Measuring your AI visibility: a repeatable method
Build your list of questions
You don’t measure AI visibility by typing in your own business name. Your future customer doesn’t know you yet: they describe what they need. Write 20 to 30 questions the way a customer would ask them, split into four groups:
- Finding a provider: “Which employment law firm in Bristol for a small business?”
- Comparison: “What’s the difference between an online accountant and a local practice?”
- Problem to solve: “How can my building firm cut down on late payments?”
- Brand check: “What do you think of [your business]? What are its prices?”
Vary the location (town, district, country), the size of the customer and the budget. These questions become your baseline: don’t change them afterwards, or two measurements won’t be comparable.
Test under the right conditions
- Use a neutral session (private browsing or an account with no history) to limit personalisation.
- Check that web search is switched on. Without it, you’re measuring the model’s memory, not what it finds today.
- Ask every question on every assistant: ChatGPT, Gemini, Claude, and Perplexity if your customers use it.
- For each answer, note: are you cited, in what position, with what description, which competitors appear, which sources are linked.
- Record any factual errors: address, opening hours, services, prices, directors.
- Repeat at a fixed interval (monthly, for example) with exactly the same questions.
Traps that skew the results
- Drawing conclusions from a single try: answers vary from one run to the next. Think in terms of presence rate across the whole set of questions.
- Only testing your brand name: you’ll almost always be found, which says nothing about whether you get recommended.
- Confusing memory with search: an answer without sources reflects the model’s training, which may be several months old.
- Changing the questions along the way: any change over time becomes impossible to interpret.
- Forgetting competitors: knowing who is named instead of you matters as much as your own score.
You can run this method perfectly well by hand in a spreadsheet. If you’d rather automate it, Saykora offers a free two-minute test (score with no sign-up, details with a work email) that asks ChatGPT, Gemini and Claude with web search on and shows their answers word for word; you can look at an example report.