AI Agent Readiness Audit: Japan's 42 Largest Retailers — Six Are Ready, None Got There Alone

Forty-two retailers, and what readiness actually looks like

As AI agents start doing the shopping, how far along are Japanese merchants? Nobody had the cross-sectional data to answer that. Coverage consisted of translated overseas case studies, or write-ups of announcements from Rakuten and Yahoo! Shopping.

So on 1 August 2026 we audited 42 of Japan's largest e-commerce and retail operators: six marketplaces, eight electronics chains, eight apparel retailers, seven grocers, seven beauty brands and six specialty retailers.

Start from the premise that there are three routes into agent-mediated purchasing: route 1 completes the purchase inside the AI, route 2 gets you discovered by an external AI and sells on your own site, and route 3 closes off outsiders and serves customers with your own AI. Being ready means a transaction actually completes on one of them. Sorting the 42 gives this:

Six companies are ready, and all six arrived on a platform's back. Not one of the 42 built agent readiness under its own steam.

How the six got there

Look inside those six and the phrase "on a platform's back" becomes concrete.

Rakuten and Yahoo! Shopping made agent purchasing work within their own marketplaces. Rakuten reports average order value up roughly 26% since its full rollout in December 2025, and Yahoo! Shopping says agent-mediated GMV has passed 15% of the total. Neither waited for an external standard. Both closed the loop with their own payments and fulfilment, which is route 1.

ZOZO and Mercari shipped official apps inside ChatGPT, on 9 March and 23 June 2026 respectively. Rather than making their own sites machine-readable, they carried the catalogue to where shoppers already are. That is also route 1, and it is the same pattern Sephora and Walmart follow on ChatGPT.

Amazon.co.jp went the opposite way and closed the door. Its robots.txt names and denies twelve of the fourteen AI crawler tokens we tested, and the buying experience completes inside Alexa for Shopping. Being unreadable from outside is not an implementation gap here. It is route 3 being executed on purpose.

Seijo Ishii alone had an external protocol in place, which is route 2 — though as the next sections show, the company did not put it there.

What the six share is that none of them opened their own site to external AI by their own hand. A marketplace surface, the ChatGPT surface, an in-house AI, or a platform's automatic switch. Every one of them sits on a road somebody else built. How each domestic player is positioned strategically is covered in the map of Japan's AI commerce landscape.

Measuring the other 36 — five stages seen from outside

From here we look at the 36 companies that are on no route at all. The yardstick is a five-stage scale describing how far a site is open to an external AI agent.

We checked six things: how robots.txt treats AI crawlers, whether protocol profiles such as /.well-known/ucp exist, and whether product pages carry structured data. Then whether prices survive with JavaScript disabled, whether an MCP endpoint is published, and whether sitemaps include product URLs. The raw data behind each score, and every place a human corrected the automated result, is on record.

The distribution — most stop at the entrance

Here is how the 36 fall out.

Twenty companies, 55.6%, sit at stage 0. An external AI agent can retrieve neither their products nor their prices. Forget agent-mediated checkout: these sites are not even at the door where an AI search answer could name their products.

The other striking feature is the hole between stage 2 and stage 4. Seven companies have structured data in place — Yodobashi Camera, Muji, Loft, Askul, @cosme SHOPPING, DHC and Rakuten Seiyu Netsuper. That is not a small number. Yet not one of them has gone on to publish a feed or an MCP endpoint.

We do not think that is a coincidence. Structured data is an extension of existing SEO work, well within what a marketing owner can decide. Publishing a feed or standing up an MCP server requires a separate decision to support AI agents, with a different budget and a different owner. In Japan today, nobody has crossed that line unaided.

They are not blocking you. They are rendering in JavaScript

To see why sites stop at stage 0, we looked at the 23 companies that cannot be read from outside — a group that includes Amazon, ZOZO and Mercari from the ready side. What we found was not what we expected.

JavaScript dependency was the most common cause, at thirteen companies. Only three block AI crawlers explicitly in robots.txt.

Concretely: Edion's search results page returns HTTP 200 and 20.8 KB of HTML. Inside it there is not one product link and not one price string. Everything renders after JavaScript executes. Nitori's search results behave the same way, returning a 15.5 KB shell.

Most AI crawlers do not execute JavaScript. These sites are effectively unreadable without having blocked anything. It is a side effect of building a single-page application, and in most cases nobody ever considered it. Closed without meaning to close is where Japanese e-commerce currently stands.

Bot mitigation accounts for seven more. Bic Camera and United Arrows accept the connection and then have the server terminate it; ZOZOTOWN returns 403 even to browser-identical headers. Controls deployed against scalpers and scrapers reject AI agents with equal efficiency. Note that ZOZO is also one of the companies that is ready, via its ChatGPT app — a reminder that an unreadable site and an unready business are not the same thing.

One of the three explicit blockers is Amazon.co.jp, from the previous section. That is a strategy being executed, which makes it a different animal from the other two, whose reasoning is not visible from outside.

The only route-2 arrival was built by Shopify

Of the six that are ready, only Seijo Ishii had an external protocol in place. It is worth looking at closely.

Open /.well-known/ucp on the company's domain and a Universal Commerce Protocol profile comes back. Cart, checkout, fulfilment and catalogue search are all declared, along with Google Pay as a payment handler and an MCP endpoint for agents to connect to.

It is unambiguously transaction-capable. The MCP endpoint, however, resolves to seijoishii.myshopify.com. In other words, the profile was generated by Shopify, not by Seijo Ishii.

To test that reading, we probed a separate sample of seventeen Japanese-language stores.

All six Shopify stores carried the same profile version: Seijo Ishii, Yogibo, ARC'TERYX Japan, Tsuchiya Kaban, COHINA and KURAND. They differ in category and in scale, yet their readiness is identical. None of the eleven non-Shopify stores had anything.

The reason is that Shopify enabled Agentic Storefronts across its eligible merchant base on 24 March 2026 — and did it on an opt-out basis. Merchants did not apply; it turned on unless they switched it off. So these six did not evaluate agent readiness and adopt it. They became ready without noticing.

That leads somewhere uncomfortable for Japanese merchants. What decides agent readiness right now is not your investment decision. It is your platform choice. On Shopify you are covered without lifting a finger. Off it, you build everything yourself. The gap opens regardless of how hard your team is working.

Electronics retail is furthest behind

Segment by segment across all 42, the differences are stark.

Six of eight electronics chains sit at stage 0, with only Yodobashi Camera reaching stage 2. These are companies selling on model numbers and prices, the most structurable data in retail.

The finding matches a separate audit we ran in July 2026, where only one of nine major electronics brands had implemented Product and Offer structured data. The segments that lagged on discovery lag on transaction too.

Apparel, by contrast, gets five of eight to stage 1, and specialty retail gets three of six to stage 2. The segment gap seems to be driven less by what they sell than by how they built the site. Read the marketplace row with a discount, since it contains companies that are ready by another route, such as Amazon and Mercari.

Almost nobody has stated a position

Even at the earlier stage of simply deciding a policy, we found a blank.

There is a file called /llms.txt, a place to write down how you want AI agents to treat your site. Two of the 42 published one, and Seijo Ishii's was generated automatically by Shopify. Just one company wrote its own: Qoo10, operated by eBay Japan. Titled "LLM Access & Citation Policy," it permits AI assistants to access and cite public pages while requiring attribution and freshness checks.

Only nine companies name AI crawlers explicitly in robots.txt. The remaining 33 leave it to the wildcard rule, which means they have expressed neither permission nor refusal. Open up, close down, or leave it to the marketplaces — from outside, there is no trace of any of those decisions having been made.

What this audit does not tell you

We observed public surfaces and interviewed no one. We sampled at most three product pages per site, which does not guarantee site-wide implementation. We also did not render JavaScript, so "no price in raw HTML" does not rule out retrieval by a JavaScript-executing crawler.

The "ready" judgements rest on published evidence. The in-marketplace agents at Rakuten and Yahoo! Shopping, the ChatGPT apps from ZOZO and Mercari, and Amazon's Alexa for Shopping all have official announcements behind them. The flip side is that an external audit cannot detect AI purchasing built into a company's own app and never announced. ACP support is likewise undetectable, because the specification has no settled public path. These figures are a snapshot taken on the audit date.

Three checks you can run today

This audit observed sites from the outside. To know precisely where you stand, run these three in order. Each takes minutes.

Structured data and feeds can wait until after those three. For a stage-0 site, what works right now is not adding structured data but becoming readable again. Get the order wrong and you are adding schema to a page nobody can fetch.

Frequently asked questions

Conclusion

Auditing 42 Japanese retailers produced three findings.

First, agent-mediated purchasing works at six companies, and every one of them arrived on a platform's back. A marketplace surface, a ChatGPT app, an in-house AI, a platform's automatic switch. The shapes differ, but no company built the road itself.

Second, the remaining 36 are on no route at all. More than half — 20 companies — stop before an external AI can read the product, and the leading cause is not deliberate blocking but JavaScript dependency. Closed without meaning to close, across every segment.

Third, read the other way, this is a wide-open position. Seven companies reached structured data and not one went further unaided. Clearing that single step would put a retailer well ahead of the domestic field. For the broader structure of agentic commerce, see what agentic commerce is.

Start by opening your own /.well-known/ucp and loading a product page with JavaScript off. Those two checks tell you where you stand.

Stellagent offers a free AI visibility audit for retail and e-commerce operators, using the same method described here.