Asking ChatGPT "do you know my brand?" one afternoon is not measurement. It is an anecdote. Measuring means repeating the same prompts under the same conditions and logging how the answers change.
Why a manual spot check misleads you
Three reasons that improvised check gives you a false reading:
- Personalisation: your account carries memory and history, so the model treats you differently from a first-time user.
- Variance: the same prompt yields different answers across runs. One sample is not data.
- Wording bias: name your brand in the prompt and the answer will name it back. That is echo, not visibility.
Step 1: build a prompt set that never names you
Write eight to twenty questions a real buyer would ask before they know you exist. Five families:
- Category: "best [category] tools for [company type]".
- Comparison: "[competitor A] vs [competitor B]: which should I choose?".
- Alternatives: "alternatives to [market leader]".
- Use case: "how to solve [specific problem] at a [industry] company".
- Purchase intent: "which [category] tool should I buy on a budget of X?".
None of them should contain your name. That is the whole point.
Step 2: fix the test conditions
Write down and always reuse the same setup: a session with no personal history, one language, one market, one model. Change a variable and last week's comparison stops being valid.
Step 3: log five metrics per run
For each prompt record:
- Do you appear? Yes or no.
- Position within the list of brands mentioned.
- Context: the literal sentence describing you.
- Competitors named in that same answer.
- Cited sources, when the assistant browses.
With twenty prompts you have a mention rate and a competitive share from day one.
Step 4: repeat weekly, not daily
AI visibility moves slowly: it depends on published content, third-party mentions and model updates. Weekly is the right cadence; daily only manufactures noise.
Step 5: turn the number into a cause
When you drop out of an answer, the cause is usually one of four:
- You have no page that answers that question directly.
- The sources the AI cites (comparisons, roundups, forums) do not include you.
- Your site is not legible: no structured data, no plain-text answers, AI crawlers blocked.
- Your category is defined inside the model with vocabulary you never use.
What to automate, and what not to
Automate execution, logging and trend comparison — doing it by hand in a spreadsheet works for a month and gets abandoned in the second. Keep the qualitative read manual: reading ten full answers a month teaches you more about how the model frames you than any chart.
AskBrandly automates exactly that layer: it reruns your prompt set weekly across several assistants, stores the history, benchmarks competitors and alerts you when something shifts. You can start with a free audit of your domain.
For the next step, read which prompts you should monitor and how to optimise your site for ChatGPT.