How do I run a repeatable AI visibility audit instead of asking ChatGPT once?
Fix the questions, fix the engines, and record the answer verbatim with the date. A one-off question tells you nothing because answers vary by phrasing and by day. A fixed panel of ten to twenty buyer questions, asked on a schedule, produces a mention rate you can compare week over week.
Why AI skips you
These 5 patterns, most frequent first.
Answers are phrasing-sensitive. Asking one version of a question and concluding you are invisible over-reads a single sample.
Answers move. A mention rate without a date and a stored answer cannot be compared to anything, so the audit becomes an anecdote.
Mentions and citations are different outcomes. Being named without a link means you are known but not sourced; being cited without being named means your page is a reference, not a recommendation.
Competitor presence is the useful half of the measurement. Knowing which domains get cited for your buyer's questions tells you which pages to build, which no single question reveals.
Personalisation makes single samples even less trustworthy. The same question can produce a different answer for two buyers in two countries, so the record has to include the market and the phrasing used, or two runs are not comparable.
What to put on the page
7 steps, in sequence.
Write ten to twenty questions in the buyer's own phrasing, covering selection, comparison, problem and price intent.
Fix the engine list before the first run. Changing the panel mid-series destroys comparability.
For each question and engine, store the full answer text, the cited URLs, the date and the engine version if shown.
Score each record on three separate axes: named, cited, and position of the first mention in the answer.
Attribute each miss to a gap type: absent from the answer, named but not cited, or cited but ranked late. Each type has a different fix.
Re-run the same panel weekly and require each fix to be validated by the next sample, not by your own opinion of the new page.
Record the market and the exact phrasing alongside every sample, so a change in mention rate can be traced to a change in the panel rather than assumed to be a change in visibility.
How to check AI's answer
Run these 2 checks to confirm.
Two consecutive runs exist with stored answers and a computed mention rate, so the trend is a number rather than an impression.
Every miss has a gap type and a named fix, and none of the fixes is optimise the content.
The checks as a tool
This is the specific step where GEO Visibility Audit takes over — it runs locally, one-time $49, no account. The reasoning stays in the steps above; the tool just repeats them. 20 weighted checks, 100 points, two axes: technical and semantic. One-time $27.55.
Questions we get
3 follow-up questions, answered.
How many questions are enough?
Ten to twenty across four intents. Fewer than ten cannot distinguish a content gap from noise, more than twenty becomes work you will stop doing.
Should I use an API or ask manually?
Either, provided the answer is stored verbatim. Manual sampling is slower but sees what a real buyer sees, including personalisation.
What mention rate is good?
There is no universal number. What matters is the direction across runs and the gap types you still have open.