For twenty years, brand teams reported Share of Voice: your slice of ads, press and social conversation. In 2026 a quieter number started deciding revenue β how often an AI assistant names your brand when a buyer asks what to purchase. Researchers at INSEAD gave it a name in mid-2025: Share of Model. Twelve months later it is on CMO dashboards, in analyst decks, and at the center of how we build measurement at GEOly AI.
This is a working guide to the metric: what it is, the benchmark numbers worth knowing, why it fragments wildly across engines, how to calculate it credibly, and how to move it.
Key takeaways
- Share of Model (SoM) measures how often, how prominently and how favorably a brand appears in AI-generated answers β the AI-era successor to Share of Voice.
- Benchmarks: across categories, the average brand is mentioned in roughly 16% of relevant AI answers; category leaders reach around 56%. The gap between average and leader is the battleground.
- SoM fragments hard by engine. In the INSEAD research, one detergent brand held ~24% SoM on Meta's Llama but under 1% on Gemini. "Visible in AI" is meaningless β visibility exists per model.
- A credible SoM number needs a fixed prompt set, multi-engine sampling, repeated runs, and scoring for mention, citation, position and sentiment β not three ad-hoc ChatGPT queries.
- You move SoM by feeding the sources engines retrieve and learn from: reviews, comparisons, communities, structured product data β the same authority engine behind GEO.
What exactly is Share of Model?
The definition from the INSEAD work: the share of AI-generated responses in which a brand appears, weighted by prominence and favorability, within a defined category of prompts. A SoM of 35% on "best wireless earbuds under $100"-type prompts means: in 100 sampled answers to that buying question across your target engines, your brand showed up 35 times.
Three components matter beyond the raw count:
Component | Question it answers | Why it matters |
|---|---|---|
Mention rate | Do you appear at all? | Absence = invisible at the decision moment |
Prominence | First recommendation or footnote? | Assistants front-load; position 1-2 captures the intent |
Favorability | Recommended, neutral, or warned against? | A negative mention is worse than none |
Share of Voice counted what you said about yourself plus earned echo. Share of Model counts what the gatekeeper says when asked for advice β a judgment synthesized from everything the model retrieved and learned. That makes it less buyable and more defensible: you cannot purchase position one in a paragraph, you have to earn the sources behind it.
The benchmark numbers
Published 2026 research puts the average brand mention rate in relevant AI answers at about 16.3%, with top brands in a category reaching 56.5%. Two practical readings:
- If you measure and land near 16%, you are average β meaning in five of six buying conversations, the assistant recommends someone else.
- The 40-point spread between average and leader is the growth channel. Nothing in paid media currently offers that kind of unclaimed share.
And the fragmentation finding is the one most teams miss: brand visibility is engine-specific. The INSEAD example β ~24% on Llama, under 1% on Gemini for the same brand β comes from different training corpora, different retrieval layers and different source preferences per engine. We see the same split daily across ChatGPT, Gemini and Perplexity tracking. Any SoM report that doesn't break out by engine is hiding the story.
How to measure it credibly

The method we use at GEOly AI, condensed:
- Build the prompt set first. 100+ prompts that mirror real buying questions in your category β best-of, vs, "is X worth it", use-case and constraint variants. This is the AI-era keyword list; our grounding queries guide explains how engines expand them internally.
- Sample across engines, repeatedly. Same set against ChatGPT, Gemini, Perplexity, Google AI Mode. Answers are stochastic β one run is an anecdote, weekly runs are a trendline.
- Score every answer: mentioned or not, cited or not, position, sentiment. Store the raw answers; you will want receipts.
- Compute SoM per engine and per prompt theme. The aggregate hides exactly the fragmentation you need to act on.
- Diagnose the losses. For prompts where you are absent, record who is present and which sources the engine cited β that citation list is next quarter's outreach and content plan.
A fixed prompt set plus repeated multi-engine sampling is also what makes the number auditable β the property that lets SoM sit next to rankings and ROAS in a board deck without an asterisk.
How to move the number
SoM responds to the source layer, not to on-page tweaks alone. The levers, in rough order of observed impact: presence and ratings on the review and comparison sites engines cite for your category; consistent entity data (name, category, claims) across the open web; original data and definitive content worth citing; community proof (Reddit, forums, YouTube) where models look for "real user" signal; and complete structured product data for shopping-type prompts β the same feed layer that now decides agentic checkout eligibility.
Expect movement in weeks-to-months, not days: retrieval-driven engines (Perplexity, AI Mode) shift first, training-data-driven visibility follows slower. That lag is an advantage for early movers β share you earn now is annoying to dislodge.
FAQ
How is Share of Model different from Share of Voice?
Share of Voice measures presence in media you or your echo chamber produce. Share of Model measures presence in the synthesized judgment of an AI gatekeeper. SoV is an input; SoM is closer to an outcome β it reflects whether the sum of your reputation actually surfaces at the decision moment.
How many prompts do I need for a reliable number?
Enough to cover your category's buying intents with repetition β in practice 100+ prompts sampled weekly per engine. Below that, variance swamps signal and month-over-month deltas are noise.
My brand scores high on one engine and near zero on another. Is that normal?
Completely β it is the norm, not the exception. Engines differ in training corpora and retrieval sources. Treat each engine as its own market with its own competitive set, and prioritize by where your buyers actually ask.
Can I calculate Share of Model manually?
A pilot, yes: 20 prompts Γ 3 engines in a spreadsheet will teach you the shape of the problem in an afternoon. A defensible, trended metric needs automation β GEOly AI runs the full loop (fixed prompt sets, multi-engine sampling, mention/citation/sentiment scoring, per-engine SoM and the losing-prompt diagnosis) continuously.
Want your baseline? Run a free Share of Model check on your brand β it takes minutes, and the per-engine split is usually the first surprise. Details on plans: pricing.

