What Is Share of Model?

Share of model is an emerging metric that estimates how strongly a large language model's internal, trained knowledge favors your brand relative to competitors. Where share of voice counts mentions in produced answers — a blend of memory and live retrieval — share of model isolates the parametric layer: if the model answers purely from what it learned in training, how often does it reach for you?

Why does the parametric layer deserve its own metric?

Because it behaves differently from everything else in the stack. Retrieval visibility changes in days and responds to publishing; model memory is frozen at each training run and responds only to what the corpus said about you during the training window. A brand can dominate live citations yet be unknown to the model itself — meaning it vanishes whenever the engine skips retrieval, which happens on a large fraction of conversational queries. Reported estimates of ChatGPT queries that trigger web search vary, but no engine searches on every prompt, so the memory layer answers a meaningful share of buyer questions on its own.

How is share of model estimated?

Since weights are inaccessible, estimation is behavioral:

1. Probe without retrieval — query models with browsing/search disabled (or via API models with no tools) using category prompts: "name the leading X tools," "compare A and B."
2. Repeat and randomize — dozens of runs per prompt, varied phrasing, to average out sampling noise from temperature and decoding.
3. Score volunteered brands — frequency, order of mention, and descriptive confidence per brand.
4. Normalize into a share — your brand's weighted mentions over the category total.

Trend it per model release: shares reshuffle when vendors ship new models, because each release trains on a newer corpus.

What moves share of model?

The training-corpus sources: Wikipedia and Wikidata presence, Reddit and community discussion (Reddit's licensing deal with Google, reported at $60M/year in 2024, made it a heavyweight corpus), sustained news coverage, and consistent brand descriptions across the open web. These compound slowly — months to years — which is why teams track share of model alongside faster retrieval metrics in a visibility program rather than instead of them.

Example

Probing GPT-class models with search off, a CRM challenger appears in 4% of "top CRM" completions while the incumbent appears in 96%. After a year of PR, Wikipedia presence, and community growth, the next model generation volunteers the challenger in 19% of runs — share of model moved with the corpus.