What Is Share of Model and How Should You Measure It?
A dashboard can make share of model look precise. The percentage is clean. The methodology behind it often is not.
Which prompts were tested? In which AI products? How many observations were collected? Did a passing brand mention count the same as a recommendation? Were citations included? Which competitors were in the denominator?
Share of model can be a useful way to measure brand visibility inside AI-generated answers, but only when the sample and counting method are explicit. Treat it as a repeatable measurement of a defined set of answers—not as a ranking position, market-share estimate, or shortcut around the rest of your measurement stack.
What is share of model?
Share of model generally describes how often a brand appears within a defined sample of AI-generated answers, usually compared with competitors or the total number of brand appearances in that sample.
The phrase has gained traction quickly. Jellyfish was publicly using and commercializing the Share of Model platform by 2024 as a way to analyze how large language models perceive brands, products, and services.
The tidy name hides an untidy reality: there is no single standardized formula everyone is using.
One methodology might calculate a brand’s share of all brand mentions. Another might track the percentage of tested answers that mention the brand. Another might combine presence, position, sentiment, citations, or other factors into a weighted visibility score.
Those can all be legitimate measurements. They are not automatically the same measurement.
That distinction matters because share of model is not a traditional search ranking. It does not tell you a brand’s rank across AI systems, its market share, or its share of potential demand.
The number describes the sample that produced it. Nothing more should be smuggled into the interpretation.
Why two share-of-model scores can disagree
Two teams can measure the same brand and get very different results if they use different prompt sets, AI products, repeat-run rules, or definitions of what counts.
That does not automatically mean one is wrong. It means methodology is part of the metric.
| Methodological choice | Why the score changes |
|---|---|
| Answers containing a brand vs. all brand mentions | Changes the denominator |
| One observation vs. repeated observations | Changes the sample |
| One AI product vs. several | Changes the measured universe |
| Mention vs. recommendation vs. citation | Changes the event being counted |
| Fixed vs. changing prompts | Changes comparability over time |
Prompt selection is particularly important. A panel built around category recommendations is measuring something different from one built around troubleshooting questions, brand comparisons, or educational searches.
The competitor set matters too. Add or remove a competitor and a relative-share calculation can move even when the underlying answers barely change.
This is familiar territory for marketers who already understand the difference between a metric and its measurement design. The same discipline applied to traditional SEO services needs to carry into AI visibility work. The acronym changed. The requirement to define what you are measuring did not.
A defensible share-of-model measurement protocol
A useful share-of-model program starts before the first prompt is run.
Define the question first
Start with the business question, not the score.
Are you trying to learn whether your brand appears in category recommendations? Whether it is mentioned when prospects describe a particular problem? Whether competitors appear more consistently for a defined set of buying questions?
A clear question keeps the measurement from expanding until the number means everything and therefore nothing.
Build and freeze the prompt panel
Create a documented group of prompts tied to the audience decisions you want to observe.
Record the major intent groups in the panel and keep the baseline set stable. If you materially revise the prompts later, version the methodology instead of quietly replacing part of the sample and drawing a trend line through the change.
There is no defensible universal rule that every company needs the same number of prompts. What matters is knowing what the panel represents and maintaining enough consistency to compare one measurement period with another.
Report AI surfaces separately
ChatGPT Search, Google AI Overviews and AI Mode, Gemini, Bing/Copilot, Claude, and Perplexity are distinct products and reporting environments.
Treat them that way.
A single cross-platform number can be created, but it should not be the default simply because executives enjoy one-cell dashboards. If you combine platforms, disclose the aggregation or weighting rule and keep the underlying platform-level results available.
First-party reporting already illustrates why the distinction matters. Google’s generative AI performance reports provide visibility data for Google’s own generative search experiences. Microsoft’s Bing AI Performance reports citations and cited pages for supported Microsoft AI experiences.
Those are not interchangeable datasets.
Define exactly what counts
“Visibility” is too broad to serve as the counting rule.
Decide whether you are measuring:
- Brand mention rate
- Share of competitive mentions
- Recommendation rate
- Citation rate
- Citations to brand-owned domains
- Sentiment or accuracy as a separate qualitative measure
Do not collapse all of those into “share of model” unless the formula explicitly explains how.
Mentions and citations deserve particular care. Microsoft says citation activity in Bing AI Performance does not indicate page importance, ranking, or placement. That is a useful guardrail well beyond Bing: count the thing you observed without quietly renaming it as authority.
Repeat and timestamp observations
AI-generated answers can vary.
Record the prompt, platform, date, run number, brands mentioned, competitors mentioned, citations, and any relevant qualitative notes. If repeated observations are part of the methodology, keep that rule consistent across periods.
Trend like against like. Otherwise, an apparent improvement may say as much about a changed sample as it does about the brand.
Version the methodology
Method changes are sometimes necessary. New products launch. Competitor sets change. Prompt panels improve.
Document them.
A change to prompts, platforms, counting rules, competitor sets, or weighting should create a new methodology version or an obvious break in the historical comparison.
A clean chart is not worth pretending two different measurement systems were identical.
What should sit next to share of model on the dashboard?
Share of model works best as one layer of measurement, not the whole stack.
For Google, platform-native reporting can show generative-AI visibility data such as impressions and appearing pages. Microsoft can provide citation and cited-page data for its supported AI experiences.
Your controlled prompt sample answers another question: How frequently does our brand appear inside the answer situations we deliberately chose to monitor?
Then come downstream business measures: referral traffic, conversions, branded demand, pipeline, or revenue where attribution actually supports the connection.
Those layers should inform one another without being mistaken for one another.
That is also the more useful way to approach broader AI Search services. An AI-visibility program should tell you what was observed, what was calculated, and what can reasonably be inferred from it.
When share of model belongs in your KPI set
If you still need a starting inventory of visibility gaps, a digital marketing audit can sit beside this measurement work. Use share of model as a KPI when the prompt sample is documented, the platforms are disclosed, the counting rules stay stable, and historical comparisons use substantially the same methodology.
Use it as a directional diagnostic when the sample is small, coverage changes frequently, or the methodology is only partially disclosed.
Skip the score, or at least refuse to make decisions from it, when a provider cannot explain the denominator, hides the tested prompt set, combines multiple platforms without explaining how, or implies that the resulting percentage proves rankings, market share, pipeline, or future revenue.
A metric does not become strategic just because somebody added a percent sign.
Share of model is useful precisely because it has limits
Share of model can tell you something valuable: within a defined set of AI-generated answers, how often is your brand showing up relative to the opportunity you chose to measure?
That is enough.
It does not need to become a secret ranking factor, a market-share proxy, or a revenue forecast to earn a place in the measurement stack.
The useful version is transparent, repeatable, and boring enough that another analyst could reproduce it. Measurement quality comes from consistent sampling and clear definitions, not extra decimal places.
If your team is trying to establish a defensible baseline for AI visibility, explore Perfect Search Media’s AI Search services.
Frequently asked questions
Is share of model the same as a search ranking or market-share number?
No. It describes how often a brand appears inside a defined sample of AI-generated answers, under a stated counting method. It is not a rank across AI systems, a market-share estimate, or proof of pipeline or revenue.
Can you combine ChatGPT, Google, Bing, and other platforms into one score?
You can, but it should not be the default. Those products are different reporting environments. If you combine them, disclose the rule and keep the platform-level results available.
When does the metric belong in a KPI set?
Use it as a KPI only when the prompt sample, platforms, and counting rules are documented and stay stable enough to compare periods. Use it as a directional diagnostic when the sample is small or the method is only partly disclosed. Do not make decisions from a score whose denominator or prompt set cannot be explained.
Sources
- Jellyfish, “Share of Model Platform” announcement (December 2024):https://www.jellyfish.com/en-us/news/jellyfish-launches-the-share-of-model-platform/
- Google Search Central, “Introducing Search Generative AI performance reports in Search Console”:https://developers.google.com/search/blog/2026/06/gen-ai-performance-reports
- Microsoft Bing Webmaster Blog, “Introducing AI Performance in Bing Webmaster Tools”:https://blogs.bing.com/webmaster/2026/2/Introducing-AI-Performance-in-Bing-Webmaster-Tools-Public-Preview/