What is Share of Model (SoM) and How to Measure It?
Share of Model (SoM) is the percentage of tracked AI prompts in which a generative engine names your brand, measured against the competitors it names instead. It is the closest thing AI search has to market share: not how often you rank, but how often you are part of the answer.
If your brand appears in 12 of 50 buyer prompts, your mention rate is 24%. If the three competitors in that set appear in 38, 31 and 27, your Share of Model — your slice of all brand mentions across those answers — is roughly 11%. That second number is the one that tells you whether you are competitive.
This guide covers the definition, the formula, a worked example, and — the part most teams get wrong — how to measure it without quietly rigging the result in your own favour.
Key takeaways
- SoM measures relative presence inside AI answers. Mention rate alone flatters you; SoM does not.
- The formula is simple. The prompt set is the hard part, and it determines whether your number means anything.
- Generative answers are non-deterministic, so a single run is not a measurement. Repeated sampling is mandatory.
- SoM should be paired with position and sentiment — being named last, sceptically, is not the same as being recommended.
- Track SoM as a trend against a fixed prompt set, not as an absolute score to compare across companies.
What is Share of Model?
Share of Model is a competitive visibility metric for generative search. You define a set of prompts that represent how buyers actually ask about your category, run them across AI engines, count which brands get named, and express your brand's share of those mentions.
The name is deliberate. Share of Voice measures your presence across the open web — mentions, coverage, ad impressions. Share of Model measures your presence inside the model's answer, which is a much narrower and more consequential space. The web has room for everyone. An AI answer has room for three or four brands.
That constraint is the entire reason the metric exists. In classic search, ranking eighth still earns clicks, so a linear metric like average position is informative. In a generated answer, the fifth-best-known brand in a category is usually not mentioned at all. Visibility is closer to binary, so you need a metric that captures presence and competitive displacement at the same time.
The underlying selection mechanics — why an engine names some brands and not others — are covered in how AI search engines decide which brands to recommend.
The Share of Model formula
The basic calculation:
SoM = (Your brand mentions) ÷ (Total brand mentions across all tracked brands) × 100
Where a "mention" is one instance of a brand being named in one answer to one prompt. Some teams count a brand once per answer regardless of repetition, which is the cleaner approach — repetition within a single answer is a stylistic artefact, not extra visibility.
A worked example
Fifty prompts, four brands, one engine, three runs each:
| Brand | Prompts mentioned in (of 50) | Mention rate | Share of Model |
|---|---|---|---|
| Competitor A | 38 | 76% | 34.5% |
| Competitor B | 31 | 62% | 28.2% |
| Competitor C | 29 | 58% | 26.4% |
| Your brand | 12 | 24% | 10.9% |
| Total | 110 | — | 100% |
Your mention rate of 24% sounds tolerable in isolation. Your Share of Model of 10.9% tells the real story: in the answers where you do appear, you are sharing space with three better-established brands, and in three-quarters of the prompts you are absent entirely.
This is why SoM beats mention rate. Mention rate can rise while your competitive position deteriorates, if competitors are gaining faster than you.
Weighted Share of Model
Not all prompts carry equal commercial weight. A purchase-intent prompt ("which tool should I buy for X") matters more than a definitional one ("what is X"). If your prompt set spans intents, weight them:
Weighted SoM = Σ (SoM per intent × intent weight)
| Intent | Example prompt shape | Suggested weight |
|---|---|---|
| Purchase | "Best tool for [use case], budget [X]" | 0.5 |
| Comparison / alternative | "[Competitor] alternatives", "X vs Y" | 0.3 |
| Discovery / informational | "How do I solve [problem]?" | 0.2 |
Adjust the weights to your funnel, then keep them fixed. Changing weights between reporting periods makes the trend meaningless — which is exactly why some agencies do it.
What SoM does not tell you
SoM is a presence metric. Alone, it is incomplete. Track it alongside:
| Companion metric | What it adds |
|---|---|
| Position in answer | Whether you are named first or listed last. First-named brands carry disproportionate weight with users. |
| Sentiment | Whether you are recommended or merely acknowledged. See the role of sentiment and credibility in AI search visibility. |
| Citation rate | Whether your own domain is used as a source. This distinguishes "the model knows you" from "the model reads you" — and the second is far easier to influence quickly. |
| Competitive gap by intent | Which prompts you lose. Losing purchase prompts while winning informational ones is a specific, fixable problem. |
A brand with 30% SoM that is always named last, with hedged language, is in a weaker position than one with 20% SoM that is consistently named first as the recommended option.
How to measure SoM without skewing the result
The formula takes five minutes. Everything below is where measurements go wrong.
1. Build a representative prompt set
This is the single largest source of error. A prompt list written by someone who wants a good report will produce one. Rules that keep it honest:
- Source prompts from sales calls, support tickets and community threads — not from your keyword tool and not from your imagination.
- Write them the way people talk: long, contextual, first-person. "I run marketing at a 40-person B2B SaaS company and we're not showing up when prospects ask ChatGPT about our category — what tools track that?" is a real prompt. "best AI visibility tool" is a keyword wearing a costume.
- Cover all three intents. A set skewed towards discovery prompts inflates SoM for content-heavy brands and hides purchase-stage weakness.
- Do not include your brand name in the prompt. "Is Geomyze good for tracking AI visibility?" measures nothing except whether the model has heard of you. Branded prompts belong in an accuracy audit, not a SoM calculation.
- Aim for 30–100 prompts. Fewer than 30 and variance dominates; more than 100 and maintenance cost outruns the insight.
2. Choose the competitor set honestly
Include the brands buyers actually consider, including the ones you dislike. A competitor set that omits the category leader produces a flattering number and a useless one. Five to eight competitors is usually right.
Also record brands you did not nominate but that appear anyway. Those are the most valuable rows in the report — they are competitors the market has, and you did not know about.
3. Sample repeatedly
Generative engines are probabilistic. Ask the same question three times and you may get three different brand sets. A single run is an anecdote.
Practical minimum: three runs per prompt per engine, averaged. If a brand appears in one run of three, record it as 0.33, not 1. Where two reporting periods differ by less than the observed run-to-run variance, report "no change" rather than inventing a narrative.
4. Fix the engine mix
SoM across ChatGPT, Gemini and Perplexity is not comparable to SoM across ChatGPT alone. Decide which engines you track, weight them by your audience's actual usage, and do not change the mix mid-programme. Report per-engine numbers alongside the blended figure — the gaps between engines are often the most actionable finding, because they tell you whether your problem is retrieval or model memory.
5. Neutralise personalisation
Run tests from clean sessions without account history, and hold location constant. An engine that has learned your preferences from months of use will over-report your own brand. This sounds obvious and is skipped constantly.
6. Freeze the baseline
Capture SoM before you change anything. Without a baseline, every subsequent number is a standalone figure with nothing to compare against, and you will end up arguing about whether 14% is good.
What counts as a good Share of Model?
There is no universal benchmark, and you should be sceptical of anyone quoting one. SoM depends entirely on your prompt set, competitor set and engine mix — all of which you chose. A SoM figure is not comparable between two companies.
What it is comparable to:
- Your own baseline. Direction and rate of change are the real signal.
- Your market position. If you hold 20% of category revenue and 4% of Share of Model, you have a visibility gap. If it is the reverse, you are punching above your weight and should protect that.
- The gap to the leader. Closing from 3× behind to 1.5× behind is meaningful progress regardless of the absolute numbers.
Set targets as relative improvements against a fixed set — "double SoM on purchase-intent prompts within two quarters" — rather than as absolute scores.
How to improve Share of Model
SoM is an output metric. You move it by fixing inputs:
- Entity clarity. If an engine cannot tell precisely what you are, it will not risk recommending you. Consistent descriptions, complete
Organizationschema, accurate third-party profiles. - Corroboration. Most of what an AI says about you was written by someone else. Third-party roundups, review platforms and community discussion move SoM more reliably than on-site content.
- Quotable content. Specific claims, clean structure, answer-first paragraphs. See how to optimize your content for Gemini and ChatGPT recommendations.
- Retrievability. If AI crawlers cannot render your pages, your best content is invisible. Fetch your own page with
curland check. - Accuracy maintenance. Stale pricing and discontinued features get repeated to prospects until the sources are corrected.
The full factor list is in 7 key factors that influence your brand's visibility in LLMs, and the broader discipline in our complete guide to Generative Engine Optimization.
Common mistakes
Measuring once. One screenshot of one favourable answer is a coincidence, not a benchmark.
Branded prompts in the set. They inflate SoM and measure recognition, not recommendation.
Changing the prompt set between periods. Then you are measuring your prompt set, not your visibility. Version it, and when you must add prompts, report the old and new sets in parallel for one cycle.
Reporting a blended figure only. Per-engine breakdowns are where the actionable findings live.
Treating SoM as a vanity score. The number is not the point. The competitive gap by intent is the point, because it converts directly into a task list.
Comparing your SoM to another company's published SoM. Different prompt sets, different competitors, different engines. The comparison is meaningless.
Frequently asked questions
How is Share of Model different from Share of Voice?
Share of Voice measures brand presence across the open web — coverage, mentions, ad impressions. Share of Model measures presence inside AI-generated answers specifically. A brand can have healthy Share of Voice and near-zero Share of Model if its coverage is not the kind engines retrieve and repeat.
How often should I measure it?
Monthly for most teams. Weekly measurement mostly captures run-to-run variance rather than real movement. Re-run off-schedule after a major launch, a rebrand, a significant piece of coverage, or a competitor's funding announcement.
How many prompts do I need?
Thirty is a workable minimum for a single category; 50–100 gives a stable read across multiple intents and product lines. Below 30, a single volatile prompt can swing the number by several points.
Can I measure Share of Model manually?
Yes, and you should do it once — it teaches you what the data looks like. Ten prompts, three engines, three runs each is 90 queries and about an afternoon. It stops scaling the moment you want repeated monthly sampling across a full prompt set, which is where automation earns its cost.
Does SoM correlate with revenue?
Not directly or immediately, and be wary of anyone claiming a fixed relationship. What it does predict is inclusion in consideration sets. The practical way to connect the two is a free-text "How did you first hear about us?" field on your demo form — "ChatGPT recommended you" is now a real answer, and no analytics platform will report it for you. For the wider commercial picture, see GEO vs SEO: how search is changing in the AI era.
Measure your Share of Model — free
You can build all of this in a spreadsheet, and for a first read you probably should. It stops being practical the moment you want repeated sampling across several engines, month after month, with sentiment and citation data attached.
Geomyze automates the loop: it generates intent-based prompts for your brand across purchase, discovery and comparison intents, runs them against the major AI engines with repeated sampling, and reports your Share of Model alongside the competitors appearing in your place.
Run your first AI visibility report free → — one full report, no card required. You will get your baseline Share of Model, a competitive gap breakdown by intent, and a prioritised list of what to fix first.
