When I first thought about designing a popular Japanese AI product site, a lot of very Japanese images came to mind: the wide-eyed mascot girl from a manga panel, handwritten POP signs outside convenience stores, those clean blue skies from anime, and an AI assistant that gives a small bow and says, “いらっしゃいませ” (Welcome). Reality is usually a little different from the picture in your head…

All of those things are charming. But one question came to me pretty quickly: would a Japanese company really hand over its team chats and business documents to a new AI product just because it has a cute character?

This is the story of how I moved past those first impressions and found an answer to that question.

First, some context. I designed both the Japanese site for Tanka and Apodex, a product aimed at the U.S. market. Both run on the same design system, which I also built. So the data and observations in this article come from products I was directly responsible for.

1. A Gut Feeling, and a Quick Experiment

The B2B AI market in Japan has a fairly concentrated group of competitors. After looking closely at the leading products, those images in my head seemed to have some basis: maybe a character-driven avatar assistant really could appeal to users in this market.

I also wanted to see how far the design system could stretch in that direction, so I quickly pushed forward with a cartoon, manga-inspired version. Five days of user observation told a different story. In fact, it didn't work at all. Users were clearly interested in the avatar and interacted with the character, but those interactions didn't lead to any meaningful conversion. People liked it, but it didn't make them trust the product more.

Looking back, I think there were two reasons. First, Japan's IP market is already a crowded red ocean. Consumers are used to an extremely high standard of character design, so expecting a new character to stand out right away was a little optimistic. Second, building a character people genuinely care about takes long-term investment and refinement, and that was exactly what this product couldn't afford. We needed something fast and lean.

There was, however, one result I had expected: the design system held up remarkably well. Even with a cartoon character, it kept the expression consistent with the brand language. Some of my colleagues in Japan went out of their way to praise it, saying they had never seen technology and character design combined in this way.

The experiment confirmed two things for me. First, the expressive range of the system was wider than I had expected. Second, what competitors like is not necessarily what users need. At the time, I was trying to answer “Do people like this?”, while the question users were actually asking may have been “Can I trust this?”

2. What Users Actually Did, and How I Interpreted It

2.1 What I Observed

So I stopped looking at what users seemed to like and started looking at what they actively went looking for.

I studied the journeys of the first 100 users who came to the Japanese site through organic growth, and used Clarity's page-visit rankings as a second source of evidence. The rankings cover the most recent 14 days. I also checked an earlier 30-day period, and the rankings were largely consistent, which suggests this isn't just a short-term fluctuation.

Tanka's English site doesn't have Clarity installed, so I used Apodex as a reference point. It runs on the same design system as the Japanese Tanka site, but the product and audience are different. This is not a controlled comparison; it can only show the direction of the difference.

One of the cleanest signals came from the terms of service. On both sites, the terms page sits in the footer, in almost exactly the same position. Under those conditions, the Japanese site's サービス利用規約 (Terms of Service) ranked ninth among all pages visited, while Apodex's more comprehensive legal terms didn't make the top ten. A footer link requires a user to deliberately scroll to the bottom and click it, which is hard to explain through layout alone.

Other rankings on the Japanese site pointed in the same direction. 会社概要 (Company Overview) ranked second, right after the homepage and ahead of the pricing page. セキュリティとプライバシー (Security and Privacy) ranked fourth, ahead of product features and use cases. Both of those pages had more prominent entry points on the Japanese site, though, so some of their traffic could come from placement. I treat them as supporting observations rather than primary evidence.

Apodex's traffic, by contrast, was concentrated on narrative content. Four of its top ten pages were blog posts, and one of them was simply titled “What Apodex Believes.” The site has complete company information, security documentation, and legal terms, but none of them made the top ten.

The same system, pages in the same positions, and yet users went in very different directions. On our Japanese site, users actively checked out the company. On Apodex, users were more interested in reading what the company believes.

2.2 How I Interpreted It

What follows is my interpretation of this behavior. It's a hypothesis based on the data, not a validated conclusion.

My sense is that users on both sides are asking the same question: why should I trust you? The answers are just different. Rather than attributing that to nationality, I tend to attribute it to how the decision gets made: is the product approved by an organization, or does an individual make the call?

When a purchase needs organizational approval, such as a ringi (稟議), where a proposal is circulated for sign-off level by level, the decision is shared across the organization, and long-term stability tends to matter more. In that situation, users need evidence they can verify and justify to the people above them: certifications, a physical address, legally required disclosures, and numbers presented honestly. The behavior on our Japanese site matches this pattern.

When the decision rests more with an individual, people are often buying into a direction and a vision. A good story convinces the decision-maker personally and also gives them something to convince others with, so the story itself becomes the evidence. The behavior on Apodex is closer to this pattern.

This understanding shaped the design decisions that followed. We weren't putting different skins on two markets. We were answering one question: what counts as evidence in each market?

3. How the Design System Carries That Decision

The principle fits in one line: one system, many expressions.

The foundation stays the same: design tokens, the component library, and the narrative structure of the page. What changes is the “evidence slot” built into that structure, and each market decides what goes there. The Japanese Tanka site fills it with information that can be checked and verified. Apodex gives the space to narrative: blog posts, a statement of beliefs, and the Frontier Program.

A common misreading is that the Japanese version puts credibility before the story, reordering the argument. That's not what the page does. The Japanese homepage still opens with a story: “あの話、どこだっけ？ TANKAが、覚えてる。” (Where was that conversation again? TANKA remembers.) The credibility signals sit right alongside it on the first screen. The order didn't change. What counts as evidence did.

Three details on the Japanese site show this most clearly.

The first is sample size. The satisfaction figure is shown with n=22. Most companies would hide a sample that small. My judgment was that for users who have to justify a decision to someone above them, being honest about the numbers matters more than how big the numbers are, so showing the sample size is itself a signal of credibility. The ISO/IEC 27001 and SOC 2 badges, the physical address in Roppongi, and the disclosure required under Japan's Act on Specified Commercial Transactions (特定商取引法に基づく表示) all serve the same purpose.

The second is the CTA. The free-start button stays, because we don't want to remove the easiest way in. But it can't be the only reason to click, so it sits next to the ISO badge instead of standing alone.

The third is the line “必要な準備は、先に。最後の判断は、あなたに。” (AI handles the preparation. The final decision is yours.) For someone who is responsible for a decision, whether AI might overstep is a real concern. This line hands the final judgment explicitly back to the person.

The problem statements follow the same thinking. Instead of feature descriptions, I wrote the thoughts users already have in their heads: “あの話、どこだっけ？” (Where was that again?), “これ、前にも聞いたっけ……” (Didn't I ask this before…), “また、最初から説明？” (Explaining it all from scratch again?), and “あれ、今日までだった……” (Wait, that was due today…). Rather than saying what the product can do, they describe how annoying things are right now. The onboarding flow repeatedly stresses that nothing has to change: 今のツールを変えずに始められる (Start without changing your current tools), まずは、ひとつつなぐだけ (Just connect one to begin), 所要時間は1分 (Takes one minute).

This is guaranteed by tooling, not by designers making the call manually each time. Take typography: when the market or the typeface changes, the structure and the rules stay the same, and only the parameters are regenerated.

4. Results and Limitations

Fuller data is still coming in, but there is one result I can already share. About half of the Japanese site's current organic traffic comes from search engines, which shows the SEO work is paying off. Another 6% comes from AI recommendations, which I credit to the GEO (Generative Engine Optimization) built in at the design-system level, making the structure and content of each page easier for AI systems to understand and cite.

It also shows that a design system serves not only users, but search engines and AI as well. When those capabilities live in the system, every product running on it gets them for free.

At the same time, these observations have clear limits. The Japanese-site data comes from the first 100 users who arrived organically. This is one of my design principles: first study the earliest potential users who find the product on their own, build a user profile from their behavior, and then use it to target the audience for later paid acquisition. The sample is limited, but it hasn't been shaped by any acquisition strategy, so it's closer to how users naturally behave. The U.S. side comes from a different product, with a different audience and product shape, so it can only serve as a directional reference, not a controlled experiment. And the cultural interpretation is my own reading of the data, not a validated conclusion.

These observations were enough to support the design decisions I made at the time, but not enough to generalize about “Japanese users” or “American users” as a whole. I'd rather see this article as a record of a process: I saw a pattern in my own products, made a judgment, and then tested that judgment against data.

5. Design and Engineering

5.1 Typography

The typography on the Japanese site wasn't tuned by hand. It was generated by a tool I built, according to a single set of rules. From the largest headline down to the smallest caption, every level gets its font size, line height, and letter spacing from the same rules, and the values scale across screen sizes so the hierarchy always stays consistent.

Letter spacing is a good example of how these rules work: the larger the type, the tighter the spacing; the smaller the type, the looser. Japanese characters are visually full, so at large sizes the gaps between characters start to feel loose and need tightening. At small sizes, the dense strokes need room to breathe.

The Japanese site uses Noto Sans JP for everything from headlines to body copy. I checked nine Japanese B2B AI product websites, and seven of them use it; it's close to a default in this space. I chose it for two reasons: users see letterforms that are already familiar from other products in the category, and pages render consistently across operating systems. Sites that don't specify a Japanese font can show the same page in two different typefaces on Mac and Windows.

The system also has a limitation I'm working on. It was originally built around Latin type, and when applied to Japanese, the line height comes out a little tight, not quite right for long-form reading. My next step is to build a separate set of line-height rules for CJK scripts.

5.2 Competitor Research: Two Judgments, Two Outcomes

Looking back, two of my judgments on this project came from competitor research.

The first was the avatar. After studying the leading competitors, I judged that a character-driven assistant would appeal to users. The user data proved me wrong: people liked it, but it didn't convert.

The second was typography. Noto Sans JP showed up a lot among competitors, so I checked nine sites to verify it. This time the data confirmed the judgment, and Tanka adopted the same choice.

The difference is in what each one was answering. The avatar answered “Do users like this?” The typeface answered “Does this feel familiar, and does the page stay stable?” The first is a preference, and competitors' choices rarely transfer directly. The second is closer to infrastructure, where industry consensus often is the user's expectation. Competitors are worth studying, but afterwards you have to separate the conventions you can borrow from the assumptions you still need to test yourself.

Conclusion

Back to the original question: can one design system serve two markets where users look for completely different kinds of evidence?

From the cartoon character, to the credibility signals on the Japanese site, to the narrative on Apodex, all of them run on the same system. A good design system doesn't make every product look the same. It gives every product room to express itself differently.

Appendix: Typography Across Japanese B2B AI Product Sites

In September 2026, I used Chrome DevTools to read the computed fonts of the h1, h2, p, and button elements on the homepages of nine Japanese B2B AI products.

Seven of the nine sites ultimately render Japanese in Noto Sans JP. The most common approach is to use it directly (five sites), followed by placing a Latin font first with Noto Sans JP behind it (two sites). About half of the sites also add Hiragino and Meiryo after the primary font, so an appropriate Japanese typeface still appears if the web font fails to load.