Share of Model: The AI Share-of-Voice Metric

What is share of model?

Share of model is a brand's share of all brand mentions in AI-generated answers to a fixed set of prompts. It measures competitive presence inside ChatGPT, Gemini, Perplexity, Google AI Overviews and similar engines, the way share of voice measures presence in advertising.

The “model” in the name is the large language model (LLM) behind an AI engine: the system that writes an answer instead of returning a list of links. When a buyer asks such an engine for the best tools, agencies or products in a category, the answer names a handful of brands. Share of model asks what fraction of those names belong to you, compared with the competitors you track.

That makes it a competitive metric rather than a presence metric. A brand can appear in a fifth of all answers and still lose ground if its competitors appear more often. Share of model captures that by putting every tracked brand’s mentions in the same denominator.

The term entered wide use with a product launch. Jellyfish, part of the Brandtech Group, launched its Share of Model™ platform in London on 4 December 2024, built to analyse how ChatGPT, Gemini and Meta’s Llama perceive brands, products and services. Jack Smyth, Jellyfish’s Chief Solutions Officer for AI, Planning and Insights, said at the launch that “LLMs are no longer just tools; they are a critical part of the customer journey.” Jellyfish uses the capitalised name as a trademark for its platform; this page uses the lower-case term for the generic metric that any tracker, or a spreadsheet, can calculate.

How is share of model calculated?

Share of model equals the brand's mentions divided by the mentions of every tracked brand, multiplied by 100. Both counts come from the same answers to the same prompts, and each brand is counted at most once per answer.

Written as one line:

The numerator and denominator must come from one log. If the brand’s mentions come from this week’s run and the competitors’ from last week’s, the ratio describes nothing. The formula is simple; the result depends on two counting decisions made before the first prompt runs. Both are this site’s method, stated here so that any reading can be reproduced.

What counts as one mention

A mention is the brand’s name, or a product name that clearly belongs to it, appearing in the text of an AI answer. This site counts a brand once per answer, however many times the answer repeats the name. One long answer that names a brand five times while describing it in detail would otherwise outweigh five short answers that each recommend it once, and the metric would reward verbosity rather than presence.

Write the accepted name variants down before the first run: the brand name, product names, common abbreviations. Apply the same list every time. A citation — a link to the brand’s site in the answer’s sources — is not a mention unless the name also appears in the text; citations get their own count, covered below as a variant.

Which brands belong in the denominator

The denominator is the sum of mentions of every brand in a competitor set chosen before the first run. That set is fixed: the brand itself plus the 3 to 5 competitors it actually loses deals to. Brands the engines name that are outside the set go into an “other brands” bucket, which is logged but kept out of the denominator.

The bucket matters because AI answers name long tails of brands. If every newly named brand entered the denominator, the share would fall each time an engine mentioned an obscure competitor, and month-to-month readings would stop being comparable. Review the bucket each quarter; if one “other” brand keeps appearing, add it to the set and start a new baseline.

Work through a share of model calculation

A brand named in 48 of 240 answers, where 4 tracked brands collect 210 mentions in total, has a share of model of 22.9%. Its mention rate on the same answers is 20%, because mention rate divides by answers, not by mentions.

The set-up: 30 unbranded prompts, run on 4 engines (ChatGPT, Perplexity, Google AI Overviews and Gemini), twice each. That gives 30 × 4 × 2 = 240 answers. Each answer is read and each tracked brand it names is logged once.

The table shows why the two metrics are reported side by side. Brand A’s mention rate says it appears in one answer in five. Its share of model says that when these engines name a brand from this category, Brand A gets 22.9% of the names, below the 25% parity line for 4 brands and well behind Brand B.

The same 48 mentions split by engine show where the gap sits. Each engine produced 60 answers (30 prompts × 2 runs).

The 4 engine values are not averaged; the page reports the pooled 22.9% with its counts. A plain average of the 4 percentages gives 22.8%, close here only because every engine produced a similar number of mentions. When one engine names brands far more often than another, an average of percentages gives that engine the same weight as a quiet one and misstates the total. Report each engine’s share with its counts, and pool the counts for the overall figure.

Read per engine, the example points to a decision: Brand A is at parity or better on ChatGPT and Perplexity and below it on the two Google surfaces, so the work belongs on the pages and sources those engines draw from.

How does share of model differ from share of voice and mention rate?

Share of model counts mentions inside AI answers; share of voice counts a brand's share of advertising or conversation; mention rate counts the share of answers that name the brand at all. Share of model and mention rate come from the same prompt log.

Share of voice is the older idea. It measures a brand’s share of a category’s advertising or public conversation, and marketers use it as a leading indicator of market share. Share of model carries the same logic into AI answers, where no one buys the placement: the engine decides whom to name. Many AI visibility trackers still label the metric “share of voice”, so check which denominator a report uses before comparing it with anything else.

Mention rate and share of model can move in opposite directions. If engines start naming more brands per answer, a brand’s mention rate can hold steady while its share of model falls, because competitors now share the same answers. Share of model is one of the 5 metrics that make up AI visibility, alongside mention rate, citation rate, answer position and sentiment.

Two variants use the same construction. Citation share divides the brand’s cited URLs by the cited URLs of all tracked brands, and shows whose pages the engines trust as sources rather than whose names they repeat. Per-engine share of model is one value per engine, as in the second table above. Where share of model sits among access, presence and outcome measures is laid out on the map of all AI visibility metrics.

How do you design a prompt set for share of model?

Design the prompt set from unbranded category questions, spread across 4 intent groups, with a competitor set fixed before the first run and identical run conditions every time. Branded prompts inflate the result because the answer names the brand by default.

A prompt is the question typed into an AI engine; a prompt set is the fixed list of prompts that every run uses. Share of model is only as meaningful as that list, because the prompts decide which answers get counted.

Use unbranded category prompts

An unbranded prompt names the category, not a company. Symphonic Digital’s example is “best project management software for remote teams”: it asks what a buyer asks before a shortlist exists. A prompt such as “is Brand A good for remote teams?” guarantees a mention of Brand A and tells you nothing about competitive standing. Symphonic suggests starting by identifying “20-50 relevant queries”. This site’s Experiment 01 uses 50 prompts on 6 engines, run weekly; it is still running and reports no results yet.

Balance 4 intent groups

Competitors win different kinds of questions, so a set weighted to one kind skews the share. Spread prompts across 4 groups:

• Recommendation: “best”, “top” and “which should I choose” prompts, where engines name the most brands.
• Comparison: “A vs B” style prompts written with generic descriptions instead of brand names.
• How-to: task prompts where a brand appears as the tool or service that does the job.
• Definition: “what is” prompts about the category, where engines name brands as examples.

Keep the proportions fixed from period to period. A set that gains recommendation prompts mid-quarter will show a rising share that the brand did not earn.

Fix the competitor set before the first run

Choose the 3 to 5 competitors before anything runs, write them into the tracking sheet and do not edit the list during the reporting period. Adding a strong competitor midway shrinks everyone’s share on paper. If the market changes and a new competitor must be added, start a new baseline from that date and keep the old series separate.

Hold the run conditions constant

Every run uses the same:

• Engines, listed by name and product surface.
• Country and language settings.
• Session state, logged out with no history.
• Runs per prompt, at least 2, so one unusual answer does not decide a week.

The full procedure for building the set and logging each answer is covered by prompt-based tracking.

How often is share of model measured and reported?

Share of model is measured weekly and reported monthly as a 4-week view. Weekly runs capture answer changes after model updates; the monthly view smooths run-to-run noise. A changed prompt set or competitor set starts a new baseline.

This is the cadence used here, matching the weekly schedule of Experiment 01; it is not an industry standard. The monthly figure pools the counts from the last 4 weekly runs and applies the formula once, rather than averaging 4 weekly percentages, for the same reason the engine values are not averaged.

A report shows 3 things per engine: the brand’s share, the parity line for the number of tracked brands, and the change against the previous 4-week view. Put the raw mention counts beside each percentage. A move from 22.9% to 25.0% on 210 mentions is a different claim from the same move on 21.

Re-baselining is the rule that keeps the series honest. When the prompt set gains or loses prompts, when the competitor set changes, or when a run condition such as country changes, the old and new figures measure different things. Mark the date, start a new series and compare only within each series.

Why does share of model change between runs?

Share of model changes between runs because AI engines generate a new answer each time and update their models without notice. The same prompt returns different brand lists on different runs, so single-run readings are not compared with each other.

An AI answer is generated, not retrieved from a fixed list. Ask the same engine the same question twice and the set of brands, their order and the sources behind them can differ. A 2026 survey of 45 generative engine optimization studies found that commercial audits “reveal low source overlap, substantial run-to-run variability, and persistent fidelity gaps.”

Model updates add step changes on top of that noise. When an engine changes its underlying model or its retrieval system, brand lists can shift for every prompt at once, and vendors rarely announce the date. A sudden move on one engine only, across most prompts in the same week, usually has that cause rather than anything the brand did.

Both effects explain the method above: repeated runs per prompt, weekly measurement and a 4-week reporting view. A single reading of share of model is one sample. Four weeks of pooled counts is a measurement.

The calculation runs by hand on a spreadsheet; at scale, software does the counting.

Which tools report share of model?

Several AI visibility trackers report share of model under the name share of voice. Semrush's Brand Performance report shows Share of Voice weekly and Peec AI tracks share of voice daily; the formula matches the one on this page.

Before relying on a tool’s figure, check the 2 counting decisions from this page: whether a brand counts once per answer or once per appearance, and whether the competitor set is yours and fixed. A tool that adds every brand it detects to the denominator produces a share that falls as engines name more brands. The comparison of share-of-model tools lists which trackers separate mentions from citations, how often they refresh and what data they export, so their share figures can be checked against a hand count on a sample of answers.