Japan's AI Commerce Landscape — Inside the Marketplaces, Agents Are Already Selling
Where Japan's AI commerce platforms actually stand. Rakuten reports average order value up roughly 26% after its full rollout, and Yahoo! Shopping says agent-mediated GMV passed 15% of the total. Meanwhile ZOZO and Mercari shipped official apps inside ChatGPT, so marketplace-internal completion and expansion onto external AI surfaces are now running in parallel.
Key takeaways
- Agentic commerce is already producing revenue in Japan. Rakuten reports average order value up roughly 26% after a full rollout, and Yahoo! Shopping says agent-mediated GMV has passed 15% of its total
- That revenue happens inside the marketplaces. Both companies keep agents within their own services rather than opening their catalogues to third-party AI
- Meanwhile ZOZO and Mercari shipped official apps inside ChatGPT. "Japan keeps everything inside the marketplace" no longer describes the market
Agents are already selling in Japan
Coverage of agentic commerce leans heavily American. Instant Checkout in ChatGPT, UCP from Google and Shopify, Amazon's manoeuvring. Most Japanese-language articles are translations of that coverage.
The numbers tell a different story. Transactions are already moving inside Japan, at a scale that is hard to dismiss.
| Company | Launched | What they did | Published results |
|---|---|---|---|
| Rakuten | December 2025 | Rolled out an agent spanning Ichiba, Fashion, Rakuma and Books | Average order value up roughly 26% |
| Yahoo! Shopping | February 2026 | Shipped an agent covering research through post-purchase | Agent-mediated GMV above 15% of total. Conversion 1.5x non-users |
| ZOZO | 9 March 2026 | Official WEAR and ZOZOTOWN apps via Apps in ChatGPT | No account required. Draws on 14M+ outfit posts |
| Mercari | 23 June 2026 | Listing search and seller drafting via Apps in ChatGPT | Multilingual search supported |
| LINE Yahoo (Agent i) | April 2026 | Umbrella brand merging Yahoo! JAPAN AI assistant and LINE AI | First extension automates multi-product comparison |
Yahoo! Shopping launched its agent in February 2026 and reports that agent-mediated GMV grew past 15% of the total, with users converting at 1.5 times the rate of non-users. Rakuten, following its December 2025 rollout, says average order value rose around 26%.
Inside the marketplaces, at least, "not yet in Japan" has stopped being true.Rakuten recommends across its whole portfolio
What distinguishes Rakuten's approach is that the agent is not a feature of one service.
The December 2025 rollout unified data across Rakuten Ichiba, Rakuten Fashion, Rakuma and Rakuten Books, which is what makes cross-service recommendation possible and what drove the order value increase. That is a different proposition from a recommender confined to a single storefront.
Expansion continued from there. In January 2026 the company embedded Rakuten AI into the Rakuten Ichiba smartphone app, putting the agent on the screen where shoppers spend most of their time.
Less noticed is that Rakuten also ships an agent to its merchants. A data analysis agent released in April 2026 was reportedly used for deeper analysis by 72.9% of participating stores. Running both a consumer-facing and a merchant-facing agent is a posture few platforms have adopted.
Underneath sits a model Rakuten built itself. Rakuten AI 3.0, announced in December 2025, is a 700-billion-parameter Japanese language model slated for deployment across the group's services. No other Japanese retailer owns the model layer as well.
Yahoo! Shopping has published the clearest numbers
LINE Yahoo has been the most specific about results.
Within months of the February 2026 launch, agent-mediated GMV passed 15% of the total and agent users were converting at 1.5 times the rate of everyone else. Those two figures establish that the agent is moving revenue, not just generating press.
The company has kept building. In April 2026 it launched Agent i as an umbrella brand, merging the Yahoo! JAPAN AI assistant with LINE AI. The first extension shipped under it compares multiple products automatically and recommends one.
What makes this structurally interesting is that the merger pulls LINE into the loop. LINE's reach in Japan is close to universal, and wiring a commerce agent into that surface has few parallels elsewhere.
Why the marketplaces keep it inside
What Rakuten and Yahoo! Shopping share is that neither is waiting for an external protocol.
UCP and ACP exist so that independent parties can transact using the same vocabulary. But both Japanese marketplaces run their agents inside their own services, settle with their own payments and ship with their own logistics. The need to conform to an external standard has never arisen.
A platform completing agent-mediated purchases entirely within its own service. Because catalogue, payments, fulfilment and identity are all in-house, external protocol support becomes unnecessary. We use the term to describe the posture Japan's large marketplaces have adopted.
It is a defensible choice. If you already own the traffic, the payments and the logistics, opening your catalogue to third-party AI invites fee negotiations and puts distance between you and the customer.
The contrast with the United States is sharp. Shopify moved toward making its merchants purchasable by external agents, distributing their catalogues outward. Japan's marketplaces have gone the other way, keeping the agent inside the walls to protect an existing economy.
That posture has a measurable limit. When we audited 42 Japanese e-commerce sites, both Rakuten Ichiba and Yahoo! Shopping scored only at the structured-data stage from an external agent's point of view. Active inside, indistinguishable from everyone else outside.
Two Japanese companies moved into ChatGPT
Others chose the opposite direction.
ZOZO launched official apps through Apps in ChatGPT on 9 March 2026. WEAR by ZOZO supplies outfit recommendations and ZOZOTOWN supplies item details, drawing on more than 14 million outfit posts to reason about occasion, weather and stated preference. No account is required, so any ChatGPT user can reach it.
Mercari followed on 23 June. Shoppers can search listings from a ChatGPT conversation, and sellers can draft listing descriptions there. Multilingual search is supported, which points at cross-border demand. In March the company had already shipped generative-AI filtering inside its own app.
This is the same pattern Sephora and Walmart run on ChatGPT: rather than pulling shoppers to your site, you take the catalogue to where shoppers already are. The claim that Japanese companies are standing still does not survive contact with it.
The nature of the move matters, though. Neither company made its own site machine-readable. They entered one specific surface through an app framework. Buying the same way from Gemini or Claude is not possible, and no general-purpose agent has been given access.
Two directions, running in parallel
| Dimension | Marketplace-internal (Rakuten, Yahoo!) | External AI surface (ZOZO, Mercari) |
|---|---|---|
| Where the purchase happens | Inside their own service | Inside ChatGPT |
| External protocol support | Not needed. Own payments and logistics close the loop | Not needed. Rides the ChatGPT app framework |
| Asset being leveraged | Traffic, payments, logistics, membership base | Outfit post archive, long-tail seller inventory |
| How external AI sees them | Structured data only. No transaction possible | Available on ChatGPT alone. Not other AI |
| Effect on sellers | Attribute completeness decides surfacing | Their products reach an external AI surface |
"Japan equals marketplace-internal" is no longer sufficient. The marketplaces are consolidating inward while ZOZO and Mercari move outward. These are not opposing camps so much as different assets seeking different returns.
Rakuten and Yahoo! own traffic, payments and logistics, so internalising pays. ZOZO and Mercari hold assets that gain value on someone else's surface: a decade of outfit posts, and a long tail of seller inventory. For them, going outward pays.
Set against the United States and China, the picture sharpens further. American players are fighting over control of open standards. China has moved toward super-apps completing cross-platform purchases in seconds. Japan resembles neither, because each player has reached a different answer.
What direct-to-consumer operators should do
Selling on a marketplace does not put you on external AI surfaces. Being surfaced by Rakuten's agent and being surfaced by ChatGPT are separate competitions.
Our audit of 42 Japanese e-commerce sites found no company that had independently reached transaction-capable status. Moving up the ladder starts with these checks.
Does the price survive without JavaScript? Disable JavaScript, load a product page, and confirm price and availability are visible. Most AI crawlers do not execute JavaScript
Is Product / Offer structured data present? When name, price, availability and model number are described as data, an AI reads them instead of guessing
Is bot mitigation rejecting AI agents too? Check whether anti-scalping controls block AI search crawlers. Permission in robots.txt is irrelevant if the WAF cuts the request first
Has robots.txt stated a position on AI crawlers? Most Japanese retailers have expressed neither permission nor refusal. Decide your policy first
Do your sitemaps include product URLs? This is how crawlers learn that the product pages exist
Structured data comes after prices are readable. Reversing the order means adding schema to pages nobody can reach.
What marketplace sellers can do
Rakuten's agent recommends using data pooled across services, which means attribute completeness decides whether you get surfaced. A listing with sparse attributes cannot satisfy a narrowly specified request.
Yahoo! Shopping's comparison feature evaluates multiple products automatically and picks one. If you are going to be compared, the material for comparison has to exist: specifications, reviews and accurate stock levels all feed directly into the outcome.
Inventory accuracy deserves particular attention. When an agent recommends something out of stock, the platform absorbs the blame. It is reasonable to expect that stores with repeated stock discrepancies get deprioritised.
Where domestic platforms stand
Support from Japan's own e-commerce platforms is the obvious next question, and here we could not confirm anything from public sources.
We found no official statements on UCP, ACP or MCP support from futureshop, EC-CUBE, MakeShop or Color Me Shop. None of the 42 companies we audited run on those platforms either.
Shopify, by contrast, enabled Agentic Storefronts across its eligible merchant base on 24 March 2026, on an opt-out basis. In our probe of seventeen Japanese-language stores, all six Shopify stores carried a UCP profile and none of the other eleven did.
Which is to say that platform choice currently outweighs merchant investment. Platform-by-platform differences are covered in our comparison of commerce platforms.
Frequently asked questions
Within those marketplaces you inherit the platform's mechanics. Whether external AI such as ChatGPT or Gemini surfaces you is a separate question, and if you also run a direct site it needs its own work.
Apps in ChatGPT is not an open self-serve programme; OpenAI applies review and technical requirements. For smaller operators the realistic first step is making your own site readable.
It depends which layer you mean. On marketplace-internal agent purchasing, Yahoo! Shopping's 15% is high by international standards. On opening a direct site to external agents, our audit found nobody.
LINE Yahoo's agent brand, launched in April 2026, merging the Yahoo! JAPAN AI assistant with LINE AI. Yahoo! Shopping's agent sits inside this framework.
We cannot forecast timing. But because platforms like Shopify enable support in bulk, the number of covered stores can rise independently of merchant decisions. Check what your platform intends to do.
Conclusion
Japan's AI commerce market is past the point of "not yet." Rakuten and Yahoo! Shopping are producing agent-mediated revenue inside their marketplaces, and ZOZO and Mercari have taken up residence inside ChatGPT.
Neither move, however, opened a company's own site to external AI. Our audit of 42 Japanese retailers found nobody who had.
If you run a direct-to-consumer site, load a product page with JavaScript disabled. If the price vanishes, that is where you start. The wider structure is covered in what agentic commerce is.
Stellagent offers a free AI visibility audit for retail and e-commerce operators.
