“I asked ChatGPT and it put us first.” Whenever we hear this, we ask the same question: “How many times did you ask?” Ask the same question an hour later, from a different account or in slightly different words, and the answer can change. A single screenshot is proof of neither good news nor bad news.
In this article we explain how we measure visibility in AI answers. It does not need an expensive tool. A spreadsheet, some discipline and a few hours a month are enough.
Why a single screenshot misleads
AI answers vary for several reasons:
- Generation involves randomness. Language models do not have to produce the exact same text every time. Getting a different answer to the same question is not a bug, it is how the method works.
- Live search results change. If the tool searches the web before answering, that day’s search results shape the answer.
- Context matters. Earlier conversation, the user’s location, account settings and the tool’s version can all change the answer.
- Wording matters. “Best” versus “reliable”, “London” versus “north London” bring different answers.
So measurement should look at the average of many answers, not a single one.
Step 1: Build a fixed set of questions
Write 20-30 questions your customers could realistically ask. Grouping them helps:
- Category questions: “Which companies offer industrial refrigeration service in Ankara?”
- Comparison questions: “What should I look for when choosing a stainless steel tank manufacturer?”
- Brand questions: “What does company X do, and where is it located?”
- Problem questions: “The temperature in my cold storage room keeps fluctuating, what should I do?”
Once the set is fixed, do not change it. If you want to add new questions, add them as a separate group so you don’t break the comparison with earlier months. The value of the measurement is in seeing how the same questions change over time.
Step 2: Record each engine separately
ChatGPT, Gemini, Perplexity, Claude, Copilot and Google’s AI Overviews draw on different sources and give different answers. If you lump them into one number, you cannot see which tool has the problem.
For each question in each engine, note:
| Field | Example |
|---|---|
| Date | 2026-09-15 |
| Engine | Perplexity |
| Question | Companies offering industrial refrigeration service in Ankara |
| Were you named? | Yes |
| Was your site cited as a source? | No |
| Competitors named | Company A, Company B |
| Was the information correct? | Phone number outdated |
| Sources shown | A directory site, a news site |
If possible, ask each question 2-3 times per engine, starting a new conversation each time. Asking while logged out, or from an account with personalisation turned off, makes the results more comparable.
Step 3: Calculate two simple metrics
Citation rate
Citation rate shows in how many of your answers your name or site appears.
Citation rate = Answers that mention you / Total answers
We suggest tracking it two ways: your name appearing in the text, and your site being shown as a linked source. They are not the same. The AI may mention you while taking the information from a directory site.
Share of voice
Share of voice shows your portion of all mentions of you and your competitors across the same question set.
Share of voice = Your mentions / Total mentions of you and the competitors you track
Worked example
The table below is purely an example and does not belong to any real business. 20 questions, 3 engines, each question asked once, giving 60 answers.
| Company | Answers naming it | Citation rate | Share of voice |
|---|---|---|---|
| Your company | 12 | 12 / 60 = 20% | 12 / 66 = about 18% |
| Competitor A | 30 | 30 / 60 = 50% | 30 / 66 = about 45% |
| Competitor B | 24 | 24 / 60 = 40% | 24 / 66 = about 36% |
The citation rates add up to more than 100% because a single answer can name several companies. Share of voice is calculated against the total mentions of all three (12 + 30 + 24 = 66). In this example you are named in one answer out of five, Competitor A in one out of two. Citation rate shows your own position, share of voice shows where you stand against competitors.
The academic versions of these metrics are more detailed. The GEO paper (Aggarwal et al.) uses metrics that factor in position within the answer and word count. For a small business, a simple count is enough to start with. What matters is using the same method every month.
Step 4: Repeat monthly
Weekly measurement produces unnecessary noise for most businesses. Repeating once a month, in the same date range, with the same question set, is usually enough. If you make a major change to your site (new service pages, a move, a domain change) you can add an extra round in between.
Be careful when reading the changes. A citation rate moving a few points from one month to the next may mean nothing by itself. A change that holds in the same direction for three months in a row is worth attention.
Step 5: Track whether your own site is cited
Being named is good, but what you really want is for the information to come from your own site. A site shown as a source controls the information and can get the click.
There are a few ways to track this:
- The source list in the answer. Perplexity, ChatGPT search and Google’s AI Overviews show sources next to the answer. Record which of your pages appear.
- Bing Webmaster Tools. In February 2026 Microsoft opened its AI Performance report as a public preview. It shows how often your content is cited in Copilot and Bing’s AI-generated answers, and which pages stand out. It only covers Microsoft’s side, but it is free, first-party data.
- Google Search Console. Google says traffic from its AI features is included in the Performance report under the “Web” search type. There is no separate breakdown, but you can follow the overall trend.
- Your analytics tool. Visits from tools such as ChatGPT and Perplexity often show up as referrals. Track them in a separate segment.
Sometimes the reason you are not cited is technical. If search bots cannot reach your site, your pages are never read during live search. See our article on AI crawlers and llms.txt for more.
Step 6: Log wrong information separately
Often the most useful part of the measurement is not the numbers but the mistakes. An old phone number, a closed branch, a service you never offered, a wrong founding year, being confused with another company. Information the AI makes up and states with confidence is called hallucination.
For each mistake, note:
- Which engine, which question?
- What is the wrong information, and what is correct?
- Did the answer cite a source? If so, does that source contain the same error?
If a source was cited, your job is easier: try to get the information on that page corrected. Sometimes it is an old page of your own, sometimes a directory, sometimes a Business Profile nobody updated. If there is no source, the error most likely comes from the model’s training data. That cannot be fixed overnight, but publishing the correct information consistently on your site and elsewhere helps over time.
Summary
- Don’t trust a single screenshot. AI answers vary.
- Build a fixed set of 20-30 questions and record each engine separately.
- Calculate citation rate and share of voice with a simple count, and use the same method every month.
- Track separately whether your site is cited and what wrong information appears.
- You can run this measurement yourself. If you’d like us to do the first round, apply through the free analysis form. Ongoing tracking is part of our visibility service.