Short answer

We see the same seven mistakes again and again, and they all share a root: treating a generated answer as if it were a fixed result. You avoid them with unbranded questions, a clean chat, more than one round, errors recorded as errors, sources read as clues rather than causes, every figure shown with its denominator and the original answer at hand.

1. Asking with your brand name

“What do you know about My Company?” measures whether the model can complete a sentence. It does not measure whether a buyer who does not know you would find you. Discovery questions carry no brand; branded questions are measured separately and serve a different purpose (representation).

2. Measuring only once

One round describes one day. Answers change with randomness, the index, the session and the wording. Two rounds a week apart are the minimum to label something as stable or variable.

3. Using an account with history and memory

If you ask from the account where you have been talking about your business for months, the model knows you, not your market. Use a temporary chat or a clean session, and note whether web search was turned on.

4. Treating errors as zeros

A timeout, a block or an empty answer does not prove that you are absent. They are recorded as errors, excluded from the denominator by a rule set in advance, and repeated.

5. Confusing citation with cause

A page appearing as a source next to an answer means it was retrieved that day. It does not prove that it caused the mention, or that it is always used. It is a clue about where to be, not an explanation.

6. Calling a sample market share

Twelve questions chosen by someone, on two platforms, on one day, are not a representative survey. “It appeared in 8 of 22 answers” is honest. “We have 36% AI visibility” is not. The mention rate is always published with its two numbers.

7. Hiding the original answer

A score without the rows behind it cannot be audited. If the report does not let you open the answer that produces each figure, you do not have data: you have to trust whoever signs it.

A checklist before you start

  • Unbranded questions, with their origin and version.
  • Platforms, language and country fixed and noted.
  • A clean chat, with search noted.
  • At least two dates.
  • An error rule decided before seeing any results.
  • Numerator and denominator for every figure.
  • The full answer and its sources saved for each observation.