Reading time: 10 minutes
Share of model is the slice your brand takes up in AI answers for its category, compared with competitors. The earliest use of the term share of model we found is by Jellyfish, a digital media agency. So in GEO, AI share of model shows how often, how prominently and in what tone AI presents a brand.
In short
- The earliest use of the term share of model we found is by Jellyfish. A 2025 Harvard Business Review article defines it by how often, how prominently and how favorably LLMs present a brand.
- It differs from share of voice, because it measures what the model answers to a prompt, not what people publish.
- To measure it, fix unbranded prompts, competitors and engines, repeat the collection every month and read the slice by engine.
- The metric has no standard formula, and the acronym SOM also means share of market. So spell out the name in reports.
What is share of model?
Share of model is the metric that shows how much AI models talk about your brand when someone asks about the category. In other words, it compares your brand’s presence with competitors’ in answers from ChatGPT, Gemini and other LLMs. It always uses a fixed set of prompts.
The term got a definition in a Harvard Business Review article from June 2025. The authors are David Dubois, a professor at INSEAD, and two Jellyfish executives, John Dawson and Akansh Jaiswal. For them, the metric measures a new kind of brand awareness:
“how often, how prominently, and how favorably a brand is surfaced by LLMs to consumers. We call this awareness Share of Model (SOM).”
David Dubois, John Dawson and Akansh Jaiswal, Harvard Business Review (2025)
Then the article proposes three lenses. The first is the mention rate in each LLM. The second is the gap in brand awareness between people and models. Finally, the third is sentiment about the brand and the category. However, the article doesn’t publish a set formula or the number of prompts used.
If GEO is new to you, start with what GEO is and where AI share of model fits in.
Who created the term share of model?
We found no source that says who created the term, but the earliest use we found is by Jellyfish, an agency in the Brandtech Group. It uses the name with a trademark symbol (™). According to the HBR article, the agency pioneered a method that measures share of model through prompts at scale. In September 2024, Jack Smyth, the agency’s chief solutions officer, described the Share of Model platform in a Jellyfish post. He wrote that the tool analyzes how LLMs respond to brand, category and product questions.
In January 2025, an AdExchanger story reported that the platform had just opened in early access. Every day, it sends the models questions like “What brand of X would you recommend?” to collect their answers. So the Jellyfish share of model platform works like a market research survey in which the respondents are LLMs. Because the agency sells the platform, its numbers come from a party with a stake in the result.
What is the difference between share of model and share of voice?
Share of voice measures how much of the conversation that people and media produce belongs to the brand. Share of model, on the other hand, measures what the model answers when it gets a prompt. In other words, the first counts published content, while the second counts generated answers, which change with every run.
The HBR authors treat the metric as an offshoot of share of search, the brand’s slice of searches, and of share of voice. The difference, they say, is that share of model reflects the LLM’s perceptions and recommendations for a prompt. That is, it doesn’t reflect human intent or available content.
| Metric | What it measures | Where the data comes from |
|---|---|---|
| Share of search | Brand searches compared with competitors | Search volume |
| Share of voice | Brand mentions compared with competitors | Social media, press and media |
| Share of model | Brand presence and tone in answers | Repeated prompts in LLMs |
In practice, however, GEO tools often call the slice of mentions in AI answers share of voice, and that slice is the core of share of model. Peec AI, for example, calculates share of voice over the mentions of all brands in AI answers. Since each tool does the math its own way, only compare numbers within the same tool. NotHuman, which publishes this blog, works in the same market.
To add what people say about the brand on social media and in the press, see our guide to brand monitoring in AI.
Are share of model and visibility the same thing?
No. Visibility, or presence rate, counts how many answers include the brand, without looking at competitors. Share of model, in contrast, looks at the same answer through competition: how much of the space given to brands goes to yours. As a result, presence can rise while the slice falls, if competitors grow faster. In fact, the HBR article uses the presence rate as its first lens, the mention rate per LLM.
How to measure share of model: prompts, runs and engines
To measure AI share of model, run the same list of prompts without the brand name on the engines your audience uses. Repeat it several times a month. Then note every brand in each answer, yours and your competitors’. Finally, calculate each brand’s slice by engine and in total.
The collection protocol, with country, language and chats without personalization, is in our guide to measuring brand mentions and citations in AI. In that collection, four choices matter most:
- Prompts: 40 to 60 unbranded questions about the category, problems and comparisons. After all, a prompt with the brand name almost always gets an answer with it.
- Competitors: 3 to 5, picked from the brands AI cites most in the first run, not just from the client’s list.
- Runs: 5 runs per prompt on each engine, spread across the month, because the answer changes every time.
- Engines: the ones your audience uses, read one by one, since each engine checks its own sources.
Freeze prompts, competitors and engines for a quarter. Also, save each full answer with its date, engine and prompt. That way, if a competitor joins the list, you can recalculate the whole series with the old answers. The math of the slice also involves its own choices, such as the denominator and the weight of position.
In a GEO strategy with many clients, AI share of model takes thousands of answers a month. At that scale, collection moves to a GEO tool or to your own script.
How do you read the results?
Read the slice in three cuts: by engine, by prompt group and by competitor. For example, the brand may lead on comparison prompts and vanish on problem prompts, where the person doesn’t know yet which product they want. This cut shows where to work on content and sources, while the average serves the monthly report.
When the brand vanishes from a whole engine, start with the most common reasons AI doesn’t cite a brand.
How do you separate mentions, citations and recommendations?
Separate the three in a spreadsheet, with one column for each. A mention is the brand name in the answer, while a citation is the brand’s site used as a source. A recommendation, in turn, is the brand suggested as an option. Each one gives its own slice against competitors, and the three tell different stories.
- Slice of mentions: this is the most direct reading of share of model: how much of the space given to brands goes to yours.
- Slice of recommendations: counts only the answers that suggest the brand, in a list or with a verb like “consider.”
- Site citations: show whether AI uses the brand’s site as a source, even when it doesn’t say the brand’s name.
A brand can have a high slice of mentions and a low slice of recommendations. This happens, for example, when it shows up in comparisons where it loses to a competitor.
The third HBR lens goes beyond counting, because it splits the reasons for a recommendation into the brand’s strengths and weaknesses. In the spreadsheet, this becomes a column with the attribute the answer links to each brand, such as price, support or quality. This way, the report shows not only how much AI talks, but also what it says.
Which changes are noise and which are real?
A change is noise when it fits within the sample’s margin of error or disappears the next month. It is real when it exceeds the margin, repeats over two periods and shows up on more than one engine or prompt group. Therefore, never read a single rise as a result of your work.
Because AI answers change with every run, each month measures only a sample. The minimum protocol has 50 prompts, 5 runs and 4 engines, or 1,000 answers a month. In that sample, the margin of error ranges from ± 3 to ± 7 points. Since both months have error, the difference between them needs to exceed about 1.4 times the margin. So only treat a change above 4 to 10 points as real, and only once the next month confirms it.
Before you celebrate a rise in your slice, ask four questions:
- Did the list change? A new competitor or a swapped prompt moves the slice even if AI changed nothing.
- Did the engine change? A new model version or web search turned on moves every slice at once.
- Does the rise show up on more than one engine? A rise on one engine alone may be sampling chance.
- Did you go up or did a competitor go down? Because the slice is relative, a rival’s drop raises yours with no action on your part.
What are the limits of share of model?
As a GEO metric, AI share of model has four limits. There is no standard formula, and the prompt list shapes the result. Also, the number doesn’t say whether the answer is right, and no public study links the slice to sales. So use the metric as a signal of consideration, next to traffic and conversions.
- No standard formula: HBR proposes lenses, not a formula. Thus, 30% in one tool doesn’t equal 30% in another.
- The list is the sample: the slice holds for the chosen questions, not for everything people ask AI.
- Confusing acronym: SOM is also the acronym for share of market, the slice of sales. So spell out the name in reports.
- No proof of sales: we found no public study, with an open method, that links the slice of AI answers to sales.
The metric also doesn’t measure whether AI tells the truth about the brand. Jack Smyth himself, from Jellyfish, described the role of share of model this way to AdExchanger:
“The job of the Share-of-Model is not necessarily to get the most accurate information from the model – it’s to help a marketer understand what the model might be saying to customers.”
Jack Smyth, Jellyfish, to AdExchanger (2025)
So AI share of model is one GEO metric, not the only one. In the report, it sits next to AI traffic and conversions, as in the digital marketing KPIs that prove results.
Want to know how your brand shows up in AI answers today?
NotHuman is opening spots for agencies and brands to track how they appear in AI answers. Join the waitlist and we’ll let you know when a spot opens up for you.
Main sources
- Harvard Business Review, article by David Dubois, John Dawson and Akansh Jaiswal that defines share of model (2025).
- Jellyfish, post by Jack Smyth about the Share of Model platform (2024).
- AdExchanger, story about the launch of the Jellyfish platform (2025).
- Peec AI, how Peec AI calculates share of voice in AI answers (2026).
