Measure your AI visibility before you change anything

The short answer

Before you edit a page, add a line of markup or write to a website, record where you stand today. Ask the same set of buyer questions to the same AI assistants, save every answer with the date, and note whether you were named. That file is your baseline. Without it, you will not know later whether anything changed.

It takes one to two hours the first time. Do it before the work, not after.

Why a baseline matters

AI answers move around. A study by SparkToro and Gumshoe asked volunteers to run the same prompts again and again. SparkToro reports that, for ChatGPT and Google's AI, the chance of two identical lists of brands in 100 runs was below 1 in 100, and that the same order came up about once in 1,000 runs. So one answer in March and one in June tell you very little. A set of answers taken on one day, and a second set taken later in the same way, tell you much more.

SparkToro's author calls the share of runs in which a brand appears a reasonable measure, and does not trust a tool that reports a single "ranking position". That is why the baseline records how often you are named, not just whether you were named once.

What to record

Keep it small. For each answer, write down:

• The date and the time.
• The assistant and, if it shows one, the model name.
• Whether web search or browsing was on.
• The exact question.
• Whether your business was named (yes or no).
• Your position in the list, if it was a list.
• Which competitors were named.
• The links the assistant showed as its sources.
• Anything about you that was wrong.

Copy the full answer text into a notes column or a separate file. You will want it later.

Which questions to use

Use 10 to 20 buyer questions: the things a real customer would type before they have heard of you. Do not put your own name in them. The next guide, "Write buyer questions that matter", shows how to build them. Add three or four questions about your own facts, such as your prices or your opening hours, to see whether the assistants state them correctly.

Do not change the questions after the baseline. If you change the wording, you are no longer comparing like with like.

How to take the snapshot

1. Open a fresh, signed-out or clean chat for each question. An old conversation can colour the answer.
2. Use the same settings every time. If you use web search for one question, use it for all.
3. Paste the question exactly as written.
4. Copy the answer and its links into your sheet.
5. Ask each question at least twice, in separate chats, and note both answers. You will usually see they differ. That difference is the point.

Gumshoe, a company that tracks AI visibility, says that one prompt on one model needs 97 answers to be within 10 percentage points, and 385 to be within 5 points, at 95% confidence and in the worst case. You will not get that by hand. A baseline from a few dozen answers shows direction, not precision. Treat it that way.

A worksheet you can copy

Make a spreadsheet with these columns, one row per answer.

date | assistant | model | search on? | question id | run (1, 2) | named? (Y/N) | position | competitors named | cited links | wrong facts | notes

Add a second tab with one row per question, and these columns:

question id | need | asked how many times | times named | share named | main competitor | main cited site

Fill in "share named" as times named divided by times asked. Write the date at the top of the sheet and call the file something like baseline-2026-10-15.

How often to repeat it

Repeat the whole exercise once a month, or after you have finished a batch of changes and given the assistants some time to catch up. Do not repeat it daily. You will only be measuring noise.

Keep every old sheet. Never overwrite the baseline.

How to check it worked

Your baseline is good enough when:

• Every row has a date, an assistant and the exact question.
• Every question was asked at least twice.
• Someone else could repeat the exercise from your notes alone.
• You have a "share named" figure for each question, even if it is 0.

What this won't do

A baseline does not say why you are or are not named. It also does not tell you what to fix. It is a starting line, nothing more.

Hand-run answers also depend on your own location, language and settings, which may differ from your customers'. And a handful of answers cannot tell a real change from ordinary variation. The next guide on tracking explains how to read the results without fooling yourself.

Our free check runs the same kind of snapshot for you on five buyer questions. The full audit does a larger one and keeps the dates and sources next to every answer.