How to Measure Share of Model: Tracking Brand Visibility in ChatGPT, Gemini, and Perplexity

Someone in your target market asks ChatGPT for the best option in your category. The model names three brands. Yours isn’t one of them. No dashboard tells you this happened. No alert fires. Your Google rank might be excellent, your ad spend might be humming along, and you’d have no idea this moment of lost consideration ever occurred.

This is the gap Share of Model exists to close. It’s a genuinely new metric category, still young enough that its methodology isn’t standardized and its relationship to revenue isn’t fully proven, but the underlying behaviour it tracks is already real and growing. This article covers what it actually is, how to measure it properly, which tools exist right now, and where the metric’s real limitations sit.

What Share of Model Actually Is (and Isn’t)

Share of Model measures how often, how prominently, and how favourably a brand appears in responses generated by large language models, relative to its competitors, when someone asks a category-relevant question.

Worth addressing directly: there’s genuine disagreement about where this term came from. Jellyfish executives Jack Smyth and Tom Roach are widely credited with introducing the concept publicly in 2024, popularising it through agency thought leadership before building a commercial tracking platform around it. Separately, INSEAD published faculty research in mid-2025 analysing brand visibility across LLMs using Jellyfish’s own SOM platform data, which several later articles describe as the term being “formalised” academically. The honest read: Jellyfish coined and popularised the term, and INSEAD gave it a more rigorous research treatment shortly after. Both matter, and I’d be cautious of any source presenting only one half of that history as the whole story.

What makes Share of Model structurally different from anything in the Share of Voice or Share of Search family is worth sitting with. A keyword ranking is fixed: you’re in position one or you aren’t, and that position holds until something changes it. Share of Model is probabilistic. Ask the same model the same question ten times and you may get a different answer distribution each time. Ask a near-identical question and the brands mentioned can shift meaningfully. There’s no “page two” on an LLM’s answer either, a brand is either part of the response or it’s invisible, with nothing in between.

Unlike a Google ranking, which is a fixed position, Share of Model is inherently probabilistic – a brand might appear in 80% of responses to one category prompt and 20% to a nearly identical one, which means a single measurement is far less meaningful than tracking the trend across a consistent panel of prompts over time.

It’s also, critically, not something you can buy. Share of Voice was always partly a function of media spend. Share of Model emerges from what a model absorbed during training and what it retrieves at answer time through retrieval-augmented generation. No media budget moves this number directly, which is exactly why it demands a different playbook than paid media ever did.

Why This Metric Exists Now

The discovery layer genuinely shifted, and it shifted fast enough that most measurement infrastructure hasn’t caught up. Zero-click search behaviour, where a person gets their answer without visiting any website, has been climbing steadily, and AI-generated overviews and chat interfaces are now frequently the first (and sometimes only) touchpoint in a buyer’s research process. When that AI system recommends three competitors and never mentions you, that’s a real loss of consideration happening entirely outside any analytics tool built for a click-based internet.

Traditional rank tracking is structurally blind to this. It was built to monitor positions in an ordered list of links. LLM outputs aren’t an ordered list. They’re a generated paragraph that may or may not include your brand, described in whatever terms the model chose, with no URL, no position number, and often no indication to the end user that alternatives even exist. You cannot retrofit SEO rank-tracking logic onto this problem. It requires its own measurement category, which is exactly what Share of Model is trying to become.

The Three Things Share of Model Actually Measures

Collapsing this into a single number hides more than it reveals, so it’s worth breaking apart into its real components before you try to measure anything.

Mention frequency is the foundational layer: across a representative set of category prompts, what percentage of responses include your brand at all. This is the “recall” component, and it’s the number most casual tools report as “Share of Model” on its own, which is a simplification worth being aware of.

Prominence and position captures where and how your brand shows up when it does appear. Being the first brand named in a three-brand list is worth more than being mentioned in passing at the end of a longer response. A mention buried in a caveat is not equivalent to a direct recommendation.

Sentiment and framing is the layer most tools handle least well. An LLM can mention your brand while describing it unfavourably relative to competitors, positioning you as the budget option, or attaching a caveat that undercuts the mention entirely. Raw mention-rate tracking misses this completely, and it’s the dimension most likely to matter for brand perception even when the raw numbers look fine.

A composite Share of Model score that blends mention frequency, prominence, and sentiment into one number is more actionable than raw mention rate alone, because a brand can be mentioned frequently while being framed unfavourably relative to competitors, a distinction a single visibility percentage cannot capture on its own.

How to Measure Share of Model: Methodology

Start with a real prompt panel, not a keyword list. This is the single most common mistake in early Share of Model tracking. Keyword lists are built for search engines that match strings. LLM prompts need to mirror how actual buyers phrase questions conversationally: “what’s the best CRM for a 20-person sales team” rather than “best CRM software.” Build 25 to 50 prompts covering your category from different buyer angles, comparison questions, recommendation requests, problem-first questions where your category is the implied solution, before you query anything.

Query consistently across models, not just one. This is where most manual tracking efforts fail. The INSEAD analysis of Italy’s laundry detergent market, using Jellyfish’s proprietary SOM platform, found one brand held nearly 24% Share of Model on Meta’s Llama but under 1% on Google’s Gemini, and a second brand had a meaningful presence on Perplexity while being entirely absent from Llama. Measuring a single model and generalising the result is a real, documented methodological error, not a hypothetical risk.

Research from INSEAD analysing brand visibility across LLMs using proprietary tracking data found that individual brands’ Share of Model varied dramatically by platform, in one documented case ranging from nearly 24% on one model to under 1% on another for the same brand, category, and time period, which means single-model measurement produces a fundamentally incomplete picture.

Run each prompt multiple times, not once. Because the outputs are probabilistic, a single query per prompt is a snapshot with high noise. Several practitioners recommend somewhere in the range of dozens of runs per prompt for a statistically reliable reading, which is precisely why manual, one-off checking doesn’t scale into anything trustworthy. This is also the strongest practical argument for using a dedicated tool once you move past an initial diagnostic pass.

Track trend, not snapshot. A single Share of Model reading tells you almost nothing on its own. The useful signal is the trajectory after a specific action, if your Share of Model on a target prompt moves from 30% to 55% after a content or positioning change, that’s a meaningful, attributable result. Track monthly at minimum, and log what changed alongside the number.

Know the difference between API queries and chat-interface scraping, because they will give you different answers for the same brand on the same day. Direct model API calls return the raw model’s output without live web retrieval. Chat interfaces like ChatGPT with browsing enabled or Perplexity pull in real-time retrieval-augmented generation, which means the answer reflects current web content, not just training data. Neither is “more correct,” they’re measuring genuinely different things, and any tool or manual process should be clear about which one it’s reporting.

Tools for Tracking Share of Model

You don’t need paid software to get a first read. A spreadsheet with a dozen prompts, run manually across ChatGPT, Gemini, Claude, and Perplexity once a month, logging mentions and sentiment by hand, is a legitimate starting point and costs nothing beyond time.

Once you need scale, consistency, or multi-model coverage without the manual burden, several categories of tools exist. Semrush has folded an AI Visibility Toolkit into its platform as a paid add-on, covering prompt research, brand performance tracking, and technical AI-crawler auditing. Dedicated AI-visibility platforms, several of which use “Share of Model” as a named metric in their reporting, offer free entry tiers (typically a few hundred tracked queries per month) before moving to paid plans for continuous, multi-model monitoring. Surfer’s AI tracker and several newer entrants report visibility score, mention rate, and average position as separate metrics rather than one blended number, which aligns with the three-component breakdown above.

One honest caveat worth flagging before you pick a tool: different platforms report different numbers for the same brand on the same day, because they collect data differently. Some scrape the live chat interface (capturing retrieval-augmented answers), others hit model APIs directly (capturing raw model output without browsing). Neither is wrong, but comparing numbers across two different tools without accounting for this will produce a misleading sense of change over time. Pick one methodology, stay consistent, and treat cross-tool comparisons with real skepticism.

What Actually Moves Share of Model

This is not the backlink-driven SEO playbook wearing a new name, and content teams who treat it that way tend to see limited results.

Structured, factually dense content performs better than persuasive marketing copy. LLMs favour clear claims, specific numbers, and content that’s easy to extract and cite accurately, over vague positioning language that reads well to a human but gives a model nothing concrete to reference.

Consensus across independent sources matters more than volume from your own properties. If your brand is described consistently and accurately across review sites, industry publications, and third-party comparisons, not just your own website, that consistency is a stronger training and retrieval signal than a large volume of self-published content saying the same thing in isolation.

Semantic association, not just keyword matching, is what’s actually being optimised. This means deliberately connecting your brand to the specific use cases, problems, and contexts your buyers actually think in, across multiple content formats, rather than repeating a target keyword across pages the way traditional SEO once rewarded.

Improving Share of Model requires building consensus across independent, authoritative third-party sources rather than concentrating content on owned properties, since LLMs weight the agreement of multiple independent sources more heavily than repeated self-published claims, a meaningfully different signal than the backlink-driven logic that shaped traditional SEO.

The Honest Limitations of This Metric

There’s no standardised measurement methodology yet, and any source claiming otherwise is overselling where this discipline actually stands. Different tools, different prompt panels, and different query methods will produce different absolute numbers for the same brand. Treat the trend within a consistent methodology as the signal, and treat any single absolute percentage with real caution.

Benchmarks vary enormously by category concentration, and there’s no universal “good” number. In a fragmented category with many competitors, 15-20% Share of Model might represent genuinely strong performance. In a concentrated category with a dominant player, the leader might hold 40-50% and everyone else is fighting over the remainder. Your own trend against named competitors is the meaningful comparison, not an industry-wide benchmark pulled from a different category entirely.

The link to actual revenue and conversion is still unproven at scale. Early analysis has found correlation between AI mention frequency and increases in direct traffic or branded search volume in specific sectors like finance, travel, and beauty, but causality across other categories remains genuinely unconfirmed. This connects directly to the broader measurement problem covered in attribution and incrementality: Share of Model is a leading indicator worth watching, not yet a metric you can confidently tie to ROI the way a mature channel’s ROAS can be validated through a properly run holdout test.

Where This is Headed

Share of Model is an early-stage metric category, and being honest about that is more useful to you than pretending it’s a settled discipline with a clean playbook. The behaviour it tracks, brands being recommended or omitted inside AI conversations that happen before a website is ever visited, is real, growing, and currently invisible to almost every other tool in a standard marketing stack.

Start with a manual prompt panel this month if you haven’t already. You don’t need a platform to find out whether your brand shows up when someone asks an AI assistant the question your best customers are actually asking. Understanding what moves that number, and how to build the measurement discipline around AI visibility properly, is exactly the kind of hands-on capability YUP’s AI for Marketers course is built to teach.

Frequently Asked Questions about Share of Model