How to measure AI search visibility

Measuring AI search visibility is harder than measuring rankings, because assistants do not give the same answer twice. A simple method, applied consistently, still gives you numbers you can trust. This is the method we use.

The problem with one answer

Ask an assistant a question and it may name you. Ask again and it may not. A measurement built on one answer per question swings from week to week for no reason, which makes it useless for decisions.

Step 1: choose the questions

• Twenty to forty questions that real buyers ask
• A mix of comparison, how-to, who-is-best and what-to-know questions
• Mostly questions that do not contain your name, plus a few that do
• A fixed list you keep for months, so results can be compared

Step 2: choose the assistants

Include the ones your customers use: ChatGPT, Claude, Gemini, Perplexity, Google's AI results and Copilot are a sensible set. Record which one produced each answer, because they differ.

Step 3: sample more than once

Ask each question several times each week on each assistant, starting each in a fresh conversation. Record whether you are named, in what position, and which sources are cited.

Step 4: calculate the share

1. For each question and assistant, find the share of answers that name you
2. Average those across questions, giving each question equal weight
3. Calculate a range around the result, so the uncertainty is visible
4. Compare the range with last week's

Step 5: read the change

• Overlapping ranges mean the change is probably noise
• A range that moves clear of the old one and stays suggests a real shift
• A shift across all assistants at once may be a model update, not your work

Step 6: track citations

Record which sources the assistants quote. Those are the pages and sites that shape the answer, so they show where to improve your own pages and where to earn mentions.

What to report

Report the share with its range, the questions where you are named by every assistant, some, or none, and the sources most cited. Leave out any single figure without a range.

Common questions

How many samples are enough?

Enough that the range is narrow enough to act on. Starting with three per question each week is sensible, and more for the questions that matter most.

Can I do this by hand?

For a handful of questions, yes. A spreadsheet and fresh conversations are enough to start. At twenty or more questions across several assistants, automation pays for itself.

Does it matter which country I test from?

It can. Answers vary by location and language, so test from where your customers are.