Share of Model: definition, calculation and reliability

Share of Model measures how often a brand appears in answers generated by answer engines such as ChatGPT, Google AI Mode, Google AI Overview, Perplexity or Gemini, across a defined set of questions. The definition is broadly agreed on. The calculation method, far less so.

Between two tools applying the same definition, the gap on a single brand can reach tens of points. The cause lies not in the measurement itself but in four decisions that are rarely made explicit: what sits in the denominator, what counts as a mention, what is excluded from the calculation, and how many readings support the figure on display.

This article sets out those decisions, then tests them against a corpus of 8,261 engine answers analysed across ten sectors of the French-speaking Swiss market between April and August 2026. Three findings emerge, two of which run counter to the prevailing view that generative answers are unstable.

Where the term comes from

Share of Model extends a well-established marketing lineage. The work of Les Binet and Peter Field on advertising effectiveness established the importance of excess share of voice (ESOV), the gap between a brand's share of voice and its market share. Then came share of search, developed by Les Binet and James Hankins around 2020, which measures the volume of queries associated with a brand.

Share of Model applies the same logic to generative engines. The term was introduced by Jack Smyth, at Jellyfish, in 2024, then popularised by Tom Roach in Marketing Week as the natural successor to share of search in the era of large language models. In its original sense, the notion also covers the associations a model attaches to a brand, not only the frequency of mentions. In practice it breaks down into measurable indicators.

One difference in nature deserves note. Advertising share of voice can be bought. Share of search can be worked on indirectly. Share of model cannot be bought: no budget places a brand inside a generated answer. It is built by what third-party sources say about the brand, which brings it closer to brand equity than to campaign performance.

This work of optimising for AI answers has a name, GEO (Generative Engine Optimization). It extends organic search optimisation, which targeted page rankings. SEO and GEO complement each other: a brand picked up by reference sources is more likely to appear in answer engines.

Three shares, three terrains

The three metrics resemble each other in form and differ in what they observe.

Share of search measures demand that has already formed: the user knows the brand and looks for it. Share of model measures an earlier stage, where the buyer does not yet know the options and asks the engine to present them. The two answer different questions and are read together.

What the metric covers

Mention and citation

One preliminary distinction governs everything else. A mention is the brand named in the text of the answer. A citation is a site retained as a source, with a link.

The two follow different logics. A mention builds the association in the buyer's mind and depends on what the model has absorbed about the brand. A citation belongs to source selection at answer time, and can generate traffic. A brand may be mentioned abundantly without its site ever being cited, and the reverse is also observed.

Share of Model concerns mentions. Conflating the two amounts to measuring two distinct phenomena under a single figure.

The four indicators

Presence in an answer is not a binary value. Four indicators describe it, two for presence, two for the quality of that presence.

Visibility, or appearance rate, is the share of measured answers in which the brand appears at least once. Share of voice divides brand mentions by the total mentions of its competitive set. Sentiment qualifies the tone attached to the brand when it is named. Position records the order of appearance within the generated answer.

These indicators are not interchangeable. A brand present in 40 % of answers but always at the end of a list, and a brand present in 20 % of answers but consistently first, describe opposite situations that a single indicator would conflate.

Brand visibility in AI answers

#	Brand	Visibility	SOV	Sentiment	Position
1	Competitor A	52%	28%	75	#1.6
2	Competitor B	44%	22%	61	#2.1
3	Your brandYou	36%	15%	72	#2.4
4	Competitor C	30%	18%	68	#2.8
5	Competitor D	22%	12%	54	#3.3

These four indicators extend and sharpen the three established answer engine KPIs: share of voice and position make measurable what "being recommended" conveyed intuitively.

How Share of Model is calculated

The formula

Share of voice is calculated as follows:

share of voice = brand mentions
                 ────────────────────────────────────  × 100
                 brand mentions + competitor mentions

The formula is trivial. The difficulty lies entirely in defining the denominator.

Choosing the denominator

Two conventions coexist, and they do not produce the same figure.

The first retains tracked competitors, meaning a list fixed in advance. The denominator is stable, the company controls it, and movements in the figure reflect real shifts rather than changes of scope. Adding a competitor mechanically dilutes share of voice, removing one reconcentrates it. This is the convention Repliq applies, and the one Peec AI also documents.

The second retains every brand detected in the answers. The denominator becomes a market share among discovered brands. The metric then describes the market as the model renders it, which makes sense for a first sector snapshot, when no competitor list has yet been fixed. But the denominator moves with every reading, according to what the models name. A brand can see its share fall without having lost anything, simply because the engines named more players that week.

Both conventions are defensible. They are not comparable with each other. A Share of Model reported without its denominator cannot be interpreted, and two tools showing 15 % and 32 % for the same brand may both be right.

What must be excluded

Three categories of readings distort the result if they enter the denominator.

Failed measurements. When an engine does not respond, the brand is obviously not mentioned. If that reading is counted, it weighs on the denominator and lowers the visibility on display even though nothing was measured. Only completed measurements enter the calculation.

Engines withdrawn from tracking. An engine measured in the past then abandoned leaves a history that, if it stays in the base, makes the figure diverge from one surface to another.

Out-of-scope questions. Questions designed to measure sentiment call the brand by name. Including them in a visibility rate amounts to observing that the brand appears when it has been named in the question.

Deduplication

The same prompt asked several times on the same day on the same engine produces several answers. Counting them all over-weights the most frequently sampled questions and makes the figure depend on the technical schedule rather than on the market. The convention retained is one reading per question, per engine, per day.

What 8,261 answers show

The preceding sections describe a method. What remains is what it produces on real data.

The corpus comes from the Swiss Atlas, the longitudinal observatory Repliq devotes to brand visibility in answer engines. Ten sectors of the French-speaking Swiss market are tracked: banking, health insurance, higher education, grocery retail, digital services, transport and mobility, residential real estate, travel agencies, luxury watchmaking, restaurants. Four engines are queried, ChatGPT, Google AI Mode, Google AI Overview and Perplexity, across 12 readings between weeks 17 and 31 of 2026.

Of 8,832 answers collected, 8,261 were exploited, or 93.5 %. Named brands were re-extracted from the full text of the answers, independently of their formatting.

Scope and limits of interpretation(5)

1. French-speaking Swiss market only. Questions are asked in French. The German-speaking and Italian-speaking regions are not covered. These results do not hold for Switzerland as a whole.
2. Market share among discovered brands. The observatory does not track a fixed competitor list: the denominator used here is the set of brands named by the engines, the second convention described above. The percentages are therefore not comparable with a share of voice computed on tracked competitors.
3. Week 17 as starting point. The preceding week corresponds to a technical ramp-up and is excluded from all calculations.
4. Extraction dated 6 August 2026. The corpus grows with each measurement cycle. A later extraction shifts values by around a tenth of a point. Gaps below 0.5 point between two close brands should not be read as a ranking.
5. Single pass. The figures are exploratory and require a second verification before contractual use.

The metric is stable from week to week

This is the most counter-intuitive observation. Generative answers are held to be unstable, on the grounds that the same question asked again rarely produces identical text. That variability is real at the level of a single answer. It almost entirely disappears at the aggregate level.

Across the ten sectors, the leading brand's share varies by a standard deviation of between 0.3 and 1.2 points from one week to the next. Total amplitude across 12 readings stays between 1.1 and 3.4 points.

Migros (grocery retail)EPFL (higher education)SBB CFF (transport and mobility)UBS (banking)

Swiss Atlas, Repliq observatory. French-speaking Swiss market, 8,261 answers from ChatGPT, Google AI Mode, Google AI Overview and Perplexity, weeks 17 to 31 of 2026.

The practical consequence is twofold. A movement of 1 point from one week to the next is noise and calls for no interpretation. A movement of 3 points sustained across several readings is signal. This stability rests on the size of the question set and the number of engines queried: narrow scopes measured on a single engine are what produce erratic curves.

Engines agree more than is assumed

A second observation runs against the grain. Each engine is held to have its own selection regime, which the data confirms at the level of cited sources. At the level of named brands, the top of the ranking converges markedly.

In 6 sectors out of 10, the leading brand is the same across all four engines. Helsana in health insurance, EPFL in education, Migros in grocery retail, Immoscout24 in real estate, Swisscom in digital services and SBB CFF in mobility lead everywhere.

Three sectors show a clear divergence, and it comes from the same surface each time.

Google AI Overview departs from the other three surfaces in two of these three cases, banking being the exception with Google AI Mode. The tenth sector, restaurants, also diverges, but its leading brand captures only 1.7 % of mentions: at that level the order of the first few varies between readings and does not support a stable ranking.

Tracking a single engine therefore remains insufficient, but the gap plays out more in the depth of the ranking than at its top.

No brand dominates its category

A third observation. The prevailing narrative presents answer engines as a machine for concentrating attention on a handful of brands. The data describes the opposite.

The leading brand captures between 1.7 % and 9.4 % of its sector's mentions. Even in grocery retail, where Migros leads clearly, nine mentions out of ten go to other players. The number of distinct brands named per sector ranges from 390 to 1,191.

Swiss Atlas, Repliq observatory. 8,261 answers from ChatGPT, Google AI Mode, Google AI Overview and Perplexity, weeks 17 to 31 of 2026.

Cumulative concentration confirms the reading. Health insurance is the exception, with 52.8 % of mentions gathered by the top ten brands. In the nine other sectors, that cumulative figure sits between 12.5 % in restaurants and 37.5 % in grocery retail. In other words, even adding up the ten most cited players, the majority of mentions goes elsewhere.

One operational lesson follows. A brand targeting 20 % or 30 % share of model is setting a goal its sector leader does not reach. Realistic orders of magnitude sit below 10 %, and progress is measured in points, not multiples.

One final observation deserves attention: comparison sites appear in the brand rankings, not only among the sources. Comparis ranks 5th in health insurance (5.5 %) and 3rd in real estate (5.8 %). The engines name them as players in the offering, which moves part of the available share outside the perimeter of the sector's brands.

Only one answer in three ranks brands

A last observation, methodological in nature. Across the same corpus, 95.2 % of answers name at least one brand, but only 33.3 % present them in a structured form, numbered list, table or bold text, allowing them to be ordered automatically.

The remaining two thirds name brands in prose, within the flow of the text. A measurement that reads only structured formats therefore covers a third of the available material, with a predictable bias: questions calling for a ranking are over-represented, open questions disappear. This is a frequent and rarely documented cause of divergence between tools.

What the metric does not say

Three limits bound what a Share of Model allows one to conclude.

It does not measure purchase intent. A brand named in an answer has been retained by the model, which presumes neither buyer consideration nor conversion. The link between share of model and commercial outcome remains to be established market by market.

It says nothing about phrasing. A mention may place the brand as one option among others, as the sector reference, or as a foil. Qualifying that tone is the role of sentiment, and share of voice alone does not substitute for it.

It depends on the question set retained. A scope skewed towards questions where the brand is strong produces a flattering figure of no practical use. The representativeness of the question set determines the validity of the measurement more than the sophistication of the calculation does.

Five common measurement errors

Deriving position from rank in an extracted list. The order in which an extractor returns the brands of an answer is not reliable. Position must be read from the text, by locating the real order of appearance.

Confusing average position with ranking. Average position indicates where the brand appears within answers. Rank indicates where it sits against the others. Both are useful, neither replaces the other.

Averaging engines without saying so. A global figure aggregating four engines remains readable, provided it is presented as an average and accompanied by the breakdown per surface.

Comparing two tools. Question scopes, denominators, frequencies and exclusion rules differ. A single tool tracked over time is more informative than several one-off readings compared against each other.

Reacting to a one-point movement. The data above places weekly noise at around 1 point. Below that, a gap carries no information.

Building your own measurement

A few principles are enough to start on solid ground.

Build a stable question set representative of the category, from questions buyers actually ask. A set that changes with every reading rules out any reading of trend.

Fix the competitor list before the first reading, and change it only knowingly, recording the date of the change.

Separate engines and languages. In a multilingual market, the same question asked in French, German or Italian does not return the same brands.

Measure over time and reason in terms of trend. An isolated reading gives a point, not a level.

Distinguish mention from citation, and track the four indicators separately.

Finally, favour data read from the engines' real interfaces over estimates, and keep a record of the measured answers: without them, an observed gap cannot be explained.

Measure my AI visibility→

An AI visibility audit establishes that starting point; regular tracking reads the trend.

FAQ

What is Share of Model? How often a brand appears in answers generated by answer engines, across a defined set of questions. The term was introduced by Jack Smyth, at Jellyfish, in 2024.

How is Share of Model calculated? By dividing brand mentions by the total mentions of its competitive set. The decisive point is the denominator: a competitor list fixed in advance, or every brand detected in the answers. Both conventions are defensible and are not comparable with each other.

Why do two tools report different figures? Because they are not measuring the same thing: question scope, denominator, frequency, deduplication and exclusion rules all differ.

How many readings are needed for a reliable figure? Across 8,261 measured answers, the leading brand's share varies by a standard deviation of 0.3 to 1.2 points from one week to the next. The metric is stable provided the question set is broad and asked across several engines.

How does it differ from classic share of voice? Share of voice measures advertising or search presence. Share of Model measures presence in AI answers. Same logic of relative share, different terrain, and a space that cannot be bought.

---

Read next:

• What is GEO?
• The 3 essential GEO KPIs
• Who answers for your brand in AI?