In short: The most reliable way to test whether ChatGPT knows and recommends your company is six fixed questions that you ask, unchanged, every month: four questions asking for providers of your services in your region, one subject question from your field and one question about your company directly. Each question goes into a fresh temporary chat without personalisation, every answer that contains your company name scores one point, and you also note who gets named instead. A single answer is a sample; only the same questions under the same conditions turn it into a measurement series.
How do I test whether ChatGPT knows my company?
With the questions a customer would ask, not the ones you would ask yourself. “What do you know about Example Ltd?” only checks whether a name can be found. Customers ask for a provider of a service, usually with a place and almost never with a company name in mind. A useful test therefore contains both kinds of question, with a clear majority of market questions.
The quick way in is the AI visibility check: five test questions, a rating per answer, a score from 0 to 10. That is a good snapshot. If you want to see whether anything moves over time, you need AI monitoring with a protocol that runs the same way every month. I have measured my own AI visibility monthly since August 2026, using the question structure this article describes.
Which six questions belong in the protocol?
Six questions of three types, each phrased the way a customer would type it:
- Questions 1 to 4, market questions: one per service line, for example “Which consultants for [service] in [region] can you recommend?”. If you offer only one service, vary the region or the customer group instead.
- Question 5, subject question: a typical problem your customers have, for example “How can I solve [problem]?”. Nobody asks for a provider here, but the AI can cite a specialist source. Good specialist articles pay into exactly this type of question.
- Question 6, name question: “What do you know about [company]?”. It checks whether the AI knows your company and describes it correctly.
Three rules for the wording. The questions come from real enquiries, meaning emails and first conversations, not from your marketing copy. They include the place if your customers search regionally; how much regional answers differ is shown in the article GEO consultants in Lower Austria. And they are fixed once and never touched again. Swapping a question starts a new measurement series.
Under which conditions do you measure neutrally?
Under conditions in which the AI knows nothing about you before the question is asked. In my measurement series, ChatGPT in the normal, logged-in chat drew on knowledge from the account's memory without being asked. Four conditions keep that out:
- A temporary chat, and one without personalisation. Temporary chats do not create new memories. According to OpenAI, however, they can use existing memories and custom instructions unless you switch personalisation off when starting the chat, and the choice cannot be changed afterwards (source: help.openai.com, Temporary Chat FAQ, as of September 2026).
- A fresh chat for each question. Otherwise the AI answers the fourth question in the context of the first three.
- Paste the wording rather than typing it, from a file, and check before sending that exactly this text is in the input field. With automated pasting, umlauts have already gone missing in my measurement series.
- Note the plan and the model. According to OpenAI, the default model and the limits on the free plan can change at any time (source: help.openai.com, ChatGPT Free Tier FAQ, as of September 2026). An answer from a different model is not a comparable value.
And only read once the answer is complete: the list of sources sometimes appears before the actual text.
How do you score the answers?
With a score nobody has to interpret: one point per question if your company name or your domain appears in the answer, otherwise zero, so six at most. In the name question your name is already part of the question; only the answer counts, and “I have no information about that” is not a mention.
Gradations such as “mentioned positively” or “named in third place” sound more precise, but they are a matter of judgement, and judgement makes months incomparable. Always write the result with its denominator, “2 of 6” rather than “named twice”.
How do you track AI mentions over months?
In a simple table with one row per question per month. These columns are enough:
- date, service, plan and model, plus anything unusual such as a notice about a limit
- question number and score, meaning 0 or 1
- the providers named, always spelled the same way
- the sources cited, meaning the pages the answer links to
- whether a map box was displayed
The last three columns are worth more than the score. Whoever appears for the same question over several months evidently owns the source the AI cites. Whoever appears for the first time has just done something right. You standardise the spelling because AI services write the same provider in different ways; otherwise you miss exactly the recurring names.
The sources column shows where you are missing, and for regional questions that is often outside your own website: for two of the six questions in my protocol, ChatGPT displayed a map box, the local pack, with listings from the Google business directory in September 2026. If a measurement condition changes, a new series begins, and values from before and after the change are not read as a trend.
How often should you measure AI visibility?
Once a month, ideally at the same time of day. AI assistants take up new sources more slowly than a search engine, effects tend to show after weeks to months, and weekly measurements mostly produce noise. Conversely, a single measurement is not enough, because the same question can produce different answers on different days. Three months is the minimum before you read a direction, and one point more or less on six questions is not yet a trend.
Does the protocol also work for Perplexity, Gemini and Copilot?
Yes, with one restriction: each service gets its own column, and values from different services are never added up into one number, because sources, map boxes and sign-in differ. Even within one service, logged in is not logged out. In my measurement series in August 2026, Perplexity named no provider at all for the same question when logged in, but did when logged out. Anyone checking visibility in Perplexity should therefore fix in advance which state they measure in. For the sources column Perplexity is particularly useful, because it shows the sources it used alongside its answers.
Can you do LLM monitoring yourself, or do you need a tool?
Do it yourself until you know which questions matter. Six questions a month take about half an hour, and you see exactly the interface your customers see. Tools that query models through programming interfaces can handle hundreds of questions, but they do not necessarily show the same answer as the app, for example without the map box. A tool pays off once the question set is fixed and you need more questions, languages or regions than one person can ask cleanly. Why you are missing is something no dashboard tells you; a first technical finding comes from the GEO scan in ten seconds.
Conclusion
AI visibility can be measured without a budget once three things are fixed: six questions word for word, neutral conditions and a score with a denominator. The real yield is not the score but the names and sources that get cited instead of you. The basics are in the article GEO: how your company becomes the answer AI gives, and how to recognise clean monitoring at a provider in the article GEO consultants in Austria. How I cut measurement, content and source building into a service is on the GEO page and under Performance Sales, and who stands behind it is in the profile.
Frequently asked questions about testing your AI visibility
How do I test whether ChatGPT knows my company?
In a fresh temporary chat without personalisation, ask your customers' questions: for providers of your service in your region, one subject question and one question about your company directly. Count how many answers contain your company name, and note who gets named instead.
Does the AI recommend my company if it knows it when asked?
Not necessarily. The name question only checks whether the AI knows your company. You are recommended once you appear in answers to market questions that do not contain your name.
How often should you measure AI visibility?
Once a month, with the same questions. A direction can be read after three months at the earliest.
Can I do LLM monitoring myself?
Yes. Six fixed questions a month and a table with score, providers named and sources cited are enough to start. A tool only pays off for more questions, languages or regions.
How do I check my visibility in Perplexity?
With the same six questions in a separate column, never added to the ChatGPT values, and in a state fixed in advance, logged in or logged out.
Note: the information on temporary chats and the free plan comes from the OpenAI Help Center (help.openai.com), retrieved in September 2026. Features and limits of AI services change frequently.