Spoke
Prompt panel tracking
Asking an assistant about yourself once tells you nothing. A fixed set of prompts, re-run on a schedule and recorded, tells you a direction — which is the most that is available.
A practitioner method, not a platform-documented one. It is described here with its limitations stated because the alternative in circulation is single-run anecdotes presented as data.
The most common way people check AI visibility is to open an assistant, ask about themselves, and react to what comes back. Generative answers vary between runs, so that produces a result that is real, unrepresentative, and frequently acted on.
A panel fixes the obvious problems cheaply.
How do you track whether AI engines mention your business?
Build a fixed panel of prompts a real prospect would ask, run each several times across the engines that matter, on a fixed schedule, from clean sessions with no history. Record four things per run: whether you were mentioned, whether you were cited with a link, which competitors appeared, and whether the characterization was accurate.
The output is a trend line, not a rank. Treat a single run as an anecdote, because that is what it is.
Choosing the prompts
The panel should be questions a buyer actually asks, not questions designed to make you appear.
- Category questions. "Best personal injury firm in Durham" — high value, high competition.
- Problem questions. "What do I do after a car accident that wasn't my fault" — where you get cited for being useful rather than for being named.
- Comparison questions. "X versus Y" — the shape with a buyer behind it.
- Brand questions. "What is [your firm]" — tests whether you are described correctly, which is a different problem from visibility.
Keep it small enough to actually re-run. A panel of a few dozen prompts that gets run every month beats two hundred that get run once.
Recording the right four things
Collapsing this into "did we appear" loses most of the information.
| Recorded | Why it is separate |
|---|---|
| Mentioned | The model knows you — possibly from training, possibly retrieved |
| Cited with a link | Retrieval found your page specifically |
| Competitors named | Share of voice is relative; absolute presence means little |
| Characterization accurate | A confident wrong description is worse than absence |
That last row is the one teams skip and the one that most often produces an urgent action item.
Running it honestly
Clean sessions. Your own account has history, memory and personalization. Of all the places to measure how you appear to strangers, it is the worst.
Several runs per prompt. Variance between runs is substantial. A single run tells you something happened once.
Same dates, same cadence. Monthly is enough for most. Quarterly is enough for small panels.
Record the date and the engine with every row. Behavior changes without announcement, and a result with no date is not evidence of anything later.
What it cannot tell you
It cannot tell you volume — how many people asked anything like your prompt. It cannot tell you conversion. It cannot distinguish an answer drawn from training data from one drawn from retrieval, except indirectly by whether a link appeared. And it cannot tell you why you were or were not chosen, because no platform publishes that.
What it can tell you is whether you are trending toward or away from being named, how you compare with the firms you compete with, and whether what is said about you is true. For a measurement that costs an afternoon a month, that is a reasonable return — provided nobody reports it as a precise number.
More in Measurement
AI visibility reporting
Google and Microsoft now report AI visibility directly, and OpenAI tags its referrals. That beats a third-party citation tracker — and each one measures something narrower than its name suggests.
The attribution gap
Being cited without being clicked is a real outcome that last-click reporting cannot see. It usually resurfaces later as direct or branded traffic, credited to the wrong thing.
AI share of voice
How often you are named across a fixed set of questions, relative to the people you compete with. It is noisy, it is the closest thing available, and it breaks the moment anyone reports it as a single number.
AI referrers
The visitors who do click from an AI answer are identifiable in ordinary analytics. It is the easiest measurement win available here, and it counts the smallest part of the channel.
Self-reported attribution
One required field on every intake. It is biased, imprecise, and the only place a human tells you directly that a machine recommended you.
KPIs that survive
Sessions as a headline metric tells a story that gets less true every quarter. These are the measures that keep meaning what they meant when the traffic mix changes underneath them.