The Share of Model (SoM) is an emerging metric in digital marketing that measures how often your brand appears in AI-generated answers. Positioned as the successor to the pre-AI Share of Voice (SOV), often used to measure organic reach, it has quickly gained popularity among marketing professionals, who are building an ecosystem of tools to track it.
The SoM metric solves an undoubtedly real problem. Hundreds of millions of people use AI tools to research products, services, and brands, so knowing how often your brand is mentioned is the first step in optimizing AI channels. But can Share of Model be accurately measured?
What share of model is trying to measure?
Share of Model purportedly quantifies how often your brand shows up in AI-generated answers to various relevant, often category-level, questions. For example, there are dozens of residential proxy providers on the market. When an LLM is asked the question “what’s the best residential proxy provider?” the amount and type of answers are quite varied, even if the same model is asked several times.
To measure our own Share of Model, for example, the answers would be analyzed to express a percentage of all IPRoyal brand mentions retrieved with AI data collection methods while also filtering different prompt intentions, question variations, and answer structures. Specialized vendors, such as Profound, Otterly, Sight AI, Share of Model, and even established marketing tools, like Ahrefs or Semrush, use some version of this formula to calculate SoM-related metrics.
SoM= Your brand mentions in AI answersTotal brand mentions in AI answers x 100
Much of this analysis can be done manually with multiple AI subscriptions and a spreadsheet. The appeal of tools is that they suggest relevant prompt sets, models, and automate the process to send queries while analyzing outputs in a timely fashion. Many CMOs are quick to make this into a new quarterly KPI.
If conversational AI becomes the primary discovery channel for customers, like search engine results are now, appearing in them is crucial for brands. SoM offers a single, easy-to-calculate signal of AI visibility that marketers can track over time and benchmark against competitors.
Whether SoM is effective as such a metric depends on measurement consistency. Significant variations in SoM percentages between models, prompts, and additional context make it difficult to rely on AI chats as a stable discovery channel.
The repeatability problem
One January 2026 study got 600 volunteers to run 2,961 prompts across ChatGPT, Claude, and Google Search’s AI Overviews to test the reliability of the SoM metric. The team used twelve prompt categories of common products and services, asking AI chats to make a best-of list 60 to 100 times per platform.
The result showed that AI tools build an identical list less than 1% of the time across repeated runs of the exact same prompt. The same list in the same order appeared less than 0.1% of the time. In other words, there’s a 99% chance the answer would come back different each time you ask.
That’s far from a minor margin of error, and should already raise red flags for those who want to justify SoM as a KPI or even a serious standalone metric. The problem becomes even worse when we account for how real users actually phrase their AI queries.
While search queries tend to be shorter, AI questions and answers are much longer. Two people with the same basic intent can produce prompts that have almost nothing in common. Additionally, the prompts rely heavily on previous conversation context with their AI chatbot.
Black box personalisation
A Share of Voice (SoV) score, in contrast, built around media mentions and search engine rankings, reflects something much more stable. A social media mention either exists or it doesn’t. Google’s rankings in a search result page, while personalized, remain at least somewhat predictable when tested from different locations with various tools.
While the systems and algorithms behind both metrics are proprietary, social media and search engines are at least observable and consistent enough to rely on. SoM proponents assume that SoV logic directly transfers to AI-generated answers, but the current data suggests the opposite, and it couldn’t be otherwise.
Large Language Models (LLMs) are probability machines built with a degree of randomness, resulting in the same inputs producing different outputs. Add the fact that models rely on accounting for previous conversation context, users’ history, location, and other factors, and we can see that this is by design.
SoM might have a nice-sounding formula, and vendors sell convenient dashboards for measuring it, but AI output differences make it a questionable metric. Google’s ranking algorithm has been reverse-engineered and studied for two decades by an entire SEO industry. AI labs are equally, if not more, protective of their business.
How a model decides which brands to mention, in what order, and under what personalization conditions is a well-guarded trade secret at best. At worst, if we agree with most radical sceptics, AI is a black box that humans may never fully understand at all.
Where share of model tools are useful?
The worst-case scenario might be too radical, and concessions must be made to SoM proponents. The already mentioned research also found that frequency of appearance is comparatively more stable than rank or exact list match. In other words, while each query ranks brands differently, how often they appear across prompts was comparatively more stable.
The same handful of brands show up in the majority of responses, even when other factors change. Tracked over enough runs, SoM might signal something useful after all. If your brand appears in 70% of relevant AI responses this quarter and drops to 40% the next, it’s definitely a signal worth investigating.
Accounting for this, the objective accuracy of SoM isn’t that important if the measure stays consistent. We can calculate the difference between values over time with SoM, known in statistics as delta, which allows marketers to estimate whether their tactics are making an impact. Often, this is enough to improve AI exposure.
The problem here is that periodic comparisons assume that LLM model brand mentions remain consistent between measurements. As we saw earlier, there are reasons to think that LLMs are largely black boxes updated and personalized without public documentation. The premise of consistency needed for the delta method is still easy to doubt.
As such, taking SoM as a metric equal to SoV is still questionable, even if there is something to report. More data could improve the picture and, in theory, AI labs could become more cooperative and share more data. The third-party SoM tools running thousands of relevant prompts daily should also have it.
If it’s technically possible and data isn’t shared, the conclusions might not be in their interests. An AI lab that discloses how a model surfaces brands invites the same reverse engineering that Google has faced from SEOs for decades. It’s safe to think they’d like to avoid this.
SoM tool vendors are also suspiciously opaque about the data they have and how many prompts we need to see more stable results. We could speculate that the data undermines the credibility of metrics like SoM.
Yet, there’s still a possibility that with enough volume, SoM becomes a better metric. Until we have the data, your marketing budget is better spent on digital marketing fundamentals that are proven to improve, among other things, AI visibility as well.
That’s not to say CMOs shouldn’t push AI initiatives and tests, but it’s too soon for them to replace traditional marketing tactics. Together with things like brand mentions, backlinks, and social presence, SoM metrics are a useful but somewhat experimental signal.
Conclusion
It’s better to treat SoM-related metrics as a rough signal, not an objective KPI. The SoM tool vendors sell the metric as offering a certainty that no available data backs up. Even if ignoring it would be a mistake, treating it as a match for SoV could be a bigger one.
Author:
Julius Narkus,
Chief Marketing Officer at IPRoyal
