Share of model is the percentage of AI assistant answers or recommendations in a product category that mention or recommend a specific brand — the AI-search counterpart to share of voice in traditional media.
When a buyer asks ChatGPT “what is the best CRM for a mid-size retailer” or asks Perplexity to compare vendors, the model names a handful of brands. Share of model measures how often yours is one of them. It is an early-stage metric — the term and its measurement conventions are still settling — but the behavior it tracks is already real: AI assistants have become a recommendation surface that sits before the website visit, before the analyst report, sometimes before the buyer knows the category’s vendor names at all.
Why Share of Model Is Becoming a Board Metric
Share of voice earned its place in board decks because ad presence predicted market share. Share of model is inheriting that role for a channel no one can buy. An AI assistant’s recommendation is not an ad slot; it emerges from what the model absorbed in training and what it retrieves at answer time. If models in your category consistently recommend three competitors and never you, you are absent from a growing slice of buying journeys — and no media budget fixes it directly.
The metric also behaves differently from paid visibility. It moves slowly (model training cycles are months apart, and retrieval authority accrues gradually), which makes it closer to brand equity than to campaign performance. That persistence is exactly why executives track it: a weak share of model today predicts weak pipeline quarters later.
How Share of Model Is Measured
No standard methodology exists yet — vendor dashboards differ, and numbers are not comparable across tools. The common measurement pattern has four parts:
- Category prompt set. Define the questions a real buyer asks — “best [category] for [segment],” “compare X and Y,” “what should I look for in a [category].” The prompt set is the denominator.
- Sampling across models. Run the prompt set against each major assistant (ChatGPT, Gemini, Claude, Perplexity, AI Overviews). A large language model is stochastic — the same prompt yields different answers on different runs — so each prompt is sampled repeatedly and results are averaged.
- Mention classification. Count whether the brand is mentioned, recommended, or cited as a source — three different strengths of presence. A recommendation (“consider Brand X”) is worth more than a passing mention.
- Trend tracking. Because single snapshots are noisy, the useful signal is share over time and share relative to named competitors.
Share of Model vs. Share of Voice
Share of voice measures presence in paid and earned media — a function of spend and PR, adjustable within a quarter, and covered under brand awareness measurement. Share of model differs on every axis: it cannot be bought, it is generated fresh in each answer rather than placed, and it changes on the timescale of model updates and retrieval authority rather than media flights. The two are complements — share of voice tracks the attention you rent; share of model tracks the recommendation you have earned in the systems buyers now consult first.
How to Move Share of Model
The lever is generative engine optimization: extractable definitional content, structured data, authoritative sourcing, and topical depth that make a brand’s content the material models cite. Entity consistency compounds it — models consolidate a brand described identically across its site and third-party sources, and fragment one that is not. For how this optimization work diverges from traditional search tactics, see GEO vs SEO.
The inward-facing half is data. A customer data platform shows which questions customers actually ask in their own words — the raw material for the category prompt set — and unified first-party data grounds the claims (real outcomes, real numbers) that make content citation-worthy rather than promotional.
Failure modes: why share of model drops
Share of model rarely collapses overnight; it erodes through identifiable mechanisms, each with a different fix. The four below account for most sustained declines teams diagnose.
| Failure mode | What causes it | How to detect it | Fix |
|---|---|---|---|
| Entity fragmentation | The brand is described inconsistently across its own site, review platforms, directories, and press, so models consolidate competing entity representations | Ask several assistants to describe the brand and compare the descriptions against intended positioning | Standardize naming, category, and differentiators everywhere the brand appears; mark them up with structured data |
| Retrieval exclusion | The content models could cite is invisible at answer time — rendered only in the browser, paywalled, or too thin to quote | Check whether any assistant cites the brand’s domain when answering category questions | Publish static, extractable pages that state definitions and comparisons in quotable prose |
| Prompt-set drift | Buyer language shifts into subcategories the tracking set never asks about, so the scorecard reads flat while presence falls | Refresh the prompt set from sales calls and search queries; compare against the original | Rebuild the prompt set quarterly and version it so trend lines stay comparable |
| Citation capture | A competitor publishes the definitional or comparison page models adopt as their default source | Track which sources assistants cite, not just which brands they name | Publish primary data and first-party definitions that give models a better source to cite |
Two properties make these failures sticky. First, absence compounds: a model that never cites a brand has no retrieval signal pointing at it, so the next answer starts from the same deficit. Second, the fix lag is long — content published today influences answers after the next training or index refresh, which is why detection cadence matters more than reaction speed.
Triage order matters because the fixes compete for the same budget. Retrieval exclusion is the first check: if assistants cannot cite the brand’s pages at all, entity work and citation capture have nothing to attach to. Entity fragmentation comes next, because it silently caps the value of every page published afterward — a model unsure which description is authoritative splits its confidence across versions. Prompt-set drift is a measurement problem rather than a presence problem, but an unrefreshed set can make real progress invisible. Citation capture is the slowest to reverse and the most valuable once reversed, since the captured source keeps feeding answers until something better replaces it.
Building a share of model scorecard
A single “share of model” number hides more than it reveals. The scorecard below separates the readings that drive decisions; each row exists because skipping it produces a specific, recurring misread.
| Metric | What it establishes | Why it matters | Failure mode when skipped |
|---|---|---|---|
| Mention share per assistant | Baseline presence in each major model separately | Models differ sharply; an average can mask a zero in the assistant a segment actually uses | Teams celebrate a blended number while absent where their buyers ask |
| Recommendation share vs. mention share | Whether the brand is named or actually shortlisted | A recommendation (“consider X”) carries more purchase weight than a passing mention | Mention-level wins get reported as recommendation-level progress |
| Source citation share | How often the brand’s domain is cited as evidence | Citations are the traffic and trust path from an answer; a name without a citation builds no pipeline | Presence looks healthy while the answer sends buyers nowhere |
| Competitor-relative share | Position against named rivals inside the same prompt set | Absolute percentages mean little without a denominator the reader can judge | Progress is claimed from a rising number that rose for the whole category |
| Share by prompt cluster | Which question clusters the brand wins or loses | Aggregate trend conceals losses in the highest-intent clusters | Budget flows to content for clusters already won |
Treat the scorecard as a panel, not a scoreboard: read the rows together, because the interesting signal is usually a divergence — strong mention share with weak citation share, or a rising aggregate over a falling high-intent cluster. A divergence is also the earliest actionable output: it points at a specific mechanism from the failure modes above rather than at visibility in general.
Where share of model meets agentic buying
The click-through collapse is what makes this metric structural rather than cosmetic. Pew Research Center’s 2025 analysis of Google browsing behavior found users clicked a traditional result in 8% of visits where an AI Overview appeared, versus 15% without one — and only 1% clicked a link inside the AI summary itself (Pew Research Center, 2025). The judgment happens inside the answer, on the exact surface share of model measures.
The surface is also gaining agency. As agentic AI moves from answering questions to completing tasks, and agentic commerce lets software research, compare, and transact on a buyer’s behalf, the shortlist an assistant produces stops being a top-of-funnel impression and becomes a pipeline input. Gartner predicted in 2024 that unofficial third-party tools powered by generative AI — tools consumers adopt themselves, not ones brands deploy — would resolve 40% of customer service issues by 2027 (Gartner, 2024); the same migration of tasks to brand-external agents is underway in buying research.
Paid presence inside AI answers — the territory of agentic advertising — does not change the scorecard. An ad slot buys ad presence; recommendation share is earned from what the model draws on. Track the two separately, because a paid placement inside answers an agent never reaches, or reaches without a recommendation, reads as visibility it does not deliver.
What market concentration means for share of model
Share of model is structurally skewed toward incumbents, and the skew has a measurable source. The CDP Institute counted 208 tracked CDP vendors in January 2026 and found the top vendors hold 67% of industry employment and 73% of funding (CDP Institute, 2026). Scale of that kind converts directly into model presence: more review coverage, more comparison mentions, more third-party pages that describe the vendor in the terms models consolidate. A model asked for category recommendations defaults toward the brands its training and retrieval material describes most consistently.
For challengers, the implication is to narrow the battlefield. Winning the category head term outright against vendors with that footprint is slow; winning “best [category] for [a specific segment or use case]” is achievable, because long-tail prompt clusters have thinner incumbent coverage and fewer consolidated third-party descriptions. The practical sequence: define the clusters where the brand has a genuine right to win, build the extractable, sourced content those clusters draw on, and measure share per cluster rather than against the head term — where the concentration effect is strongest.
FAQ
How do you measure share of model?
Sample a category prompt set across the major AI assistants and count brand mentions. Define the questions buyers ask in your category, run them repeatedly against ChatGPT, Gemini, Claude, and Perplexity to average out run-to-run variation, and classify each result as a mention, recommendation, or citation. Track the percentage over time and against competitors; single snapshots are too noisy to act on.
What is the difference between share of model and share of voice?
Share of voice is rented; share of model is earned. Share of voice measures presence in paid and earned media and responds to spend within a quarter. Share of model measures presence in AI-generated answers, cannot be bought, and moves on the slower timescale of model training cycles and retrieval authority. Treat them as complementary reads on visibility, not substitutes.
What is a good share of model?
There is no established benchmark — the useful reading is relative and directional. The metric is too new and methodologies too varied for absolute targets. In practice, teams compare their share against the top competitors in their own prompt set, check consistency across models (strong in one assistant, absent in another signals fragile authority), and treat sustained upward trend as the success criterion.
How long does it take to move share of model?
Weeks to months, not days — the metric moves on the timescale of model updates and retrieval authority. A single answer is stochastic, so no one run proves anything; what moves the number is sustained citation-worthy content, entity consistency across third-party sources, and the next model training cycles absorbing them. Teams that see movement fastest usually win a narrow prompt cluster first, then expand. Treat a change in one sampling round as noise until it persists across several.
Can you buy share of model?
Not directly — ad spend buys ad slots, not the model’s recommendation. Paid placements inside AI answers exist, but they are advertising inventory tracked separately from organic recommendation share. What raises share of model is the material models draw on: extractable definitions, structured data, cited sources, and consistent entity descriptions across the web. Budget still helps — it funds the content and data work — but the purchase is indirect, and conflating the two misreads the scorecard.
Related Terms
- Answer Engine Optimization — Winning the direct-answer surfaces where share of model is contested
- LLM SEO — The technical mechanics that determine whether models can find and cite your content
- AI Search Optimization — The umbrella visibility program that share of model scorecards
- AI Search Visibility — The measured state of how present a brand is across AI answers, which share of model quantifies at the brand level
- Marketing Analytics — The measurement discipline this metric extends into AI channels