Share-of-Model: measuring whether AI even mentions you
The single most useful GEO metric is also the easiest to fake. Here's a defensible way to sample the prompt space, log citations and report a number you can stand behind.
A share-of-model figure is only credible if its prompt sample, engine set and date window are disclosed alongside it.
Share of model answers a deceptively simple question: when your buyers ask an AI engine about your category, how often does it mention you — and how does it describe you? It's the GEO equivalent of share of voice, and it's becoming the number boards ask for first.
The trouble is that the figure bends entirely to whoever picks the prompts. Ask flattering questions and your share looks heroic; ask neutral ones and it collapses. A defensible methodology removes that discretion.
Define the prompt space first
Before sampling anything, enumerate the realistic questions a buyer would actually ask. Group them by intent — discovery, comparison, validation — and weight each group by how often it occurs in real demand.
"If you can't show the prompts, you can't show the share. Sampling is the methodology."
Sampling that's honest
Run each weighted prompt across every engine in scope, repeated enough times to smooth out non-determinism. Record presence, citation and sentiment per response. Below: a representative run across five engines.
Report the blended figure and the spread. A 48% average hides the fact that one engine never mentions you — and that gap is exactly where the optimisation work lives.
Logging citations
For each prompt-engine pair, record three things: (1) presence — were you mentioned at all? (2) citation — was your domain linked or attributed? (3) sentiment — was the characterisation accurate and positive, neutral, or negative/hallucinated?
Reporting the number
A share-of-model figure is only credible if you disclose five things alongside it:
Without these five, a share-of-model number is marketing, not measurement.