Share of Model
The metric clients are starting to ask for in place of rankings. Some call it Share of Model. At AIOInsights we call it AI Share of Voice, and the difference is not just branding, as the section below sets out. Either way the underlying question is the same: across the buying questions in your category, how often does an AI name you rather than a competitor. The idea is sound and the demand is real. The catch is that the number is dangerously easy to fake, because AI answers vary between runs, and a percentage reported without its denominator is not a measurement at all.
Share of Model is your slice of the AI answer: across a defined set of buying questions in your category, how much of the time does a model name you rather than someone else. It is the AI-era analogue of share of voice, and it is only as credible as the method behind it. AIOInsights reports this as AI Share of Voice.
Where the Term Came From
Share of Model is not a generic industry phrase. It was pioneered by Jack Smyth at the agency Jellyfish in 2023, and set out publicly in an Adweek piece in March 2024 under a line that explains the whole idea: every good CMO knows their share of the market, but do they know their share of model. In December 2024 Jellyfish launched a Share of Model platform to analyze how different models perceive brands and products, with Danone and Chivas Brothers among the named early clients.
A later write-up co-authored by INSEAD's David Dubois with two Jellyfish authors gave it the definition most people now quote: the metric measures how often, how prominently, and how favorably brands appear in AI-generated responses. Worth knowing when you cite it: the origin story and the academic framing both come substantially from the people who invented and sell it. That does not make the metric wrong. It does mean the burden of proof sits with whoever reports the number.
Watch: Meet the New Metric of Marketing Success by the Content Marketing Institute, in which Robert Rose walks through Share of Model as a way to gauge a brand's standing by its mentions across one or more large language models. Source: YouTube.
Why We Call It AI Share of Voice
You will see the two names used interchangeably, and in most vendor dashboards they now mean the same thing. There is a real distinction underneath, and it is the reason we report under the second name rather than the first.
Share of Model, in its original framing, is about what the model knows: the associations carried in its weights, how it perceives your brand against competitors, and why it would recommend you. AI Share of Voice is about what the model says: your share of the mentions across live answers to real category questions. The first is a perception audit of the model itself. The second is a tally of observable outputs.
The practical consequence is decisive. Only the second one can be measured by anyone from the outside. Nobody outside a model lab can inspect what is in the weights, so nobody outside one can honestly claim to report what a model knows. What you can do, rigorously and repeatably, is ask the buying question many times across several engines and count who gets named.
So when a report says Share of Model and shows you a percentage, that percentage is almost always a mention tally: the second thing, presented under the name of the first. That is a legitimate measurement wearing a label that overstates it. We use the name that matches what we actually observed. If a vendor tells you they measure what a model knows, ask them how they got inside the weights.
Why the Number Is So Easy to Fake
The single biggest problem with Share of Model is that AI answers are not stable. Ask the same question twice and you can get two different lists of businesses. That is not a bug in the measurement, it is a property of the thing being measured, and any method that ignores it produces a number that looks precise and means nothing.
The evidence here is specific. In an analysis published in April 2026, Keller Maloney of Unusual found that changing a single word in the prompt, from "best" to "top", moved a brand's share of voice by 17 percent, and that only 16 of 100 prompt pairs returned identical results. A July 2026 arXiv paper by D. Żatuchin decomposed the variance across 12,933 responses covering 20 brands, 8 languages, and 3 models, and reached a blunter conclusion: query language explained 26.5 percent of the variance while brand identity explained 1.5 percent, and the intraclass correlation of 0.0146 means a single AI answer carries almost no brand-discriminating signal. That paper studies sentiment rather than mention share, and its author is affiliated with a commercial vendor, so treat it as strong supporting evidence rather than the final word. The direction it points is not seriously in dispute.
Three failure modes follow, and every one of them is common in the reports being sold right now. A number produced from a handful of runs is noise. A number whose prompt set was chosen by the vendor has a denominator nobody can audit, and a vendor with an incentive can pick questions their client happens to win. And a percentage published without stating how many runs it came from cannot be checked, reproduced, or defended in a room.
Publish the denominator
Share must be appearances divided by the number of runs that actually succeeded, and that count must be printed alongside the percentage. If an engine errors out on three of ten runs and you still divide by ten, you have understated the share. If you quietly drop the failures without saying so, nobody can reproduce your figure.
Name the engines and the models
A brand can dominate one model and be invisible in another, so a single-engine figure is not a share of model, it is a share of one model. State which engines were queried, which model versions, and on what date. Model versions change underneath you, which makes an undated figure worthless six weeks later.
Fix the questions before you run them
Write down the money questions your buyers actually ask, agree them with the client, and keep them constant between reporting periods. A prompt set edited after the results come in is not measurement, it is selection. Holding the questions fixed is also the only way a change between months means anything.
Why It Matters
Rankings are losing their meaning as an executive reporting number. When a buyer asks an engine who the best option is and reads one synthesized answer, position four on a results page nobody scrolls is not the thing that decides the sale. Being named is. Share of Model is the first metric that reports on that directly, which is why boards are asking for it and why it is displacing rank tracking in quarterly decks.
It also reframes the competitive picture usefully. A share figure is inherently relative, so it tells you not just whether you appear but who is taking the space you are not occupying. That is a more actionable brief than a ranking ever was, because the named competitors in an AI answer are a concrete list of sources the model trusts more than yours, and trust is something you can go and build.
A Share of Model figure with no run count, no engine list, and no date is not a measurement. It is a number chosen to look like one. Ask any vendor quoting you a share for all three before you put it in front of a client.
How AIOInsights Reads This Signal
We measure the observable form and we report it under its accurate name, AI Share of Voice. For a fixed money question in a client's category, we query the answer engines repeatedly, force each response into a structured ranked list rather than scraping prose for names, group the naming variants of a business so that one firm is not counted as three, and report appearances divided by the number of runs that succeeded. Every figure ships with the question, the engines, the model versions, the run count, and the date attached. If an engine fails partway through, the failures are excluded from the denominator and the reduced count is stated, so a partial outage can never inflate a share.
Two honest limits. Because answers vary between runs, a share is an estimate with a spread around it, not a fixed fact, and we present it that way rather than implying a precision the underlying system does not have. And AI Share of Voice is a separate thing from the AIOInsights score: the score is deterministic, computed from observable signals on your site, and identical every time you run it, whereas a share is sampled from a system that is not deterministic at all. We keep them apart deliberately. Mixing a sampled number into a deterministic score would quietly corrupt both.
Check Whether AI Can Find Your BusinessKeep reading the lexicon: Citation Share, Deterministic Scoring, and Answer Engine.