What is Share of Model?

At a glance

• Share of model is measured on prompts where the assistant answers without searching, so the answer reflects what it absorbed during training.
• The unit is a brand name in generated text, with no link and no source list attached.
• Gains here follow from being written about widely, over a long period, in places that end up in training data.

Asking a model what it knows without letting it search

The measurement depends on getting an answer that does not trigger retrieval. Broad category prompts do this: name some brands that make X, who are the main players in Y, what should I consider when buying Z.

Capture the full answer, list every brand named, and your share is your mentions over all brand mentions in that answer. Run the prompt set, total it up, and you have a distribution across the category.

Two details move the result more than people expect. Order matters, since the first brand named tends to be the default association. And wording matters, because “best” pulls a different set from “affordable” or “for professionals”.

Search-enabled surfaces muddy this. If the assistant runs a live search, you are measuring retrieval, not recall, and the levers are different.

Slow to move, and hard for a rival to take back

What makes share of model appealing is what makes it frustrating: it barely responds to a content sprint.

Default associations come from being mentioned repeatedly across the wider web, in journalism, forums, documentation and other people’s comparisons, over years. You cannot publish your way into that in a quarter, and neither can a competitor with a bigger budget.

That stability is the value. A position in default recall survives a bad month, a redesign and a crawl problem. Retrieval-based visibility survives none of those.

Treat share of model as your long-term brand measure and share of voice in AI search as your quarterly one. Judging a three-month program on share of model makes good work look like failure.

Build a share-of-model baseline from a prompt list

1. Write 15 category prompts with no brand names in them. If your brand appears in the prompt, you are testing the model’s manners, not its recall.
2. Turn search off where the interface allows it, and note the setting for every run so later comparisons are fair.
3. Record every brand named, in order. First position is worth tracking separately from being on the list.
4. Divide your mentions by total brand mentions. That fraction is the number, per assistant, never blended across them.
5. Repeat quarterly, not monthly. The metric does not move fast enough to justify weekly reporting, and frequent checks invite overreaction to noise.

From Parsnipp’s research: a brand’s share of a model’s answers is not its share of the model’s picks. Bose appeared in 82 of 320 answers and was recommended in none, on both GPT-5.6 and Gemini 3.8 Flash. Method: 320 simulated shopper conversations about headphones under $150, on GPT-5.6 and Gemini 3.8 Flash, in two independent batches of 160 (Parsnipp, October 2026). Read the research

Share of model beside share of voice

How this term is used

Share of model is usually reserved for what a model says from its own training, with no web search behind the answer. Share of voice is the broader measure, including answers built from live search.

There’s no standard yet for measuring AI visibility, and vendors name the same metric differently. Whatever the label, these metrics track the organic branch of AI search marketing.