How to check what a machine says about you
Four questions, two languages, twenty minutes. The whole method for finding out how assistants describe your company, and what to fix first depending on what comes back.
August 5, 2026
You can find out what machines say about your company this afternoon. It takes about twenty minutes, costs nothing, and almost nobody has done it. Here is the whole method.
I am not going to pretend this is a science. It is closer to taking a temperature: you ask a set of questions, write down the answers and the date, and repeat it later to see whether anything moved. That is a low bar. It is also more than most organizations know about themselves right now.
Ask four kinds of question, not one
Most people test this by typing their own name, deciding the answer looks fine, and stopping. That tells you the least useful thing. Somebody who already knows your name was never the problem.
- The name question. What is [your company]? This checks whether a model can resolve you at all, and whether it has you confused with someone else who shares the name.
- The category question. Who does [what you do] for [who you serve]? You are looking for whether you appear unprompted, and who does appear instead. That list is your real competitive set, and it is often not the one you have in your head.
- The problem question, phrased the way a buyer would actually phrase it. Not your service names. The thing that is going wrong for them at 11pm.
- The comparison question. How does [you] compare to [a competitor]? Uncomfortable, and the fastest way to find out what the internet thinks your weakness is.
Run each one in at least two different assistants. They do not share an index and they will not agree, which is itself information.
Ask again in every language your customers use
If part of your market reads in Spanish, run the whole set again in Spanish. Not translated by you into English-shaped Spanish. The words a person would actually type.
In my own testing the two languages routinely return different companies for the same question. There is far less structured Spanish-language material about most US organizations, so the model has less to work with and leans harder on whatever it can find. Sometimes that is a directory entry from 2019.
Four things go wrong, and they need different fixes
When you read the answers back, sort what you find into four piles. They look similar and they are not.
- Absence. You are simply not in the answer. Usually this is not a content problem. It is that nobody else on the internet has written about you, so there is nothing to retrieve.
- Staleness. You are there, described as you were two years ago. Your site has moved on and nothing else has. This is the most common one and the most fixable.
- Conflation. You have been blended with another organization sharing your name. Fix this with consistent naming and links out to profiles that are unambiguously you.
- Invention. The model states something about you that is not true anywhere. Rarer than people fear, and the hardest to correct, because you are competing with a confident guess.
Write down the answer, not your reaction to it
Paste the reply verbatim into a document with the date, the assistant, and the exact question. Do not summarize it. Summaries lose the wording, and the wording is the thing that changes.
Do it again in a month. The value is not in any single answer, which will be noisy. It is in the direction of travel over three or four rounds, which is the only honest way anyone can currently claim this is working.
20 minutes is the whole exercise for one organization, in two languages, across two assistants.
Four questions, run twice, written downFix in this order
Whatever you found, the sequence is the same, because each step makes the next one work better.
Start with staleness, because it is the cheapest thing on the list and the most embarrassing to leave. Make sure the plain description of what you do exists somewhere obvious and current.
Then fix conflation: one spelling of the name, one canonical address, links out to the profiles that already exist for you. This is an hour of work and it is the step almost nobody has done.
Absence comes last, because it is the slow one. Being in the answer means other people have written about you, and that cannot be shortcut with markup. Partner pages, directories, coverage, customer stories. It compounds over months.
What this does not tell you
This method has real problems and you should know them. Answers vary between runs, so a single test tells you very little. Assistants personalize, so what you see is not exactly what a stranger sees. And none of it gives you a number, because there is no number to give.
What it does give you is the difference between guessing and knowing. Right now most organizations have never once looked at what a machine says about them, which means their entire understanding of their own reputation stops at the edge of their website.
Twenty minutes. Four questions. Write it down.
