulled through Apify actors (harvestapi/linkedin-post-search, apidojo/tweet-scraper) at a cost of $8.54. Reading, grading, image review and drafting were done with Claude. The pull script, the grading rubric and every verdict file are kept alongside the corpus, so the same cut can be rebuilt or re-graded with a different rule.

The evidence ledger

Every number quoted in these chapters, with who measured it and how solid it is. Sort your own reading by the grade column: a vendor launch claim and a clickstream panel are not the same kind of fact.

128 numbers pulled from the 1,500 posts, de-duplicated: 28 first-party or research, 66 named but secondhand, 34 vendor-reported.