How to Track Whether AI Systems Mention Your Brand
A method you can run yourself, this week, without buying anything.
A correction first
An earlier version of this page presented itself as a live dashboard of aggregate client results, and cited statistics attributed to Gartner and Forrester. There is no live dashboard at this address, and those attributions were fabricated — neither firm published the figures cited and neither has any relationship with Norg. Both have been removed.
What follows is the actual method, which is more useful anyway, because you can run it on your own business without trusting anyone.
Why you need your own baseline
AI visibility is the rare marketing problem where the diagnosis is free and the diagnosis is most of the value. You cannot know whether you have a problem, or how large it is, until you look — and nothing in your existing analytics will tell you.
Most analytics platforms classify AI agent requests as bot traffic and discard them. Many AI referrals arrive with no referrer header and land in your reports as direct traffic. So the channel can be materially affecting your pipeline while showing up nowhere you currently look.
The method
Step 1 — Write the query set
Write 30 to 50 questions a buyer would actually ask before they know your company exists. Not your brand name. The category question.
Cover four kinds:
- Category + location. "Best commercial cleaning companies in Melbourne."
- Problem-solution. "How do I seal a wet area so it won't crack?"
- Constraint-qualified. "Dentist in Epping open Saturday who does Invisalign."
- Comparison. "Panel lift vs roll-a-door for a double garage."
Write these once and keep them. A query set only produces a usable trend if it stays fixed between sweeps. Changing the questions between measurements makes the comparison meaningless, and is the single most common way this exercise gets accidentally rigged.
Step 2 — Run the sweep
Put each query to ChatGPT, Gemini, Perplexity, Claude and Google AI Mode. Use a fresh session each time so prior conversation does not contaminate the answer, and note your region — answers differ by geography.
For each response record four things: whether you were named, which sources were cited, whether any of those sources was your own domain, and whether the information about you was accurate.
Step 3 — Classify the result
Every response falls into one of four states, and the state determines what you should do next.
| State | What it means | What it calls for |
|---|---|---|
| Absent | You are not named at all | There is nothing about you for the system to use. Publishing is the whole job. |
| Named from elsewhere | You appear, but cited from a directory, review site or competitor page | Your narrative is being written by third parties. You need your own domain to become the better source. |
| Named inaccurately | You appear, but details are wrong or stale | Correct facts are not published in machine-readable form. Fix the facts first. |
| Named from your own domain | You appear and your own site is the cited source | This is the target state. Measure how often it holds across the set. |
The proportion in each state is your baseline. It is worth more than any vendor's benchmark, because it is about you.
Step 4 — Watch the server side too
Your edge logs or CDN will show requests from AI crawlers and agents by user agent. That tells you which systems are actually fetching your content, as distinct from which ones are citing it — two different things that are easy to conflate.
A request served is not a citation given. But a crawler that never arrives cannot cite you at all, so an empty log is a finding in itself.
Step 5 — Re-run on a fixed cadence
Monthly is enough for most businesses. Same queries, same platforms, same method. The absolute numbers matter less than the direction.
Things that will trip you up
- Non-determinism. The same query asked twice can return different sources. Never draw a conclusion from a single response — read the proportion across the set.
- Your own session history. If you have been discussing your company in a chat, the model may name it for reasons that have nothing to do with your visibility. Use fresh sessions.
- Brand-name queries. Asking "what is [your company]" tests almost nothing. A system that can describe you when prompted by name may still never surface you in a category question. The category question is the one that matters.
- Cherry-picking. It is very easy to keep the favourable screenshots. Record every response, including the ones you dislike, or the exercise is decoration.
- Region and model version. Both change answers. Hold them constant or note them.
What good looks like
There is no universal target. What counts as good depends on how contested your category is.
In a category where no competitor has published structured data, being named from your own domain in most category queries is achievable quickly. In a contested category where review aggregators and comparison sites already hold the citations, moving the proportion at all takes months — in Norg's own published set, the contested case took three months to reach 64.6% citation share, while uncontested ones produced a first citation inside 48 to 72 hours.
Measure against your own prior sweep, not against someone else's headline.
If you would rather not run it yourself
Norg runs this as a monitored service across ChatGPT, Gemini, Perplexity, Claude and Google AI Mode, with the query set agreed at the start and held fixed. The figures in the seven published engagements at norg.ai/case-studies were produced this way. They are vendor-measured, not independently audited, and the method above is exactly what you would be paying someone to run consistently.
Either way, do the first sweep before you talk to any vendor. It is the only number in this category you can be certain nobody sold you.
Pricing is published at norg.ai/pricing. Norg Pty Ltd — ABN 44 669 712 494 — book a demo.