Key facts
- Use the same prompts every time, or trends mean nothing.
- Ask each prompt more than once, because answers vary.
- Track each engine and market separately.
- Report ranges and real change, not single numbers.
Why you cannot just check by hand
Asking ChatGPT a question once and taking a screenshot feels like tracking, but it is not. The same question can name different brands a minute later. Answers also change with the country, the language, the wording and whether the engine searched the web. Real tracking means asking many questions, many times, in a consistent way.
Step 1: build your prompt set
- 1List your products or services and the markets you sell in.
- 2For each, write the buyer questions: best, top, versus, alternatives, pricing, how to and problem questions.
- 3Include questions without your brand name, because that is where new customers come from.
- 4Add a few branded questions to check how AI describes you.
- 5Tag each prompt by product, topic and funnel stage so you can slice results later.
Step 2: choose engines and markets
Track the engines your buyers use, usually ChatGPT, Gemini, Claude, Perplexity, Copilot and Google AI Overviews and AI Mode. Run prompts as a person in each country and language would ask, because local answers name local brands.
Step 3: record what matters in each answer
- Whether your brand is named, and in what position.
- Which competitors are named.
- How the answer describes you: positive, neutral or negative, and why.
- Which sources are cited, by domain and page.
- Any wrong facts about your brand.
Step 4: turn answers into metrics
- Visibility: the share of answers that name you.
- Average position: how early you appear when named.
- Share of voice: your mentions as a share of all tracked brand mentions.
- Sentiment: the tone of what is said about you.
- Citation rate: the share of answers that cite your site.
Show a likely range on each number. With a few dozen answers, a score of 30% might really be anywhere from 24 to 36%. Without the range, normal variation looks like progress or decline.
Step 5: report change honestly
Compare the same prompts between periods. Call a change real only when it is larger than normal variation. Report by engine and market, explain why numbers moved with the sources and pages behind them, and list the next actions.
Doing it with a tool
You can run a small version of this in a spreadsheet, but it gets heavy quickly: 100 prompts across seven engines, twice a week, is 1,400 answers. Tracking tools automate the asking, recording and scoring, and keep history so you can see trends.
Put this into practice with Variask AI visibility and Variask Prompt research. See Variask pricing or read the Variask FAQ.