Share of model is the percentage of AI-generated answers, across a defined set of prompts, engines and time window, in which your brand appears. The formula is simple: answers mentioning your brand divided by total answers, times 100. It is becoming the AI-era equivalent of share of voice, but it only means something if you measure it with a frozen prompt set and repeated runs.
Key Takeaways
- Share of model = brand mentions ÷ total AI answers × 100, across a fixed prompt set, engines and time window.
- Brand strategist Tom Roach credits Jack Smyth at Jellyfish with introducing the term; Jellyfish now sells a Share of Model platform.
- Single screenshots are useless because LLM answers vary between runs; measure rates, not moments.
- Separate mentions, recommendations and citations. They move differently and need different fixes.
- Share of model measures presence, not persuasion. Pair it with sentiment and AI referral traffic.
- A 30-prompt, 3-engine, 5-run panel is enough for most brands to start.
Where the Term Came From
Marketing has a history of "share of" metrics that arrive with a new dominant channel: share of market, share of voice in the media era, share of search in the Google era. In a January 2025 post, Tom Roach wrote that share of model is the next in that line, crediting Jack Smyth at Jellyfish with introducing it. Roach described it as the number of mentions of a brand by one or more LLMs as a proportion of total brand mentions in the same category.
Jellyfish (part of Brandtech) has since built a product under the Share of Model name. You do not need that product to use the concept, but credit where it is due.
Why It Is Replacing Share of Voice
Share of voice measured how loud you were in paid and earned media. Share of search, popularised as a leading indicator of market share, measured how often people searched your brand relative to competitors.
Both assume a person scans multiple options. In an AI answer, the person often sees three to five brands in one paragraph, picked by the model. If you are not in that paragraph, you are not in the consideration set. That is why teams now ask a new question: when a buyer asks an AI engine about our category, how often are we in the answer?
The honest caveat, which Roach also raised: we do not yet have strong public evidence that share of model predicts market share the way share of search was argued to. Treat it as a visibility KPI with commercial logic, not as proven proof of revenue.
The Formula
At its simplest:
Share of model (%) = answers mentioning your brand ÷ total answers × 100
A more useful category version:
Category share of model (%) = your brand mentions ÷ all brand mentions in the category × 100
The first tells you how often you are present. The second tells you how much of the conversation you own relative to competitors. Report both.
Worked example
You run 30 prompts × 3 engines × 5 runs = 450 answers.
| Brand | Answers mentioning brand | Presence rate | Share of all mentions |
|---|---|---|---|
| You | 135 | 30.0% | 135 ÷ 600 = 22.5% |
| Competitor A | 225 | 50.0% | 37.5% |
| Competitor B | 150 | 33.3% | 25.0% |
| Competitor C | 90 | 20.0% | 15.0% |
| Total mentions | 600 | 100% |
(Illustrative numbers, not client data.) Presence rates add to more than 100% because one answer can mention several brands.
A Repeatable Measurement Protocol
Storylake's measurement protocol is one of the clearest public write-ups, and I use a close variant.
1. Build and freeze the prompt set
Write 20-50 prompts in real buyer language, split across:
- Discovery: "best online data science course for working professionals"
- Comparison: "Brand X vs Brand Y for beginners"
- Problem-framing: "how do I switch from non-tech to a developer job"
Source them from Search Console queries, sales calls, support tickets and community threads. Freeze the list for at least a quarter. Changing prompts breaks comparability.
2. Choose engines and report them separately
ChatGPT, Perplexity, Gemini, Google AI Overviews and Claude draw on different sources. Averaging them hides the story. A brand can be strong in Perplexity and invisible in ChatGPT.
3. Run each prompt multiple times
Storylake suggests a floor of 10 runs per prompt per engine, with 30 preferable. If budget or time is tight, I start at 3-5 and state that clearly in the report. Fewer runs mean wider error bands.
4. Define classification rules before you look
Decide in advance what counts as:
- Mention: brand name appears anywhere.
- Recommendation: brand is suggested as a good option for the user's need.
- Citation: your URL is linked as a source.
Write down edge cases: does a mention in a "cons" list count? Does a sub-brand count?
5. Score presence first, then recommendations
Presence rate is the headline. Recommendation rate is the one that matters commercially. Citation rate tells your content team which pages are doing the work.
6. Publish the method with the number
Every report should list prompt set, engines, run counts, dates, classification rules and a variance range. Without that, the number cannot be challenged or reproduced.
Mentions vs Recommendations vs Citations
| Level | What it tells you | Main lever |
|---|---|---|
| Mention | The model knows you exist in the category | Brand building, PR, third-party coverage |
| Recommendation | The model thinks you fit the need | Reviews, comparisons, clear positioning |
| Citation | Your page is a source | Answer-first content, original data, crawl access |
In edtech, I have seen brands with high mention rates and low recommendation rates, because the model associated them with outdated pricing or old controversies. That is a positioning problem, not a content-volume problem.
Pitfalls to Avoid
- Brand-name prompts. "Is [brand] good?" inflates your number. Keep them out of the headline metric.
- Screenshot reporting. One run is anecdote.
- Blended engines. Averages hide where the problem is.
- Ignoring sentiment. Share of model measures presence, not whether the description is flattering or accurate.
- Ignoring location and language. Answers differ by market. An Indian brand selling in the US needs separate panels.
- Moving prompts. If you edit the prompt set, restart the baseline.
How to Report It to Leadership
A one-slide format that works:
- Headline: category share of model this quarter vs last, with the range.
- By engine: presence rate per engine.
- By intent: discovery vs comparison vs problem prompts.
- Top cited domains: the third-party sites engines lean on for your category.
- Actions: three specific tasks, such as "get included in the two listicles ChatGPT cites most," "refresh the comparison page," "fix outdated pricing on review sites."
Pair it with AI referral sessions from GA4 and branded search trends, so leadership sees presence, traffic and demand together.
How to Improve Share of Model
- Earn third-party mentions on the sites engines already cite for your category.
- Publish original data that others reference. When I worked on Masai School's social growth (Instagram 26K to 117K, LinkedIn 50K to 160K), the lasting value was a body of distinctive content others could point to, not just follower counts.
- Write answer-first pages with direct answers under question headings.
- Fix factual drift on directories, review sites and your own old pages.
- Keep crawl access open for AI search bots.
Tools vs DIY
A manual panel of 30 prompts, 3 engines and 3-5 runs is achievable in a spreadsheet each month. Once you need 10+ runs, multiple markets or weekly cadence, an AI visibility tracker makes sense. I compare them by budget in a separate post. Whatever tool you pick, check that you can export raw answers so you can apply your own classification rules.
FAQ
What is share of model in marketing?
Share of model is the share of AI-generated answers, across a defined set of prompts, engines and time period, in which your brand appears. It is often framed as the AI-era successor to share of voice.
How do you calculate share of model?
Divide the number of answers mentioning your brand by the total number of answers, then multiply by 100. For a competitive view, divide your mentions by total brand mentions in the category.
Who coined the term share of model?
Tom Roach credits Jack Smyth at Jellyfish with introducing the term. Jellyfish has since launched a platform under the Share of Model name.
How many prompts do I need to measure share of model?
Most protocols suggest 20-50 prompts in real buyer language, frozen for a quarter. More prompts help only if they reflect genuine demand.
How many times should each prompt be run?
Storylake's protocol recommends at least 10 runs per prompt per engine, and 30 if possible. Fewer runs are acceptable for a first baseline if you report the wider uncertainty.
Is share of model the same as AI share of voice?
They are used almost interchangeably. Some tools use "AI share of voice" for the competitive ratio and "visibility" for presence rate. Always check how a tool defines its metric.
Does share of model predict revenue?
Not proven yet. It has clear commercial logic because AI answers shape consideration sets, but public evidence linking it to market share is still limited. Pair it with AI referral traffic and branded search.
Should share of model be reported per engine?
Yes. Different engines use different sources and user bases, so a blended number hides where you are strong or weak.
Want to Set Up Your Share of Model Baseline?
I have 4+ years of marketing experience in organic growth, SEO and AEO for edtech and startup brands, and I help teams build measurement they can actually defend in a leadership meeting. If you would like to see my work or set up your first prompt panel, reach out via the contact form at younusfardeen.in.