Most founders have done the vanity version of this. You open ChatGPT, type your company’s name, and read a pleasant paragraph about yourself. It feels like good news. It is not news at all.
An assistant asked about you by name will talk about you. The question that matters is the one your buyer asks before they have heard of you: who should we consider for this? Whether your name comes back from that question is the whole test.
Here is how to run it properly, by hand, in about an hour. It works the same way in Claude and Gemini, and you should run it in all of them.
Write the questions your buyer asks, not the ones you would
The test is only as honest as the questions. Write them in the buyer’s words, before they know you exist.
Ten questions is enough. Spread them across three kinds.
- Category questions: “What are the best options for this problem for a company our size?”
- Problem questions: “How do teams like ours usually solve this, and who helps with it?”
- Comparison questions: “What should we look for when choosing a provider for this, and who fits?”
Leave your name out of all of them. Leave your competitors’ names out too. The moment a question contains a name, the answer is about that name, and you have stopped testing whether you get found.
If you do not know the words your buyers use, look at the first message in your last twenty inbound conversations, the questions people ask on sales calls, and the phrases in your won deals’ notes. Buyer language is usually plainer and more specific than the language on your homepage.
The moment a question contains a name, the answer is about that name.
Ask each one twice: from memory, and with the live web
The same assistant can give you two different answers, depending on whether it searched.
Start in a clean session. Log out, or open a new chat with memory and personalisation switched off, so the assistant is not tailoring its answer to what it knows about you.
Ask each question once with web search off. That answer comes from the model’s memory: what it learned about your market in training. Then ask it again with web search on. That answer comes from the pages the assistant reads today. Where an assistant will not let you switch search off, note on the sheet that the answer searched.
Ten questions, twice each, in ChatGPT, Claude and Gemini, is sixty answers. Each takes about a minute to ask and record once you have a rhythm. If an hour is too long, do five questions. Five questions honestly recorded beat fifty half-remembered.
What to write down for every answer
Four columns. Resist the urge to summarise until every row is filled in.
| Column | What to write | Why it matters |
|---|---|---|
| Named | Yes or no, and where in the answer | The headline count |
| Facts | Right, or which fact was wrong | Wrong facts are worse than absence |
| Instead | Every company named in the answer | Your real competitors in the answer |
| Sources | Which pages were cited, if any | Tells you what the live web reads |
Add two more columns if you have the patience: the date, and the exact wording of the question. Next month you will want to ask the same questions the same way, and you will not remember the wording.
Four results, four different fixes
The count is not the point. The pattern in the count is.
Named from memory, missing with the live web. Your reputation is ahead of your pages. The assistant searched, read what ranks for the question, and you were not in it. Look at the sources column: those are the pages you need to be on or to beat. This is the fastest fix of the four.
Named with the live web, missing from memory. Your pages work and your footprint is thin. Not enough of the web says clearly who you are and what you do. This is slow work: consistent descriptions of you in directories, reviews, press and other people’s pages, over months.
Named, with wrong facts. Old pricing, a retired product, a category you left. Find where the wrong fact lives on the web, correct it at the source, and say the right fact clearly and often on your own site. Then check again.
Missing everywhere, and the same three names keep appearing. Those three are who the assistants think owns your category. Read what the cited pages say about them. That is the bar.
Keep the count honest
Sixty answers is a small sample. Treat it like one.
If you were named in two answers out of sixty, write 2 of 60, not 3%. The fraction tells you, and anyone you show it to, how little data sits behind it. Ask the questions that surprised you a second time before you act on them, because assistants do not give the same answer twice in a row.
Then put the sheet away for a month and run it again, same questions, same modes. The difference between the two sheets is the only trend worth reporting, and it is one you can defend line by line.
Typing your own name into ChatGPT tells you it can read. Asking your buyer’s question tells you whether you are on the shortlist.
Questions people ask next
Why does ChatGPT describe my company but never recommend it?
Because asking about you by name is a different question. An assistant will describe any company you name; whether it names you unprompted, for a buyer’s category question, is what decides the shortlist.
Should I test with web search on or off?
Both. With search off you are reading the model’s memory; with search on you are reading what today’s web says. The gap between the two tells you whether to fix your pages or your wider footprint.
How often should I run the check?
Monthly, with the same questions worded the same way. Answers vary from run to run, so a single check is a snapshot; the change between two careful checks is the signal.
Throughline reads across every tool you already pay for and narrates the part that moved. Request a run to see your own line.