Insights · Japanese keyword research
Japanese keyword research: four assumptions that quietly break it
If you run SEO in English, keyword research has a shape you trust. Pull a list, check volumes, sort, pick. That shape mostly survives a move into French or German. It does not survive a move into Japanese.
Published 2 September 2026. Figures measured by SuzuLabo in August 2026. Search volumes are Google Keyword Planner estimates, not observed traffic. Domain and page authority figures, where mentioned, are third-party Ahrefs estimates.
What breaks, and in what order
The reason is not that Japanese is mysterious. It is that four things you can safely assume in English stop being true, and none of them announce themselves. Your list still looks like a list. The numbers still look like numbers. The work fails quietly, six months later, when the pages you built do not appear.
This is what breaks, in the order you are likely to hit it.
Scope: this is about finding the Japanese terms your buyers type, for a Japanese-language site you are building or already have. It is not about reading Japanese pages in English, and it is not about translating a finished site. If you are deciding between translation and localization first, that decision comes before this one.
Assumption 1. Translating your keyword list gives you Japanese keywords
Translation gives you a Japanese phrase. Keyword research needs the phrase people actually type. Those are different jobs, and translation is not built to do the second one.
Take a term as ordinary as "quote" or "estimate", the thing a prospect clicks before they contact you. A translator will hand you one Japanese word for it and move on, correctly, because that is the job. What a translator will not tell you is that the word arrives in several written forms, that Japanese companies use different ones, and that the choice is not stylistic.
You now have one form out of several, no way to tell whether it is the one with demand, and a list that looks finished.
The same applies to loanwords. Many business terms exist in Japanese twice: once as a word borrowed from English and written in katakana, once as a compound built from Chinese characters. Both are correct. They are not interchangeable in search, and which one your buyer uses tends to track their industry and their seniority rather than anything about the language.
What this means in practice: treat a translated list as a list of candidates, not of keywords. Every entry needs to be resolved against real Japanese search results before it earns a page.
Assumption 2. One spelling per term is enough
In English, a term has a spelling. Occasionally two, and the difference is regional and predictable.
In Japanese, a single business term routinely appears in several written forms, all of them normal, all of them in use on real corporate sites. Here are three kinds of split, with the actual forms, because the problem is much easier to see than to describe:
- How much of the ending gets written out. "Estimate" or "quote": 見積もり / 見積り / 見積. Same word, same meaning, three spellings.
- Whether a polite prefix is attached. "Inquiry": お問い合わせ / 問い合わせ / お問合せ / 問合せ. The prefix is optional and the ending contracts independently, so one term yields four forms. This is the word on your contact button.
- Whether a long vowel is marked. "Computer": コンピューター / コンピュータ. "Server": サーバー / サーバ. Both conventions are established in Japanese technical writing, and large publishers differ on which they use.
If you do not read Japanese, nothing in that list is visible to you. A translator hands you one of them. A spellchecker will not object to any of them.
None of these are errors. You cannot resolve them by picking the most correct one, because they are all correct.
The tools will not show you the split
We tested this. In August 2026 we submitted 11 written forms of four business terms to Google Keyword Planner, with the location set to Japan.
The tool returned 5 rows.
It had merged our 11 inputs into 5 keywords. Which forms it merged, and on what basis, is not shown anywhere in the interface. When we repeated the run on 2 September 2026 with the figures visible, the five surviving rows were 問い合わせ (22,200 monthly searches), お問い合わせ (9,900), 見積もり (18,100), コンピューター (33,100) and サーバー (27,100). So the tool kept the polite-prefix pair apart, but folded the three spellings of "estimate" into one row and the long-vowel variants of "computer" and "server" into one row each. The spelling it shows as the label is the tool's choice, not evidence of what people type. The variation that decides whether your page matches what someone typed is variation the tool quietly hides from you.
What this means in practice: decide the primary form deliberately, then make sure the alternates appear naturally somewhere real on the page. Do not spray all of them into a title.
Assumption 3. The number the tool gives you is the number
This is the one that costs the most, because the output looks authoritative and is easy to paste into a deck.
The location setting changes the answer, and not by a predictable amount
We measured the same keyword set twice in August 2026, changing only the location. Same account, same period, same language setting.
| Keyword | Location: Japan | Location: United States |
|---|---|---|
| japanese translation services | 40 | 260 |
| japanese website design | 20 | 110 |
| japan seo services | 10 | 50 |
| japan seo agency | 40 | 50 |
| japanese seo company | 70 | 70 |
| japan market entry consulting | 20 | 10 |
Google Keyword Planner, monthly average, July 2025 to June 2026. Measured by SuzuLabo, August 2026. These are English-language keywords; we use them here because they isolate the location variable cleanly.
Read the last three rows before the first three. One term moved by more than six times. One did not move at all. One moved down.
There is no rule here you can learn and then skip the measurement. "Japanese numbers are smaller" is wrong as often as it is right. The only reliable move is to set the location to the market you are selling into, every time, and to distrust any number you inherit without knowing which location produced it.
"No data" is not "no demand"
In the same run, 4 of our 14 keywords returned no figure at all. Not zero. A dash.
Those four were not fringe terms. They were phrases describing services companies actively buy. The tool simply declined to estimate them, which is what it does below a volume floor it does not publish.
If you treat a dash as a zero, you will systematically drop exactly the specific, low-volume, high-intent phrasing that a B2B buyer uses, and keep the broad terms that a student uses.
The account you query from changes what you see
One more thing that is not documented anywhere obvious. We ran the same lookup from two different Google Ads accounts on the same day. One returned exact figures. The other returned nothing usable at all.
We are not going to claim we know Google's exact rule here, because we have not been able to verify it against official documentation. What we can say is that the granularity you get depends on the account you are querying from, so two people can run "the same" research and get different answers, and neither will know why.
Assumption 4. Pick the biggest number
Volume tells you how many people search. It does not tell you whether any of them buy from you.
We see this on our own site. Over 28 days in mid-2026, our English pages drew 59 distinct queries. Not one of them contained a word describing a service someone purchases. The traffic was real. It was people finding out how search works in Japan, which is a useful thing to be found for and a poor thing to forecast revenue from.
The Japanese equivalent of this trap is worse, because the highest-volume Japanese terms in most categories are consumer terms, and a B2B buyer's phrasing sits far down the tail where the tool has already stopped reporting numbers, as above.
What this means in practice: rank candidates by how close the phrasing sits to a purchase, then use volume to break ties. Not the other way round.
What to do instead
The order matters more than the tools.
- Fix the market and the language setting first. Location Japan, Japanese. Write down which setting produced every number you keep. Most inherited keyword sheets fail this step alone.
- Treat translation output as candidates. One English term becomes several Japanese candidates. Expect the list to grow before it shrinks.
- Open the actual Japanese results for each candidate. (The settings for doing that from abroad are in How to see what Japanese buyers see in search.) This is the step that cannot be skipped and cannot be done without reading Japanese. You are looking at what kind of page ranks, whether the intent matches what you sell, and whether the companies there are competitors or media. If you are new to the Japanese search landscape, what the search engine data actually shows is the short background.
- Collapse the written variants deliberately. Choose the primary form, note the alternates, and check that your choice matches what the ranking pages use.
- Rank by distance to purchase, then volume. Keep the dashes. A term with no reported volume and obvious buying intent is worth more than a term with 1,000 and none.
- Write down what you could not resolve. Every real keyword set has terms you are unsure about. Recording them is what makes the next round cheaper.
What you can do yourself, and where it stops
Being honest about this is more useful than pretending the whole thing is easy or that all of it is specialist work.
You can do this yourself: steps 1, 2 and 6. Setting locations correctly, generating candidates, and keeping an honest record of open questions need care, not Japanese.
You will struggle without Japanese: steps 3, 4 and 5. Reading a Japanese results page tells you things no tool exports. Whether the top result is a manufacturer or a comparison site. Whether the phrasing is what a buyer would use or what a job seeker would. Whether a term is drifting out of use. Machine translation of a results page loses exactly the signal you came for, because it normalizes the variation described above into one English phrase.
That is the split. If you want the second half done rather than approximated, that is what our Japan SEO work covers, and the written Snapshot is the smallest version of it.
FAQ
Frequently asked questions
Can I just use an AI assistant to generate Japanese keywords?
It will produce fluent candidates, which is useful at step 2. It will not tell you which form carries demand, because that requires current search data it does not have. Use it to widen the candidate list, not to close it.
Do I need a separate page for each written variant?
No, and doing it creates pages that compete with each other. Choose one primary form per page and let the alternates appear naturally in the body.
Is Google Keyword Planner enough on its own?
It is the right starting point, because it is free and unlimited, and in our experience it reports actual figures rather than ranges when queried from an account with active spend. Google does not document that rule, so treat it as an observation. It is not sufficient, for the three reasons in Assumption 3.
Does any of this change because of AI search?
The research does not change much. What changes is what you do with the result: a page that answers a specific question with checkable specifics has a better chance of being cited than one built to rank for a head term. Search volume was always a proxy. It is a weaker proxy now.
Our Japanese distributor already gave us a keyword list. Is that enough?
It is a strong starting point and better than a translated list, because it comes from someone who reads the market. Ask which location setting produced the numbers, and when. Lists assembled for a different purpose often carry the wrong location or are several years old.
How long does this take?
For a focused set covering one service line, the research itself is days rather than weeks. The part that takes longer is agreeing which terms you are willing to build pages for.
If you already have a list
Most Japanese keyword sets we are asked to review are not wrong because someone was careless. They are wrong because the method that works in English was applied faithfully to a language where four of its assumptions do not hold.
If you have a list already and want to know which parts of it survive, send it over with the URL of your site. We will tell you what we would keep, what we would re-measure, and why, and if the answer is that your list is fine, we will say that.