What is the Share of Model formula?

Share of Model = (answers naming your brand ÷ total answers sampled) × 100. Total answers sampled equals prompts × models × runs per prompt. Hold all three constant between measurement cycles or the numbers are not comparable.

How is Share of Model different from share of voice?

Share of voice counts impressions or mentions across media you can largely buy into. Share of Model counts inclusion in answers a model generates on the buyer's behalf, which you can only influence through the sources the model trusts. Share of voice measures exposure; Share of Model measures recommendation.

What is a good Share of Model?

There is no universal benchmark, because it depends on how many credible vendors exist in your category. Judge it relatively: your share versus named competitors on the same prompt set, and your own trend over time. In a shortlist of five plausible vendors, an even split is roughly 20%.

How many runs are needed for a reliable Share of Model?

More than one, always. Model output is stochastic — the same prompt can return different brand sets minutes apart. Sample each prompt several times per cycle and average. Single-run checks produce swings that look like performance changes but are variance.

Should Share of Model be tracked per model?

Yes. ChatGPT, Claude, Gemini, and Perplexity return materially different brand sets for the same question, and they update on different schedules. A blended figure can stay flat while your position collapses on the one model your buyers actually use.

Does a mention always count?

Count it, but classify it. A recommendation, a neutral listing, and a caveated mention are all 'named' and mean very different things commercially. Track favorability alongside frequency so a rise in mentions is not mistaken for a rise in preference.

Metric

What is Share of Model?

The measurement layer of LLM optimization: how frequently AI models recommend, mention, or position your brand across the questions your buyers actually ask.

Updated: August 2026

The formula

Three parameters must be declared and then frozen:

Prompt set — the buyer questions, in buyer language, tagged by stage.
Model set — which models, at which settings, with retrieval on or off.
Window — the sampling period, e.g. one week.

A worked example

A mid-market vendor tracks 40 buyer questions across 4 models, sampling each question 3 times a week. That is 40 × 4 × 3 = 480 answers per cycle.

Brand	Answers naming it	Share of Model
Competitor A	178 of 480	37.1%
Competitor B	141 of 480	29.4%
This vendor	62 of 480	12.9%
Everyone else	99 of 480	20.6%

Illustrative example using round numbers, not observed data.

Now split the same 480 answers by stage and the diagnosis changes. If the vendor's 62 mentions are concentrated in discovery questions and near zero in comparison questions, the problem is not awareness — it is that no source the model trusts compares them to Competitor A.

How to break it out

By buyer stage — discovery, evaluation, decision, retention. The single most useful cut.
By model — coverage differs sharply between ChatGPT, Claude, Gemini, and Perplexity.
By competitor — absolute share means little without the set you are being compared against.
By favorability — recommended vs listed vs caveated.

8,644
Buyer questions observed: 24,965
Cited sources captured: 993
Distinct domains cited: 4
Models sampled

Buyer stage	Questions	Cited sources	Share of observed answers
Evaluation	4,087	10,986	44.0%
Discovery	2,521	7,541	30.2%
Decision	2,034	6,421	25.7%
Retention	2	17	0.1%

How observed AI citations distribute across buyer stages in Monroya's scan data. Evaluation questions produce the most cited sources — and are where most B2B brands lose Share of Model.

Common measurement mistakes

Running each prompt once and treating the result as a reading.
Changing the prompt set mid-quarter, then comparing to last quarter.
Reporting one blended number across four models with different behavior.
Counting caveated mentions as wins.
Using keyword-shaped prompts no buyer would ever type.

Next: how to improve Share of Model →

Frequently asked questions

What is Share of Model?: Share of Model measures how frequently your brand is recommended, mentioned, or positioned by AI models across the questions your buyers ask. It is expressed as the percentage of sampled answers that name your brand for a fixed prompt set, model set, and time window.
What is the Share of Model formula?: Share of Model = (answers naming your brand ÷ total answers sampled) × 100. Total answers sampled equals prompts × models × runs per prompt. Hold all three constant between measurement cycles or the numbers are not comparable.
How is Share of Model different from share of voice?: Share of voice counts impressions or mentions across media you can largely buy into. Share of Model counts inclusion in answers a model generates on the buyer's behalf, which you can only influence through the sources the model trusts. Share of voice measures exposure; Share of Model measures recommendation.
What is a good Share of Model?: There is no universal benchmark, because it depends on how many credible vendors exist in your category. Judge it relatively: your share versus named competitors on the same prompt set, and your own trend over time. In a shortlist of five plausible vendors, an even split is roughly 20%.
How many runs are needed for a reliable Share of Model?: More than one, always. Model output is stochastic — the same prompt can return different brand sets minutes apart. Sample each prompt several times per cycle and average. Single-run checks produce swings that look like performance changes but are variance.
Should Share of Model be tracked per model?: Yes. ChatGPT, Claude, Gemini, and Perplexity return materially different brand sets for the same question, and they update on different schedules. A blended figure can stay flat while your position collapses on the one model your buyers actually use.
Does a mention always count?: Count it, but classify it. A recommendation, a neutral listing, and a caveated mention are all 'named' and mean very different things commercially. Track favorability alongside frequency so a rise in mentions is not mistaken for a rise in preference.