How to Measure AI Visibility: A Baseline You Can Repeat

What a baseline is

A baseline is one complete measurement, taken before anything is changed: a fixed set of your customers' real questions, run across the platforms they use, with what happened recorded verbatim.

Without it, no later claim of improvement can be checked. This is not a technicality — it is the single most common reason GEO engagements end in an argument.

Step 1: build the question set

Three sources, in order of usefulness:

Sales and support conversations. The customer's own wording is already there. This is by far the best source.
The assistants themselves. Ask a broad question and see what follow-ups the answer raises.
Competitor content. What they explain is usually what customers ask.

Aim for 15 to 30 questions covering awareness, comparison and decision intent. Write them as questions, not keywords — models match on the question. More on why in the guide to building a semantic keyword system.

Step 2: freeze it

Write the list down, punctuation included, and store it separately from the keyword list you use for content planning. The content list can grow; the measurement list cannot change.

Changing the question set between rounds is the most common way GEO reporting becomes meaningless, and it usually happens by accident rather than by design.

Step 3: standardise the conditions

Record web-enabled and non-web-enabled runs separately. Do not average them.
Keep account state consistent — all logged in or all not.
Run one round within a day or two, not spread across weeks.
Run each question at least three times per platform and record the proportion of runs in which the brand appears. Generated answers vary; a single run is not a measurement.

Step 4: record four fields, no more

For every question-and-platform combination:

Named or not
Mention context — positive, neutral or negative
Other brands appearing in the same answer
Source domains cited by the answer

More fields and execution drifts by round three. These four support every conclusion worth drawing.

Step 5: compute three numbers

Mention rate — appearances divided by total combinations.
Share of voice — your mentions divided by all brands' mentions. This filters out category-wide movement.
Source coverage — the proportion of cited domains that are yours. This usually moves before mention rate does, which makes it the useful early signal.

Report them per platform. An average across six hides which work paid off.

Two rules that are easy to break

A negative mention is not a win. Being named as the cheap option or alongside a complaint has to be counted separately, or the headline number improves while the commercial effect gets worse.

Count symmetrically. Apply the same matching rules to competitors as to yourself, and exclude questions that name your own brand in the prompt — those are not a contest.

FAQ

Questions teams ask before they start

Do I need a tool for this?

No. A spreadsheet and an afternoon are enough for a first baseline. Tools help with repeating it at scale and with keeping the raw answers organised, but nothing about the method requires one.

How many questions is enough?

15 to 30. Fewer and a single question swings the percentage; more and the rounds stop getting run because they take too long. The consistency of the set matters far more than its size.

Should we include our brand name in the questions?

Keep those separate. Questions naming your brand in the prompt measure how you are described, not whether you are chosen; they are useful, but counting them in the mention rate inflates it and removes the competitive meaning.

How do we record the mention context?

Three values only: positive, neutral, negative, judged on what the sentence around the mention says. Being named as the cheap option or alongside a complaint is recorded as negative, not as a mention.

What if a platform gives a different answer every time?

That is expected and is why each question is run at least three times. Record the proportion of runs in which you are named rather than a single result; the proportion is the measurement.

Who should carry out the runs?

Whoever can keep the conditions identical across rounds. Consistency matters more than seniority: the same account state, the same run count, the same one-to-two-day window each round.

How do we store the raw answers?

Verbatim, with the question, platform, run number, date and whether search was enabled. Storing only the scored result removes the ability to check any figure later, which is the main thing a baseline is for.

Measuring your baseline