Getting ChatGPT and Claude to Describe Your Company Correctly

Enquiries that open with "I asked ChatGPT and your name came up" are now a real category.

So are the other kind. A company asks a model about itself and gets a description that is wrong. The wrong address, services it does not offer, an outdated company name.

Both have the same cause. The model is assembling information written about the company from fragments, without checking a source of record. You can fix that from your own site. In order.

First, this is not just SEO with a new name

It is close, and some of the requirements genuinely differ.

A search engine ranks pages. A person picks one and clicks. The goal is getting a person to the page.

A model reads pages and synthesises an answer. The user may never see the page. The goal is whether what is on the page gets quoted accurately.

That difference produces a few requirements, and the biggest is this. Write something ambiguously and it does not come back as a vague summary. It comes back as a confident error.

A person reading a vague sentence thinks that was unclear. A language model takes the most probable reading of a vague sentence and generates a definite one. Ambiguity in your writing gets converted into misinformation about you.

First: put the facts in one place

This is the highest impact item and the least often done.

Put the basic facts about the company on one page, in plain sentences, as a list.

The legal name and how it is written. Year founded. Location. What the business does. The name of each service and what it actually is. Languages you work in. Contact details. Who runs it.

If you already have a company profile page, go and look at it. Most of them are a table with short phrases in the cells. "Business: design." A person can read that. A model cannot expand it correctly, because there is no context to expand from.

Write it as a sentence. "We are based in Tokyo and do product design and AI product development for companies" gets quoted as it stands. "Design" on its own means the model fills in the rest.

Second: structured data

Add JSON-LD describing your Organization, using the schema.org Organization type.

This gives a model an unambiguous source. Name, URL, logo, address, contact points, social profiles. The ideal state is the same facts present both in the prose and in the structured data.

Service and article pages have their own types. Article or BlogPosting for writing, Service for a service, FAQPage for questions.

It is a one time implementation. Put it in the template and it maintains itself. On the sites where I have added it, the accuracy of AI descriptions changed afterwards. I cannot prove strict causation, but as a mechanism for correcting a specific error it works.

Third: check that crawling is actually allowed

This gets missed, and I have seen sites that are never quoted by anything for this exact reason.

robots.txt. AI crawlers arrive under different names from search crawlers. GPTBot, ClaudeBot, PerplexityBot, Google-Extended and others. Check whether you allow them by name.

Your CDN. This is the real trap. Cloudflare has a setting that blocks AI crawlers by default. Allowing them in robots.txt achieves nothing if they are stopped in front of it.

So you can be in a state where you believe you have opted in and not a single line has been read. Check your server logs, or your CDN dashboard, for whether AI crawler traffic exists at all.

Fixing this requires no content changes. It is configuration, and the effect appears on the next crawl rather than in a quarter.

Fourth: one subject per page

This one is about writing.

Pack several subjects onto a page and the model cannot settle on what the page is about. It gets quoted partially in response to everything and authoritatively in response to nothing.

One page, one subject. And state the subject in the first paragraph.

I hold my own site to this. The first paragraph of an article contains the argument of the article. There is no decorative introduction in front of it. Human readers prefer that too, and for a model it is decisive.

Fifth: publish in both Japanese and English

The most common gap among the Japanese companies I talk to.

A Japanese only site does not get quoted in response to an English question. If you want enquiries from outside Japan, you need the English.

And machine translation pasted in is not sufficient, for a specific reason. If the translation reads badly, it gets quoted as written, and your company is introduced to the world in awkward English.

Declare the relationship between the versions with hreflang, so a model can tell which is the original and which is the translation.

What not to bother with

Hidden text for models. Text placed for a model and hidden from people is treated the same way as hidden text for search engines. Do not.

Publishing volume. Thin articles at scale lower the standing of the whole site. What gets quoted is a page with specific facts on it, not the twentieth article on a topic.

Expecting much from llms.txt. It is worth having. There is no guarantee anything reads it. The five items above are far more reliable.

How to check whether it is working

Once a month, ask the major models about your own company.

Ask by name. Ask by what you do. Ask the way a customer would, something like "companies in Tokyo that do X".

Log the errors in what comes back and fix the page responsible. That is the fastest feedback loop available. Keep it up for a few months and the answers change.

It is new work, but it is nearly the same work as existing communications: check how the company is being described, correct what is wrong. Only the medium changed.

If you do one thing

Do the third one, the crawler check.

The other four take time to have an effect. That one is configuration, so it changes on the next crawl. And if that is where you are blocked, doing the other four changes nothing at all.