Why AI cannot tell you what Japan incorporation costs

Ask ChatGPT, Perplexity or Gemini what it costs to incorporate in Japan with professional help, and you will get a confident-sounding range. Ask it which firm charges what, and the answer thins out fast — a couple of named prices, then hedging.

That is not the model being cautious. It is the model having nothing to quote.

We list 54 firms that help foreign founders set up in Japan. Only 25 of them publish any prices at all on their own websites. The rest ask you to make contact first. This page sets out what we measured, because the consequence is now larger than it used to be: a firm that publishes nothing is not merely harder to compare. It is close to invisible to the tools people increasingly use to shortlist.

What the 54 firms publish

29 of 54. That is the industry norm, not an aberration, and it is worth saying plainly that publishing nothing is legal, common, and defensible — engagements genuinely differ, and a “from” price that turns out to be wrong is its own kind of harm. Several of the firms that publish nothing are among the most experienced on this list.

But the founder’s side of that trade has changed.

The crawlers are already there

We keep our own access logs and verify every crawler against the IP ranges its operator publishes, because a user agent string is not evidence — in the same sixteen days below, a single address worked through ChatGPT-User, GPTBot, PerplexityBot, OAI-SearchBot and bingbot in turn while probing for .env files. None of those were real.

Counting only the fetches that came from an operator’s own published IP ranges, between 20 August and 4 September 2026:

Those crawlers reached 36 of our 37 firm listings, 85 times in total. Not the front page — the individual listing pages, by name.

To be exact about whose pages those were: these are fetches of our listing for each firm, on this site. They do not tell us what any AI system said afterwards, and they are not a measurement of the firms’ own websites. What they do establish is that when these tools go looking for a firm on this list, something gets read — and what gets read is whatever has been written down.

Almost nobody is blocking them

The obvious explanation for thin AI answers would be that firms are shutting the crawlers out. We checked. On 5 September 2026 we fetched the robots.txt of all 54 firms’ own sites and parsed the rules that apply to ten AI crawlers.

• One firm blocks any of them. CROSSCOOP disallows GPTBot, ClaudeBot, Google-Extended, CCBot, Applebot-Extended, Bytespider and Amazonbot, while still allowing the two crawlers that feed ChatGPT search. That is a coherent position, deliberately taken.
• One other firm mentions AI crawlers at all. Forvis Mazars Japan names five of them and permits them.
• The remaining 35 have no AI policy in their robots.txt. Most have the default WordPress rules and nothing more.
• Seven have no usable robots.txt to read. Three return 404; three return a HTML page instead of a text file, which a crawler cannot parse; one did not respond to us at all.

So the door is open at essentially every firm on this list. The reason an AI answer cannot tell you what a given firm charges is not that it was refused entry. It is that it went in and found no number.

What this means if you are the founder

Treat an AI price range as a floor built out of the minority of firms that publish. It is skewed towards them by construction, and the firms that publish are not a random sample — several are online-first operations whose model depends on a visible price, which is a different kind of business from a long-established practice that quotes after a consultation.

Two practical consequences:

• A quoted range is not the market. If a tool tells you incorporation runs JPY 200,000 to 350,000 in professional fees, that is roughly the published band on this list — but nearly two thirds of the firms here are not in it.
• Absence of a price is not a red flag. It is the majority position. What it does mean is that you cannot shortlist on price without making contact, so budget the time for three or four conversations rather than expecting to compare on a screen.

Our published fees page lists every figure we have been able to source from a firm’s own site, with the date we checked it. Where a firm publishes nothing, we say so rather than estimating.

What this means if you run one of these firms

The decision not to publish prices was made when the alternative was a competitor reading your price list. That is still true. What is new is that the same silence now removes you from a set of answers you are not present at.

You do not have to publish a full price list to change this. A single figure with its scope attached — “company incorporation, GK, from JPY X, excluding the JPY 60,000 registration licence tax” — is quotable. A page that says “fees on enquiry” is not, and no amount of crawler access fixes that.

The same goes for anything else a founder asks before choosing: whether you take clients who are still overseas, which languages your staff actually work in, whether the first consultation is free. On this list, 22 of 54 firms state a position on clients who are still overseas. Six do not say which languages they work in, and four of those make no mention of English support anywhere on their own site.

If a field on your listing here says “Not published” and you would rather it did not, publish the information on your own site and tell us. We will re-check and update the listing at no charge, whether or not you ever advertise with us. That is set out in how we list.

Method

• Fee disclosure is recorded per firm from the firm’s own website, with the check date shown on each listing. Re-checked on a rolling basis: every listing here was verified between 2 August and 2 September 2026.
• Crawler figures come from this site’s raw server access logs for 20 August to 4 September 2026 (7,612 requests). Every hit is matched against the operator’s published IP ranges — OpenAI’s gptbot.json, searchbot.json and chatgpt-user.json, and Perplexity’s perplexitybot.json — and unverified hits are excluded. Crawlers whose operators publish no IP ranges, including ClaudeBot, Applebot and CCBot, are not counted in either direction.
• robots.txt was fetched from each firm’s own domain on 5 September 2026 and parsed by user-agent group. We report only what the file says. Whether a firm’s server or CDN separately refuses a crawler is not visible in robots.txt and we make no claim about it.
• Counts marked as live figures update automatically as listings change; the fixed numbers in this piece are as at 5 September 2026.