What Share of Model is and how to put it in your reporting
A new metric is showing up in the marketing reports of 2026: Share of Model. The term is starting to circulate in the industry — in articles, in agency proposals, in the odd client conversation that reaches them before you do — and, as happens with any young metric, it circulates with different definitions depending on who's using it. If you're going to put it in your reporting, you'd do well to define it properly, calculate it with a method, and present it with its limitations up front. That's what this article is.
The definition, no smoke
Share of Model: the percentage of AI answers about your category in which your brand appears.
Taken apart piece by piece:
- "AI answers": what ChatGPT, Gemini, Perplexity or Claude reply when someone asks them.
- "About your category": not just any question, but the ones relevant to your business — "which accountant do you recommend in Seville?", "what's the best invoicing software for freelancers?".
- "Your brand appears": the answer mentions or recommends your client.
If you ask 100 things about the category and the brand comes up in 23 answers, the Share of Model is 23%.
The analogy your client will get instantly: it's the share of voice of old — what slice of your category's conversation you occupy — applied to a new channel: the answers from AI models. Where classic share of voice measured presence in media or in search results, Share of Model measures presence in what the AIs reply (the AI version of share of voice, explained in the glossary). The name comes from there: share of the model instead of share of voice.
An honest note on the term: it's starting to be used in the industry and still has no standardised definition nor any body to set it. Different tools and agencies calculate it with different nuances. That doesn't invalidate it — classic share of voice wasn't born standardised either — but it forces a rule: in your reporting, always define how you calculate it. A metric without a methodology beside it is marketing, not measurement.
How it's calculated: sampling, not a census
There's no register of "all the answers the AIs give about your category" — each answer is generated on the spot and nobody sees them all. So Share of Model isn't looked up: it's sampled. The method has four steps:
1. Define the prompt set. A fixed list of questions that represent your category: 30-75 prompts mixing recommendation ("which X do you recommend in Y?"), comparison ("X or Z?") and problem ("I need to solve such-and-such, who can help me?"). It's the most important decision in the whole calculation — how to choose them well, here.
2. Run the set at a fixed frequency against each AI. For example, every week against ChatGPT, Gemini, Perplexity and Claude. The fixed frequency matters: answers change over time, and irregular sampling mixes different periods into the same figure.
3. Record the appearances. For each answer: is the brand mentioned? The finer versions also record position (first recommendation or fifth?), competitors mentioned and sentiment — but the base metric is binary: appears or doesn't.
4. Divide. Appearances over total answers in the period. 40 prompts × 4 AIs = 160 weekly answers; if the brand comes up in 24, that week's Share of Model is 15%. Report it aggregated and broken down by AI: it's common (and revealing) to have 30% in Perplexity and 4% in Gemini, because each model draws on different sources.
A cousin metric not to confuse it with: citation share measures what percentage of the sources cited in the answers are yours (your site as a reference), whereas Share of Model measures brand mentions in the answer. You can be cited without being recommended and vice versa; mature reports carry both.
The limitations, before your client finds them
Putting a metric in the reporting obliges you to tell its weaknesses on day one — the alternative is explaining them defensively the day the number drops.
Volatility. The same question, to the same AI, on the same day, can give different answers: the models aren't deterministic, and their sources and versions change without warning on top of that. A weekly Share of Model swinging between 12% and 18% without anyone having done anything is normal. Practical consequence: the signal is in the trend over several months and in moving averages, never in the one-off figure.
Dependence on the prompt set. The number only means something relative to the questions chosen. With generic category prompts it'll come out low; with niche prompts where the client is strong, high. That's why two tools (or two agencies) give different figures for the same brand, and why comparing your 15% with the 40% another tool showed the client is comparing different thermometers. Operating rule: a fixed, documented set; if you change it, recalculate the baseline and say so in the report.
It doesn't measure volume. A 20% Share of Model doesn't tell you how many people ask those questions — the AIs don't publish query volumes the way the classic search engine publishes data. It measures presence in the answer, not the audience for the question. Present it as a coverage metric, not a traffic one.
And the commercial consequence of all three: Share of Model is a metric to be reported, not committed to. Promising "we'll get you to 30%" is promising the behaviour of systems nobody controls. In the proposal, the commitments go on verifiable work and the metric goes in as an indicator of progress — the full split between committable and reportable KPIs is here.
How to present it to a client
Four rules so the metric adds credibility instead of subtracting it:
- The definition first, in their language. One line in the report: "out of every 100 answers the AIs give about your sector, you appear in X". Without the word "model" if need be.
- Trend, not snapshot. A monthly evolution chart with a baseline from day one; the weekly figure stays in your kitchen.
- Always with the competition alongside. A 12% in isolation says nothing; a 12% when the direct competitor has 31% tells a story and justifies a plan. The comparative version is what sustains renewals.
- Stuck to the actions. The number next to what was done that month (content published, sources fixed) — if the report doesn't connect metric and work, the client will conclude the number moves on its own. Where it fits inside the full report is in the AI visibility report in minutes.
As for the operations: calculating it by hand (dozens of prompts × 4 AIs × every week × every client) doesn't scale beyond the first client. A platform like Surfeo does exactly this sampling — 40-75 prompts per client depending on tier, up to 4 AIs, weekly execution — and gives you the evolution and the competitors already calculated, with a PDF report for the client. Your work stays where it adds value: choosing the prompt set and explaining what the number means.
Frequently asked questions
Is it an industry-standard metric?
Not yet. The term is spreading, and the underlying idea (sampled presence in AI answers) is common to almost all AI visibility tools, but there's no standardised definition or single methodology. Treat it the way social media metrics were treated in 2010: useful, emerging, and demanding methodological transparency from whoever presents it to you.
What Share of Model is "good"?
There's no universal benchmark, and be wary of anyone who gives you one: it depends on the prompt set, the sector and the competition. The useful references are internal: your day-one baseline, your trend, and the gap with your direct competitors in the same set. As general context on the starting point in Spain: in our study of 9,865 SMEs, 91% appeared in only 1 of the 4 AIs — low initial numbers are the norm, not the exception.
Can I commit to raising a client's Share of Model?
Committing to a specific number, no — the volatility and opacity of the models make it a promise you can't keep. What you can commit to is the work that historically moves it (citable content, sources, authority, structured data) and rigorous measurement of its effect. It's the same distinction serious SEO has been making with rankings for years.
How does it differ from classic share of voice?
In the channel and the method. Classic share of voice measures your share in media, mentions or search results, usually with volume data behind it; Share of Model measures your share in AI-generated answers, via sampling with a prompt set, and with no question-volume data. They share the idea — what part of your category's conversation you occupy — and complement each other in a complete report.
Want to see the starting point before putting the metric in any report? Take the free AI visibility test for your client (or your own agency): you'll get their first snapshot of presence in the AIs and the argument for measuring it properly from there on.
