How to Track Whether AI Engines Mention Your Brand

Traditional search gives you instrumentation. Impressions, average position, click-through rate, referrer strings — you can see the shape of your visibility even when it is bad.

AI answers give you none of that. If ChatGPT recommends four competitors and not you, nothing anywhere records the event. You cannot fix what you cannot see, and by default you cannot see any of this.

Key takeaways

• AI mentions produce no impressions, no rank and usually no referrer. Passive analytics will not find them.
• The only reliable method is active testing: ask a fixed set of questions on a schedule and record the answers.
• One check is noise. Answers vary between runs for identical prompts; only the trend across weeks is meaningful.
• The metric that matters most is share of voice — not just whether you appear, but who appears instead of you.

Why can't you measure this passively?

Because in most cases there is no click. An assistant answers the question inside the conversation, names a few sources, and the user never leaves. No click means no referrer, no session, no analytics event — the mention happened and your instrumentation never saw it.

Some assistants do pass a referrer when a citation is clicked, so a trickle shows up in analytics. That trickle is a rounding error against the mentions that never converted into a visit, and treating it as your measurement is how teams convince themselves AI search does not matter yet.

What should you actually track?

Four things, in increasing order of usefulness:

The last one is the most actionable and the least tracked. If the same three domains keep getting cited for your core questions, you now know exactly what you are competing against and can go read what they did.

How do you choose which questions to track?

Track the questions a buyer asks when they do not yet know your brand exists. Category questions — "what is the best tool for X", "how do I fix Y" — reveal whether you are discoverable. Branded questions only tell you whether the engine knows about a company it was already handed the name of.

A workable starting set is 10–25 questions, split roughly:

1. Category questions — "best tool for auditing landing pages". Where discovery actually happens.
2. Problem questions — "why is my landing page not converting". Where you demonstrate expertise.
3. Comparison questions — "X vs Y". High commercial intent.
4. A few branded questions — to check the engine describes you accurately and is not hallucinating features you do not have.

That last category catches a real failure mode. An engine confidently misdescribing your product is worse than not mentioning it.

How do you run the test?

1. Write your question list and freeze it. Changing questions mid-stream destroys comparability.
2. Decide which engines matter to you — ChatGPT, Perplexity and Google AI Overviews cover most of it.
3. Run every question against every engine on a fixed schedule.
4. For each answer, record: was your domain cited, at what position, which other domains appeared.
5. Roll the results up daily so you have a trend rather than a pile of transcripts.
6. Review weekly. React to sustained movement, not single-day swings.

Doing this by hand for 20 questions across 3 engines is 60 prompts a day. That is exactly the kind of task worth automating, and it is why our AI Visibility module exists — you define the question set, it runs daily, and you get the trend and share-of-voice comparison instead of a folder of screenshots.

How do you read the results?

Look for direction and for who is beating you, not for absolute numbers.

• Visibility flat and low — usually a structural problem. Your pages are probably not extractable. Fix answer structure before writing more content.
• Visibility rising — whatever you changed is working; keep going and do not change five things at once.
• You appear but late — you are eligible but not the best source. Usually a depth or credibility gap.
• Same competitor always first — go read their page for that question. The gap is normally concrete and copyable.
• Answers describe you wrongly — an accuracy problem, not a visibility one. Your own pages are ambiguous about what you do.

What do you do when the number will not move?

Almost always, the page cannot be extracted from. Before writing anything new, check the pages that should already be winning:

1. Is the target question the actual heading?
2. Is there a 40–60 word self-contained answer directly under it?
3. Does FAQPage schema exist and parse?
4. Does Organization schema resolve who published it?
5. Is there a named author on advice content?
6. Do you cite any primary sources?

Our free AI citation check answers all six from your live HTML in a few seconds, no signup. It runs no AI models, so every result is something you can verify yourself in dev tools.

Start with the questions, not the tooling

The most common mistake is buying a tracker before deciding what to track. Write down the ten questions you most want to win. Ask them yourself, once, in ChatGPT and Perplexity. Write down what happens.

That hour will teach you more about your position than any dashboard, and it will tell you whether this is worth automating for your business yet.

Related reading: How to get cited by ChatGPT and AEO vs SEO.