How we measure AI visibility

The exact prompts, engines and formula behind every scorecard — nothing hidden, no black box. See it applied in a real example report.

The 12 buyer-intent prompts

Every check asks the same battery, with your category, brand, audience and region filled in (defaults: audience "everyday use", region "Europe"). Verbatim answers appear in your report.

Discovery	`What are the best places to buy {category} online?`
Discovery	`Recommend some good online stores for {category}.`
Discovery	`I want to buy {category}. Which brands or shops should I look at?`
Discovery	`Best {category} in 2026 — what do you recommend?`
Evaluation	`Is {brand} a good place to buy {category}?`
Evaluation	`What do people say about {brand}?`
Comparison	`Compare the top online shops for {category}.`
Comparison	`{brand} vs alternatives for {category} — which is better?`
Audience fit	`Best {category} for {audience}?`
Audience fit	`Affordable {category} with good quality — where should I buy?`
Gifts & logistics	`I need {category} as a gift. Where should I order from?`
Gifts & logistics	`Where can I get {category} delivered to {region}?`

Engines & models

We ask each prompt once per engine, through the same public APIs the assistants run on, with no system prompt or steering — the answer is what the model gives an ordinary user. Every answer in your report carries the exact model version and a UTC timestamp.

ChatGPT — OpenAI API (gpt-5-mini class)
Gemini — Google AI API (Gemini Flash class)
Claude and Perplexity — added to checks as engine coverage rolls out; your report lists exactly which engines were used

The score

score = 100 × ( 55% × mention rate + 25% × avg prominence + 20% × share of voice )

Mention rate — share of all answers that name your brand or domain.
Prominence — how early you appear in an answer: first-third mentions score near 1.0, later mentions scale down. Averaged over answers that mention you.
Share of voice — your mention rate divided by yours + your competitors' combined, so the score reflects who the AI actually recommends, not just whether you exist.

We also record sentiment around your mentions (cue-word window, marked positive/neutral/negative), which answers cite your own domain as a source, and every "gap prompt" where a competitor was named and you weren't.

Honest limitations

Answers vary run to run. LLMs are not deterministic; the same prompt can name you today and skip you tomorrow. Scores are directional, not decimals to obsess over.
We sample. 12 prompts per engine, sampled weekly — not continuous daily tracking like enterprise tools. A snapshot of buyer intent, not the universe of questions; the deep audit widens the battery.
Sentiment is heuristic. Cue-word matching near your mention — right in the clear cases, crude in subtle ones. The verbatim answer is always in the report so you can judge yourself.
Model versions move. Engines ship new models without notice; that's why every answer is stamped with the model that produced it, and why monitoring re-runs matter.