How visible is your business in AI answers? The self-check
Want to know whether ChatGPT, Gemini or Perplexity know your business? A single attempt tells you very little, but a planned self-check tells you a great deal. This guide shows you step by step how to capture a reliable starting point using real customer questions.
pr-geo.com editorial team · 7 min read · As of 9 November 2026
Key points at a glance
- Do not test with your company name but with the questions customers ask before they buy.
- A fixed set of around 20 questions per system makes results comparable over time.
- For every answer, note the language, region, date, system, search mode, the exact prompt and the sources shown.
- Count brand mentions and source links separately, because each can occur without the other.
- Every result is a snapshot, and no cause can be automatically derived from it.
Why a spontaneous test is misleading
Almost everyone has tried it: type your own company name into ChatGPT and see what comes back. That is understandable, but it says little about your visibility. Anyone who already knows your name will find you anyway. The interesting moment comes earlier, when someone has a problem and no provider in mind yet.
On top of that, AI answers fluctuate. The same question may be answered differently tomorrow, depending on the system, the search mode, the wording or the location. A single hit is therefore neither a success nor a reason to worry. The test only becomes meaningful when you plan it, log it and repeat it under the same conditions.
Step 1: Choose the systems and set the framework
First decide which systems you want to check. It makes sense to pick the ones your customers are likely to use, for example ChatGPT, Gemini, Perplexity, Microsoft Copilot and the AI Overviews in Google Search. Two or three of them are enough to start with.
Then set the framework that you will keep for every repetition:
- Language: the language your customers use, plus a second one if you sell internationally.
- Region: the place you are testing from, or the region you name in the question.
- Search mode: note whether web search was active, as far as the system shows this or lets you set it.
- History: ideally a new chat without earlier conversations, so that previous questions do not colour the answer.
Before you start, it is worth taking a quick look at the technical side of your website.
So ask your web service provider whether robots.txt is accidentally blocking search crawlers. If it is, that explains some test results before you even start thinking about content.
Step 2: Put together the question set
The heart of the self-check is a fixed set of questions. The study on pr-geo.com suggests using around 20 real customer questions per system. Real means phrased the way customers actually ask, not the way you would describe your service in a brochure. Good places to look are emails, phone notes, sales conversations and reviews.
Spread the questions across five types so that you cover different situations:
| Question type | What it stands for | Sample question |
|---|---|---|
| Problem question | The customer knows the problem but not the solution | “What can I do if my patio stays wet for a long time after rain?” |
| Provider question | The customer is looking for a service provider | “Who lays natural stone patios in the Augsburg area?” |
| Comparison question | The customer is weighing up options | “Natural stone or wood for the patio: which lasts longer?” |
| Cost and process question | The customer is making concrete plans | “What does the cost of a new patio depend on?” |
| Brand question | The customer already knows you | “What does company X offer, and where is it based?” |
With 20 questions, four per type is a good split. The brand questions show whether the facts about you are correct. The other 16 show whether you appear when nobody knows you yet.
Phrase each question once, in its final form, and save it in a file. Every time you repeat the test, paste it in word for word, because even one different word can change the answer.
Step 3: Log every answer
Set up a table, ideally in a spreadsheet. Each question and system gets its own row. These columns make a solid framework:
| Column | Content |
|---|---|
| Date and time | When you ran the test |
| System and search mode | For example Perplexity with web search |
| Language and region | German, test location Augsburg |
| Prompt | The question in its exact wording |
| Brand mention | Yes or no, plus where it appears in the answer text |
| Source link | Yes or no, plus the linked address |
| Other sources | Which other providers or pages appear |
| Factual accuracy | Are the details about your business correct? |
| Valid | Yes or no, for example if there was an error message or the answer broke off |
| Note | Anything unusual and the file name of the screenshot |
Take a screenshot of every answer or copy the text including the sources. Answers often cannot be called up again later, and without evidence every evaluation turns into a question of memory.
Count brand mentions and source links separately
This is where a reasoning error easily creeps in. An answer can mention your business in the text without linking to any of your pages. Conversely, an AI can show an article about you as a source without mentioning your name in the answer text. These are different signals, and they belong in separate columns.
That gives you four cases:
- Mention with source link: your business appears in the text, and a page by you or about you is linked.
- Mention without source link: your name comes up, but a matching source is missing.
- Source link without mention: a page by you or about you serves as a source, but your name does not appear in the text.
- Neither: the answer does without you.
For source links, also note what kind of page is linked: your website, an expert article in a magazine, an industry directory or something else. That way you can later see which publications in your subject area appear as sources at all.
The evaluation: mention rate as a snapshot
You can derive a simple metric from the log. For the mention rate, divide the number of answers with a brand mention by all valid answers in your test set. For source links, do the same calculation separately.
For every evaluation, state the absolute numbers and the number of repetitions. “3 out of 18, tested once” is more honest than a percentage without context.
Repeat the check at fixed intervals, for example quarterly or a few weeks after a publication. At the same time, record what else has changed: a website relaunch, new ads, seasonal demand. Only then can you cautiously interpret changes later.
What to do with the results
The self-check is not a report card but a map. It shows you which questions you are missing from, where details about you are inaccurate and which sources the systems draw on instead.
- Inaccurate facts: first check your website and your profiles for outdated details.
- Missing topics: questions without you are candidates for new expert articles. You will find inspiration in the ten topic ideas from your own business.
- Unfamiliar sources: look at which magazines and portals appear as sources in your subject area. This will help you choose media later.
If you first want to understand how press work and AI answers are connected, start with the introductory article on PR GEO. If you work with a PR agency, it can keep track of observations like these as a separate module in the coverage report. The complete measurement concept with a 90-day plan is in the secondary study, available to download.
Frequently asked questions
What is the minimum number of questions I need?
Ten well-chosen questions are enough for a first impression. For comparisons over time, the study recommends a fixed set of around 20 questions per system. More important than the exact number is that the set stays the same.
Should I ask each question several times?
That makes sense, because AI answers fluctuate. If you ask each question three times, for example, the picture becomes more stable. Note the number of repetitions in the evaluation so that you compare fairly later.
What do I do if an AI gives incorrect details about my business?
Record the answer with the date and a screenshot, and check whether outdated information is published anywhere. Possible causes are old profiles, directory entries or contradictory details on your own website. Correct these sources and test the same question again later.
Can I hand the self-check over to my agency?
Yes, as long as the rules are clear: a fixed question set, the same conditions, separate counts for mentions and source links, complete evidence. Ask for the raw log and not just a summary. That way you can verify the results yourself at any time.
Sources
- Google Search Central: AI features and your websiteaccessed 9 November 2026
- OpenAI: Overview of OpenAI Crawlersaccessed 9 November 2026
Next step
If you want to get started
We show you which packages suit your business, with prices and services compared.
Provider disclosure: pr-geo.com is a knowledge portal of NexTao GmbH. pr4geo belongs to our group, which is why we expressly recommend it here. No package guarantees a mention in AI answers. Our principles
