We put the questions your customers actually type to four AI assistants. This is what came back. Half of it is good news.
This was a read-only review. We changed nothing on your site. Everything below is taken from recorded answers on 31 August 2026 — not estimates.
Ask where to buy pearls in Tokyo, and only one of four AI assistants says your name.
Ask in Japanese, and none of them do.
And one of them is using your own writing to recommend your competitors.
We used the kind of short, imperfect phrase a real person types — not a polished question.
| Assistant | Did it name you? | Who it sent them to |
|---|---|---|
| Perplexity | ✓ Own section, listed first | You, then OHARA PEARL |
| Gemini | △ Quoted your site, recommended others | Ginza, Okachimachi, department stores |
| ChatGPT | ✕ Never mentioned you | MIKIMOTO, TASAKI, OHARA PEARL, others |
| Claude | ✕ Never mentioned you | MIKIMOTO, TASAKI, OHARA PEARL, others |
Why such different answers to the same question? Because the four are not reading the same material.
Perplexity went and read your website, which is why it got you right. Gemini leans on Google's own map and shopping data. ChatGPT and Claude lean mostly on third-party "best pearl shops in Tokyo" articles and on Google's business listings.
The pattern is clean: wherever your own website is the source, you appear. Wherever somebody else's list is the source, you vanish — because you are not on those lists.
Your showroom is three minutes on foot from Roppongi station. But when someone searches in Japanese, not one of the four assistants brings you up.
All four give the same two answers: Ginza (MIKIMOTO, TASAKI, OHARA PEARL, Nichibo Pearl) and Okachimachi (PEARL SEIWA, GENERAL PEARL, Shinju Ryutsu Center).
Gemini spells the split out: Ginza = luxury brands, Okachimachi = wholesale prices.
Your showroom is in Roppongi — neither of those places.
Why zero rather than simply "low"? In Japanese, all four assistants build their shortlist from the same two places: Google's business listings, and Japanese-language articles about where to buy pearls. You are in neither.
There is also a very concrete reason. Every shop that made those lists has 真珠 or パール — the Japanese words for pearl — in its name: 真珠専門店リュウホウ, 東京真珠, オハラパール, PEARL SEIWA. Your Japanese name on Google is アミットトレーディング — Amit Trading, with no mention of pearls at all. In Japanese you read as a general trading company.
In English the opposite is true: your name literally is Pearls. That asymmetry lines up exactly with the results — found in English, invisible in Japanese.
The comfortable explanation is that you don't have enough reviews yet. We pulled the real numbers for every shop on those same lists. That explanation does not survive.
| Shop | Google rating | Reviews | Recommended? |
|---|---|---|---|
| Shinju Senmonten Ryuho (Okachimachi) | 4.8 | 4 | ✓ Yes |
| PEARL SEIWA (Ueno) | 4.9 | 82 | ✓ Yes |
| Tokyo Pearl (Ginza) | 4.9 | 140 | ✓ Yes |
| Pearls.jp (Roppongi) | 5.0 | 144 | ✕ No |
| OHARA PEARL (Ginza) | 5.0 | 256 | ✓ Yes |
| MIKIMOTO (Ginza) | 4.3 | 316 | ✓ Yes |
| WSP (Ginza) | 4.9 | 937 | ✓ Yes |
A shop with four reviews makes the list. You, with 144 reviews and a clean 5.0, do not. Even MIKIMOTO's 4.3 places ahead of you.
Your reputation is not the problem. The problem is that AI believes pearl shopping happens in Ginza and Okachimachi. Roppongi is not on that map, and no amount of work on the website changes that by itself.
So can it be fixed? In large part, yes. The technical approaches available today address a substantial share of this, and we would expect a real improvement from them. Some of it is not a technical problem at all.
Worth setting out plainly, because the last few sections could otherwise be read as "go compete for the Japanese domestic market." That is not what we are recommending.
The data is clear that your customers are English, Chinese and Spanish speakers worldwide, plus foreign visitors who happen to be in Japan. Japanese domestic shoppers are not the prize here, and we would not put your budget there.
The machine-readable information on your Japanese pages still has to be right — because its audience is not Japanese people. It is the search engines and the AI systems themselves.
AI does not handle information the way conventional software does, record by record. It arrives at an understanding and answers from that. So what your Japanese pages say feeds into its overall read of whether this is a pearl specialist based in Tokyo — and that read is what governs how you are treated in English, Chinese and Spanish searches.
This is the most uncomfortable thing we found.
It read the Japanese buying guide you invested in, leaned on it to support its conclusions, and then published a shortlist that leaves you off.
The same Perplexity puts you first when the question is in English. In Japanese you stop being the answer and become the raw material.
MIKIMOTO ¥572,000. TASAKI ¥1,144,000. Pearl for Life ¥181,500. Jewelry Tsutsumi ¥112,000.
Not one of them is yours.
Asked who runs Pearls.jp and how long it has been trading, all four assistants could answer. One of them got caught on your own website.
"The company's own history describes the business as having passed through three generations of pearl expertise, although Dori is the current second-generation principal." ChatGPT · recorded 31 Aug 2026
The AI did not get this wrong. It read your FAQ page, which says three generations, and your Our Story page, which says second generation, found that they disagreed, and printed both.
The About page is stranger still. On a single page it says both "Two generations" and "Three generations, from Kobe to Tokyo."
"All of this comes from the company's own website … so it's their self-reported history." Claude · recorded 31 Aug 2026
Founded in 1969, JPEA member since 1971, generations of expertise — none of it is confirmed anywhere else online. You are the only source. AI discounts what it cannot verify against a second source.
"Google rating: 5.0, 144 reviews. The business's own website currently says 5.0 from 145+ Google reviews, so there's a slight discrepancy between the live business listing and their website." ChatGPT · recorded 31 Aug 2026
The true figure is 144. Three numbers are in circulation right now:
This section isn't AI. It's the record inside your own Google account: 46,300 visits from search over the past 16 months, the longest period Google keeps.
| Where the visitor was | Visits from search |
|---|---|
| Japan | 13,485 |
| United States | 8,530 |
| Spain | 1,558 |
| Singapore | 1,509 |
| Australia | 1,491 |
| Hong Kong, Taiwan & Malaysia | 2,063 |
| Israel | 888 |
Japan comes first, at 29% of everything. Now look at what those people actually typed — the seventeen most common searches, in order:
Not one of the top seventeen is in Japanese. They are in English, Chinese and Spanish.
So the people in Japan finding you are searching in English and Chinese. Almost certainly foreign residents and visitors — not Japanese domestic shoppers.
Not the technical causes. Just what is already happening.
Your showroom is walking distance for anyone already in Roppongi. But in Japanese-language searches, all four assistants leave you out. That audience is not being lost to a competitor — it was never reached in the first place.
The Japanese pages exist and the translation was done. The content is good enough that AI quotes it. The problem is not the website; it is that AI thinks pearl shopping happens in Ginza and Okachimachi.
Ask a full, well-formed question and you come first. Change it to the short phrase a real customer types and three of four assistants drop you — ChatGPT hands the citation to OHARA PEARL's English buying guide instead.
That is not a market you have won. It is a set of phrasings you have won. OHARA PEARL is running the same playbook and closing the gap.
This is better news than we first thought. Now that we've seen the account: you have been connected to Google Shopping since 2021. Your roughly 137 designs are listed once per language, so Google counts about 1,300 listings — all approved and live — and your shipping and return policies are both filled in.
One thing is missing — your store details for Japan. Without them you aren't eligible for Japan-facing shopping results, and those are exactly the results AI draws on when it shows a customer things they can buy.
Separately, your live product listings have fields missing.
Two answers for how many generations. Three numbers for your review count. Two figures for how many countries you supply. These are not internal notes — they are things AI reads back to customers.
The cost is not just one wrong sentence. AI decides whether a business is trustworthy by checking whether the same number holds up in several places. When it doesn't, confidence drops across the board.
This one is not a fault. The currency switcher works, customers can move to yen themselves, and the prices are correct. It is here because changing it would noticeably improve things.
As it stands, a Japanese product page opens showing $533. A customer in Japan has to convert it in their head before they know what the necklace costs.
The effect on AI is more direct: the product data it reads is fixed in US dollars regardless of the visitor's language. Ask an assistant in Japanese what a necklace costs and the answer comes back in dollars.
Google knows about roughly 20,000 addresses on your site and has indexed 4,700. Of the rest:
4,600 are dead links that open to "page not found." Most are left over from an earlier change to how your multilingual addresses were built — the old ones were never pointed at the new ones.
3,700 were looked at and passed over, mostly the same product repeated once per currency and once per language.
The effect: most of Google's time on your site goes to addresses that lead nowhere, so the pages that matter get visited less often and updated more slowly.
English buttons are still sitting on the Japanese, Chinese and Hebrew pages, and Japanese characters have slipped into the Chinese homepage. Not damaging, but for a house of this age it is the kind of roughness a customer notices in the first few seconds.
None of this needs changing, and everything that follows should be built on top of it.
About this review Everything here comes from testing on 31 August 2026: a read-only inspection of www.pearls.jp, and questions put to ChatGPT, Claude, Perplexity and Gemini while signed in. We made no changes to your site. Fifteen full transcripts are on file.
One concern we ruled out We had expected to find your managing director's name recorded incorrectly in Japanese-language business directories — one directory does carry the wrong name. In testing, this turned out not to matter: asked about the company in Japanese, all four assistants cited your own site and gave the name correctly. The bad entry exists, but nothing is picking it up.
我们拿真实问题，去问了四个 AI。这是它们实际给出的答案。结果一半是好消息。
本次只做只读检查，没有对贵司网站做任何改动。全部结论来自 2026 年 8 月 31 日当天的实测记录，不是推测。
问「东京哪里买珍珠」，四个 AI 里只有一个会说出你的名字。
用日语问，一个都没有。
而其中一个，正在用你网站上的内容，推荐你的竞争对手。
我们用一句最普通的话去问——就是真人会打出来的那种短句，不是完整工整的问句。
| AI | 有没有提到你们 | 它推荐了谁 |
|---|---|---|
| Perplexity | ✓ 单列一节，排第一 | 你们，然后才是 OHARA PEARL |
| Gemini | △ 引用了你的网站，但没推荐你 | 银座、御徒町、百货店 |
| ChatGPT | ✕ 完全没提 | MIKIMOTO、TASAKI、OHARA PEARL 等 |
| Claude | ✕ 完全没提 | MIKIMOTO、TASAKI、OHARA PEARL 等 |
同一个问题，为什么差别这么大？因为这四个 AI 看的不是同一份资料。
Perplexity 直接去读了你们的网站，所以答得最准。Gemini 靠的是 Google 自己的地图与购物数据。ChatGPT 和 Claude 主要参考第三方写的「东京珍珠店推荐」文章，以及 Google 的商家名录。
规律很干净：凡是以你们官网为依据的地方，你们就出现；凡是以「别人写的名单」为依据的地方，你们就消失——因为那些名单里没有你们。
展厅在六本木，日本客人走三分钟就到。但他们用日语找珍珠店时，四个 AI 没有一个提到你们。
四个 AI 的答案一致：银座（MIKIMOTO、TASAKI、OHARA PEARL、日貿真珠）和御徒町（PEARL SEIWA、GENERAL PEARL、真珠流通センター）。
Gemini 甚至把市场明确分成两条路——银座＝高端品牌、御徒町＝批发价。
而你们的展厅在六本木——不属于这两条路里的任何一条。
为什么是零，而不是「靠后」？在日语里，四个 AI 都从同样两个地方取名单：Google 的商家名录，和日本人写的珍珠店推荐文章。你们两个都不在。
还有一个很具体的原因。进了名单的每一家，店名里都带着「真珠」或「パール」——真珠専門店リュウホウ、東京真珠、オハラパール、PEARL SEIWA。而你们在 Google 上的日文名是「アミットトレーディング」，一个「真珠」都没有。在日语里，它读起来就是一家普通贸易公司。
英文那边正好相反：你们的名字字面就叫 Pearls。这个不对称，和实测结果完全吻合——英文找得到，日文找不到。
最容易的解释是「评价太少所以排不上」。我们查了同一份名单上每家店的真实评价数——这个解释站不住。
| 店铺 | Google 评分 | 评价数 | AI 推荐了吗 |
|---|---|---|---|
| 真珠専門店リュウホウ（御徒町） | 4.8 | 4 条 | ✓ 推荐了 |
| PEARL SEIWA（上野） | 4.9 | 82 | ✓ 推荐了 |
| Tokyo Pearl（银座） | 4.9 | 140 | ✓ 推荐了 |
| Pearls.jp（六本木） | 5.0 | 144 | ✕ 没有 |
| OHARA PEARL（银座） | 5.0 | 256 | ✓ 推荐了 |
| MIKIMOTO（银座） | 4.3 | 316 | ✓ 推荐了 |
| WSP（银座） | 4.9 | 937 | ✓ 推荐了 |
一家只有 4 条评价的店进得去，你们144 条、5.0 分进不去。连 MIKIMOTO 的 4.3 分都排在你们前面。
问题不在你们的口碑，在于 AI 认为「买珍珠」这件事发生在银座和御徒町。六本木不在这张地图上，而这一点光靠改网站解决不了。
那这能解决吗？很大一部分可以。现有的技术方案能处理其中相当一部分，带来实实在在的改善。也有一部分根本不属于技术问题。
这一点值得单独讲清楚，否则前面几节容易被读成「你们该去抢日本本土市场」——我们并不这样建议。
数据已经说明，你们的客人是全球的英语、中文、西班牙语买家，以及人在日本的外国访客。日本本地消费者不是这里的目标，我们也不建议把预算投在那儿。
但日语页面上那些「机器读的资料」仍然必须写对——因为它的读者不是日本人，是搜索引擎和 AI 本身。
AI 处理信息的方式，不是传统软件那种逐条读取、逐条比对，而是先形成理解、再据此作答。所以日语页面上写了什么，会进入它对「这是一家位于东京的真珠专门店」的整体判断——而这个判断，决定了你们在英文、中文、西班牙语搜索里被怎么对待。
这一条是本次检查里最刺眼的发现。
它读了你们花力气写的日文购珠指南，用其中的内容支撑自己的结论，然后列出了一份不包含你们的推荐名单。
同一个 Perplexity，在英文提问时把你们排在第一位。换成日语，你们从「答案」变成了「素材」。
MIKIMOTO ¥572,000、TASAKI ¥1,144,000、Pearl for Life ¥181,500、Jewelry Tsutsumi ¥112,000……
没有一件是你们的商品。
问四个 AI「Pearls.jp 是谁在经营、开了多久」，四个都答得上来。但其中一个，被你们的网站难住了。
「公司自己的历史页面说这门生意传承了三代，尽管 Dori 是现任的第二代负责人。」 ChatGPT · 2026-08-31 实测原文翻译
这不是 AI 弄错了。是它读了你们的 FAQ 页（写三代）和 Our Story 页（写二代），发现对不上，只好把两种说法都写出来。
About 页更特别：同一页里既写了「Two generations」，又写了「Three generations, from Kobe to Tokyo」。
「以上全部来自该公司自己的网站，所以这是他们的自述历史。」 Claude · 2026-08-31 实测原文翻译
1969 年创业、JPEA 会员资格、三代传承——这些关键事实在网上找不到第三方佐证，只有你们自己在说。AI 会因此降低对这些说法的信任权重。
「Google 评分 5.0，144 条评价。该公司自己的网站写的是 5.0 分、145+ 条，所以线上商家资料和他们的网站之间存在出入。」 ChatGPT · 2026-08-31 实测原文翻译
真实数字是 144。而目前对外的说法有三个：
这一节不是问 AI 得来的，是你们自己 Google 后台的记录：过去 16 个月，来自搜索的 46,300 次访问。16 个月是 Google 保留数据的上限。
| 访客所在地 | 来自搜索的访问 |
|---|---|
| 日本 | 13,485 |
| 美国 | 8,530 |
| 西班牙 | 1,558 |
| 新加坡 | 1,509 |
| 澳大利亚 | 1,491 |
| 香港・台湾・马来西亚 | 2,063 |
| 以色列 | 888 |
日本排第一，占全部访问的 29%。但再看这些人实际输入的是什么——搜索量前十七位的关键词：
前十七名里，没有一个是日语。全是英语、中文和西班牙语。
也就是说，在日本找到你们的这些人，是用英文和中文在搜。几乎可以肯定是住在日本的外国人和来旅行的客人，不是日本本地消费者。
不讲技术原因，只讲已经在发生的结果。
展厅在六本木，日本本地客人步行可达。但四个 AI 在日语查询里全部没有提到你们——这部分客源目前等于零，而且不是被抢走的，是从来没进过战场。
日文页面是有的，翻译也做了，内容甚至好到被 AI 当资料引用。问题不在网站，在于 AI 认为「买珍珠」发生在银座和御徒町。
用完整工整的问句问，你们排第一。但换成客人真实会打的短句，四个 AI 里三个不提你们，ChatGPT 还把引用给了 OHARA PEARL 的英文攻略。
这不是「已经赢了」，是在一部分问法上赢了。而 OHARA PEARL 正在用同样的打法追上来。
这一条比我们最初判断的要好。看过后台之后确认：你们 2021 年就接好了 Google 购物。你们约 137 款设计，每种语言各上架一条，Google 那边合计约 1,300 条，目前全部审核通过、正常在售，配送与退货政策也都填好了。
只差一样——日本的门店资料没填。没有它，你们就没资格进入面向日本的购物结果，而 AI 展示「可以买的东西」时用的正是这类结果。
另外，目前在售商品的资料存在字段缺失。
「第几代」有两种说法、评价数有三个版本、供货国家数有两个数字。这些不是内部资料，是 AI 当着客人的面复述出来的东西。
后果不只是这一条被讲错——AI 判断一家店可不可信，靠的是「同一个数字能不能在多处对上」。对不上，它就整体降低信任。
这一条不是错误。货币切换器工作正常，客人可以自己切换到日元，标价本身也没有问题。列在这里，是因为改动之后体验会明显更好。
目前打开日文商品页，价格默认显示 $533。日本客人要自己心算一次，才知道这条项链大概多少钱。
AI 那边影响更直接一些：它读取的商品数据里币种固定为美元，不随访客语言变化。所以有人用日语问 AI「这条项链多少钱」，拿到的仍是美元报价。
Google 知道你们网站上大约 2 万个网址，但只收录了 4,700 个。剩下的里面：
4,600 个是死链，打开显示「页面不存在」。它们多半是网站早年改过一次多语言网址结构留下的旧地址，改完之后没有指向新地址。
3,700 个 Google 看过后决定不收录，大多是同一件商品因为切换货币、切换语言而重复出现。
后果是：Google 每次来，时间大都花在了通向死胡同的网址上，真正重要的页面反而被看得更少、更新更慢。
日文、中文、希伯来文页面上还留着没翻译的英文按钮，中文首页混进了日文假名。不致命，但对一家五十多年的老店来说，是客人第一眼就会注意到的粗糙感。
这些不用改，而且后面所有工作都要建立在它们之上。
关于这份报告 所有结论来自 2026 年 8 月 31 日的实测：对 www.pearls.jp 的只读检查，以及对 ChatGPT、Claude、Perplexity、Gemini 四个 AI 的登录状态提问。我们没有对贵司网站做任何修改。共 15 份完整问答记录存档备查。
已排除的一个说法 我们曾担心日语世界里贵司负责人姓名被第三方企业名录写错（该名录确实写错了）。登录状态实测下未发现这个问题——用日语问公司名，四个 AI 引用的都是贵司官网，负责人姓名正确。该风险存在但目前没有生效。