How do I measure whether AI is recommending my brand?

Rank tracking gave the industry twenty years of comfortable certainty. One keyword, one number, same for everyone, checkable daily. Generated answers destroy that. The same question produces different answers for different users, in different sessions, on different model versions.

So the honest starting point is this: you cannot measure AI visibility with rank-tracker precision, and any tool promising that is selling you a false sense of control. What you can do is measure it well enough to make decisions, which is a lower bar and a genuinely achievable one.

The four measures worth running

1. Citation frequency on a fixed prompt set

Write down twenty to thirty prompts a real buyer would type. Not keywords, prompts. "Who are the best commercial interior fit-out firms in Kuala Lumpur." "I need an industrial supplier in Penang for X." "What should I look for when choosing Y in Malaysia."

Run the same set across the major answer engines on a schedule. Monthly is enough. Record whether you appear at all. The number that matters is what percentage of your prompt set names you, tracked over time.

Keep the prompt set frozen. The moment you start editing prompts, the trend line becomes meaningless.

2. Share of voice against a named competitor set

On the same prompts, record which competitors appear. Your share of voice is your appearances divided by total brand appearances across the set.

This is the more useful number of the two, because it separates your performance from category-wide changes. If everyone's citation frequency drops because a model started answering more conservatively, share of voice stays honest.

3. Branded search volume

The most underrated proxy available, and the one most people already have data for. When someone encounters your brand inside an answer, a meaningful share of them go and search your name afterwards. They do not click through from the answer, so no referral is recorded, but the intent shows up as branded search two minutes later.

Pull impressions for queries containing your brand name in Search Console and chart the trend. Rising branded impressions with flat non-branded impressions is one of the cleanest available signals that your generative visibility is improving.

4. Structural readiness

A scored baseline of the things you fully control: crawlability, schema, entity consistency, passage quality, crawler access. This is the only one of the four you can move directly and see change quickly, which makes it the right leading indicator. The free scan produces one.

Putting it in a report someone will read

Report all five together. Any one of them alone tells a misleading story, and the last one is the only one your finance director cares about.

Three traps

• Chasing a single prompt. One query is noise. A frozen set of twenty-plus is a signal. Never let a board conversation hinge on one screenshot.
• Comparing across models. Different systems retrieve differently. Track each separately or your trend line is measuring model differences rather than your progress.
• Expecting monotonic improvement. Model versions change and answers move for reasons that have nothing to do with you. Judge quarterly, not weekly.

Attribution, honestly

You will not get clean attribution from answer engines. Most do not pass referral data in a usable form, and even when they do, the person who read about you in an answer often searches your name and arrives via organic instead.

So add one question to every enquiry form and every first call: how did you hear about us. Self-reported attribution is imprecise and it is dramatically better than nothing. It is also the only channel that will tell you someone found you through a chatbot, because nothing in your analytics ever will.