Share of Model

Share of model is the proportion of an assistant’s answers, across a fixed set of questions, in which a given business is named. It is a share of a finite space: an answer names two or three businesses, not ten, so one company’s gain is usually another’s loss.

Why it is measured per assistant

Averaging across assistants hides the thing worth knowing. The models are trained differently, retrieve differently and have different habits about how many sources they name, so a business can be a regular recommendation in one and entirely absent from another. A single blended number reports neither situation.

Kept separate, the same measurement says something actionable: absence from one assistant while performing well in the others is usually a retrieval or source problem specific to that assistant, not a reputation problem.

How it differs from share of voice

Share of voice is the older idea: the proportion of a channel’s total conversation or advertising that belongs to one brand. Share of model applies the same shape to a channel where the total is not published and has to be constructed: the denominator is the question panel, which means the number is only as meaningful as the questions chosen.

That dependence is worth stating plainly rather than hiding. A panel weighted towards questions a business already wins will report a high share that means nothing. The panel has to be the questions its customers actually ask, written before the measurement, and then left alone.

One question people ask

Does a rising share of model mean competitors are losing theirs?

Often, but not always. The number of businesses an assistant names in a given answer is small but not fixed, so a share can rise because an answer that previously named two now names three. Reading a rise as a competitor’s loss without checking who else was named is a guess. This is why the recorded answers keep the full list of businesses named beside each one, not just whether the subject appeared.

Sources

The question panel and the recording described here are this site’s own method, asked of one assistant, Claude; the methodology page sets it out. No third-party figures are used on this page.