Most AI visibility reporting is not wrong so much as unfalsifiable. A composite "visibility score" moves and nobody can say why, or whether it would move again if you ran it twice. This is the method we use instead. It is not proprietary and you can run it without us.
Step 1: build the prompt set
The prompt set is the whole experiment. Get it wrong and nothing downstream means anything.
Use the phrasing buyers actually use with a chatbot. People type differently
into an assistant than into a search box — longer, more conversational, and
loaded with constraints. "crm" is a search query. "what CRM should a 12-person agency use if we already run Xero" is a prompt. Your rank tracker
contains the former and almost none of the latter.
Good sources for real phrasing:
- Sales call recordings — the questions asked before a demo is booked
- Support tickets, especially pre-sales ones
- The "what did you consider" answers in win/loss notes
Cover the families that precede a purchase. In practice these are:
comparison (X vs Y), alternatives (alternatives to X), constrained
recommendation (best X for Y under Z), and validation (is X any good,
is X worth it). Informational top-of-funnel prompts are cheap to win and rarely
worth anything.
Size it for repeatability, not coverage. Sixty to two hundred prompts is enough. The number matters less than never changing it once sampling starts — adding prompts mid-programme makes the trend line meaningless, so new prompts go into a clearly separated second cohort.
Step 2: decide what you record
Per prompt, per engine, per run. This is the part people cut, and it is the part that makes the result diagnostic rather than decorative.
| Field | Why it earns its place |
|---|---|
| Mentioned | Named at all, with or without a link |
| Cited | Named with an attributed source link |
| Source URL | Which page earned it — tells you what to make more of |
| Competitors named | Share of voice needs a denominator |
| Verbatim answer | Lets you diff week to week, and catch misstatements |
| Sentiment | Being named unfavourably is a distinct problem |
Recording the full answer text is the highest-value field and the one most often skipped. Without it you can see that a number moved but never why.
Step 3: sample properly
Every engine, not one. Results diverge substantially between them, and a baseline drawn from a single assistant tells you about that assistant.
On a fixed cadence. Weekly is enough for most programmes. The cadence matters more than the frequency, because you are looking for a delta.
With sessions isolated. Personalisation and conversation history will contaminate results. Fresh session per prompt, no logged-in account carrying memory, and be aware that geography changes answers — pin a location if you serve one.
Repeat before you trust a change. These systems are non-deterministic. A prompt that named you last week and does not this week may be noise. Sample each prompt more than once per run, or treat single-run changes as unactionable.
Step 4: reduce it to two numbers
Everything above collapses to:
- Citation rate — of the prompts sampled, the share that named you. Per engine and overall.
- Share of voice — your mentions against those of a fixed, named competitor set. Fixed matters; a competitor set that changes when convenient is a reporting artefact.
Trend both monthly. Report the query-level detail beneath them so a movement can always be traced to specific prompts.
The mistakes that invalidate the result
Changing the prompt set. The most common one, usually well-intentioned.
Sampling once and calling it a baseline. One run is a snapshot of a stochastic system.
Counting brand-name prompts. Asking an engine about your company by name will usually get you named. It measures nothing about discovery. These belong in a separate accuracy check, not in the citation rate.
Reporting a composite score. If a number cannot be decomposed into which prompts moved on which engine, it cannot be acted on.
Attributing through last-touch analytics. Referral traffic from AI surfaces exists and is worth tracking, but it undercounts badly. Most of the effect arrives as a buyer who is already convinced. Add a "how did you hear about us" field — for most companies that is where it first becomes visible.
What to do with the result
The baseline tells you which of three situations you are in.
If you are absent across the board, check crawler access before anything else. Robots rules excluding the retrieval agents are the most common single cause and cost nothing to fix.
If you are mentioned but rarely cited with a link, the gap is usually extractability — the model knows of you but your pages give it nothing specific to attribute.
If competitors are consistently named and you are not, look at what is being cited. It is frequently a third-party page neither of you owns, and the work is to be present on those sources rather than to publish more of your own.