- "Share of model" means the percentage of AI-generated answers in a category that name a specific brand, run across a fixed set of prompts and one or more language models.
- The term is vendor-coined, not an academic or standards-body metric. Jellyfish, a marketing agency, is the source most consistently credited, with a product launch dated December 4, 2024.
- There is no standardized prompt set, model selection, or counting method across vendors. A "share of model" number from one tool isn't directly comparable to the same metric from another.
- This is the same measurement already covered on this site as "consideration-set tracking": running a fixed prompt set across models and tracking how often a brand shows up. "Share of model" is a newer marketing label on an existing practice, not a new technique.
- SparkToro's research already shows AI brand lists are highly inconsistent between runs of the identical prompt. That instability applies directly to any "share of model" number: it's a snapshot of one prompt set on one day, not a stable baseline, unless it's tracked as a trend across many runs.
"Share of model" is one of the newer pieces of vocabulary showing up in AI-visibility pitches, usually presented as if it were an established industry metric with agreed-upon benchmarks. It isn't. Here's what the term actually refers to, where it came from, and why the specific percentages attached to it in vendor marketing deserve more scrutiny than they usually get.
What does "share of model" actually measure?
The basic idea: run a set of prompts that a buyer in your category might plausibly type into ChatGPT, Gemini, or another model, then calculate what percentage of the resulting answers mention your brand by name, relative to the total number of brand mentions across all answers in that same set. It's a direct borrowing of the older advertising concept "share of voice" (what percentage of category media coverage or ad spend belongs to you), with "model" substituted for "voice" and AI-generated answers substituted for media mentions.
Mechanically, it's the same thing this site already covers under consideration-set tracking: a fixed, repeatable prompt set, run across one or more models, counted for brand mentions. "Share of model" is a newer label vendors have put on that existing measurement practice, framed as a standalone KPI rather than one component of a broader monitoring approach.
Where did the term come from?
The attribution that shows up most consistently in vendor and trade coverage credits Jellyfish, a marketing agency, with coining the term, and names Jack Smyth, the company's Chief Solutions Officer, as the person behind it. Jellyfish launched a product called "Share of Model" on December 4, 2024, with Danone and Chivas Brothers named publicly as beta participants.
That attribution is worth treating with the caution it deserves: it comes from vendor blogs and trade press covering a company's own product launch, not from an independent or academic source verifying who used the phrase first. It's plausible and consistently repeated, but it isn't the kind of claim with a primary-source paper trail behind it, the way, for instance, the term GEO traces cleanly to a specific 2023 arXiv paper. Treat "coined by Jellyfish" as the best-supported account available, not a verified fact with a citable original source.
Is there an agreed-upon way to calculate it?
No, and this is the part that matters most if someone quotes you a number. "Share of model" as a concept is simple; the implementation choices behind any specific reported figure are not standardized at all:
| Variable | Why it changes the number |
|---|---|
| Prompt set | A prompt set skewed toward your strongest category terms will show a higher share than one skewed toward adjacent or competitor-leaning terms. |
| Model selection | Mention rates vary enormously by platform. Muck Rack's Generative Pulse research already shows citation behavior differing sharply between ChatGPT, Claude, and Perplexity on the same underlying content; the same holds for raw mention rates. |
| Counting method | Does a mention of a product line count as a mention of the parent brand? Does an indirect description without the brand name count at all? Vendors differ. |
| Run count | A single prompt run captures one instance of a probabilistic, non-deterministic process. SparkToro's research found under 1-in-100 odds that two runs of the identical prompt return the identical brand list. |
Two vendors measuring the same brand, in the same category, at the same time, can report meaningfully different "share of model" numbers, because none of the four variables above is fixed across tools. That's not a flaw unique to this metric; it's the same problem the AI visibility score already has, under a different name.
Are the published benchmarks for "good" share of model reliable?
Vendor write-ups on the term cite different typical ranges for what counts as a healthy share of model, and none of the ones found in researching this piece disclosed a large, independently verifiable sample with a transparent methodology behind the number. That's the same pattern the fake-AI-stat piece on this site keeps finding elsewhere: a number gets quoted confidently in marketing copy, and tracing it back lands on another vendor's blog post rather than a disclosed study.
That doesn't mean the underlying measurement is useless. It means a single snapshot number, compared against someone else's published "typical range," tells you very little. What's actually useful is tracking your own share of model, on your own fixed prompt set, across the same models, over time, the same discipline the five-stage framework already recommends for Stage 5 monitoring generally. The trend is the signal. The absolute number, especially compared against a vendor's unverified "benchmark," isn't.
What should you actually do with the term?
Treat "share of model" as a repackaged name for a measurement practice this site already covers in more operational detail: building a fixed prompt set that reflects real buyer language, running it across enough prompts and enough runs to get a trustworthy read, and tracking it over weeks rather than treating one run as a scorecard. If a vendor pitches a "share of model" dashboard, the right questions are the same ones worth asking about any AI-visibility tool: what's the prompt set, which models, how is a mention counted, and how many runs back up the number being shown.
Frequently asked questions
What does "share of model" mean?
The percentage of AI-generated answers in a given category or query set that mention a specific brand, measured across one or more language models. It's modeled on the older marketing concept of "share of voice," applied to AI answers instead of media mentions.
Who coined the term?
Jellyfish, a marketing agency, is the most consistently credited source; its Chief Solutions Officer Jack Smyth is named in vendor coverage of the term, and the company launched a "Share of Model" product on December 4, 2024, with Danone and Chivas Brothers named as beta participants. That attribution comes from vendor and trade coverage, not an independent or academic source.
Is there a standard way to measure it?
No. Every vendor citing a "share of model" number ran it on their own fixed prompt set, their own choice of models, and their own counting method (exact brand mention vs. a looser match). Two vendors measuring the same brand in the same category can and do report different numbers, because the inputs aren't the same.
Is a 10-30% share of model good or bad?
There's no answer to that independent of which vendor's benchmark you're using. Different vendor write-ups report different typical ranges, and none of them discloses a large enough, methodologically transparent sample to treat as an industry baseline. Treat any specific number as meaningful only relative to your own prompt set, tracked over time, not against someone else's published average.
Sources referenced
- Conductor: Share of Model
- Jellyfish: Share of Model platform launch coverage
- SparkToro / Gumshoe.ai AI recommendation consistency study (Jan 2026, 2,961 prompt runs)
For the fuller measurement methodology behind any "share of model" number, see consideration-set tracking in AI search, how many prompts do you actually need for a trustworthy read, and building a fixed prompt set that reflects real buyer language.
