Blog/Guide · 4 min
How to measure share of voice in AI for iGaming brands
Mention rate, share of voice, citations, position, sentiment — what each AI visibility metric tells a casino or sportsbook, and which ones hold up.
The five metrics people use
| Metric | What it counts | What it's good for |
|---|---|---|
| Mention rate | Share of answers that name your brand | The cleanest single read of visibility |
| Share of voice | Your mentions as a share of all brand mentions | How crowded the answer is, and your slice of it |
| Citations | How often the model links to your pages or domains | Whether your own content is feeding answers |
| Position | Where in the answer you appear | Rough prominence — noisy, see below |
| Sentiment | Whether the model describes you positively | Reputation signals — hard to measure well |
Mention rate and share of voice answer different questions. If a model names eight brands in every answer, you could be in all of them and still own only an eighth of the voice. If it names two, the same mention rate means half.
Get the denominator right
The question set
Everything rests on the questions. Ask "best online casino" and you measure one thing; ask "which casino pays out fastest with iDEAL" and you measure another. A defensible set:
- comes from what players in that market actually search, weighted by volume
- is written in the local language
- never names a brand, so it doesn't steer the answer
- stays fixed, so this week is comparable with last week
Questions you add yourself are useful for tracking specific topics. They shouldn't be blended into a benchmark score, because then the benchmark measures your choices.
The model panel
Players don't all use the same assistant. Measure several models and report each separately before you combine them — a brand can be strong in one model and absent in another, and an average hides that.
Repeated samples
Ask the same question to the same model twice and you can get two different shortlists. So one answer per question tells you very little. Sample each question several times and the rate settles into something you can put a range around.
A worked example
(Illustrative numbers.) You run forty questions through six models, five samples each: twelve hundred answers. Your brand appears in three hundred of them, a twenty-five per cent mention rate. Across those twelve hundred answers, models name brands a total of four thousand times; you account for three hundred, so your share of voice is seven and a half per cent.
Now a rival shows a twenty-three per cent mention rate. With twelve hundred answers, the ranges around twenty-five and twenty-three overlap. You aren't reliably ahead, and you shouldn't report that you are.
Metrics to treat with caution
Position in the answer
"Named first" sounds valuable, and sometimes is. But models format answers differently — numbered lists, paragraphs, tables — and the order can shift between runs of the same question. Treat position as colour, not as a headline KPI, unless it's sampled heavily and shown with its spread.
Sentiment
Sentiment scoring means using one model to grade the tone of another model's answer. That's two layers of judgement before a number appears. Gambling answers often include responsible-gambling caveats, which a naive scorer reads as negative. If you use sentiment, read the underlying answers before you act on the score.
How miraindex.ai measures it
miraindex.ai measures mention rate per brand per market. Each market has a fixed, demand-weighted question set in the local language. Every question goes to six models — ChatGPT, Claude, Gemini, Perplexity, Grok and DeepSeek — five times each, weekly. Every figure carries a Wilson ninety-five per cent interval, and brands whose intervals overlap are not shown as ranked.
Custom questions are supported but never enter the score. And because every brand in a market is scored from the same shared scan, you see your competitors on exactly the terms you see yourself.
It does not report sentiment or answer position. Full detail is on the methodology page.
A checklist
- Fixed question set, from real demand, in the local language
- No brand names in the questions
- Several models, reported separately first
- Several samples per question per model
- Confidence ranges on every number
- No ranking where ranges overlap
- Custom questions kept out of the benchmark
See your own number
Run the free visibility scan for your brand and market.