How we score

Your AI Visibility Score is a 0–100 number built from real AI engine answers to the questions your buyers actually ask. No panel of experts, no self-reported survey — we ask the engines, read what comes back, and score it. Here is every step, including the parts most tools keep vague.

The questions: 11 per audit

Every audit asks two families of questions — because "what does AI say about you?" and "does AI bring you up at all?" are different measurements.

5 branded questions — they name your brand and measure perception:

• Tell me about {brand}. What do they do?
• Is {brand} a good and reputable choice? What are the pros and cons?
• What are the best alternatives to {brand}?
• Who are {brand}'s main competitors?
• Would you recommend {brand}? Why or why not?

6 category questions — they name only your market category, never your brand, and measure visibility: does the engine surface you unprompted? Three are the same for everyone:

• What are the best {category} options?
• Recommend a {category} — who should I consider?
• Who are the top {category} providers?

The other three are tuned to the kind of business you run (software, local service, online store, professional services, or a general set when none fits), so a dentist and a SaaS tool aren't asked the same thing.

On a subscription, your own tracked questions run on the same engines — your topics and your buyers' phrasing, not just our templates. They are shown separately and do not change your score.

The engines: 4 live, up to three runs each

Prompts run against the live engines — ChatGPT, Perplexity, Gemini, and Claude. Each of them answers every prompt up to three separate times. That's up to 132 real AI answers behind a single audit.

Majority vote, biased against inflation

When a prompt runs multiple times, the runs vote: your brand counts as mentioned only if a strict majority of real answers mention it. Ties go to "not mentioned" — when in doubt, we under-count you rather than flatter you. A failed API call is never treated as an answer and can never outvote a run that actually responded.

Hallucination guards

• Deflection guard (English + Russian): "I'm not familiar with this brand" never counts as a mention — even when the answer politely echoes your brand name back.
• Grounding sentinel: branded prompts explicitly instruct the engine to say it has no grounded information instead of confidently making things up. When an engine can't ground your brand, that answer scores as unknown — not as a win.
• No-visibility cap: if visibility can't be measured at all, the score is hard-capped at 30. An engine that has never heard of you cannot accidentally produce a 90.

The formula: 70/30

score = 70% visibility + 30% perception. Visibility — how often you surface on the category questions, where the engine wasn't prompted with your name. Perception — how often the engines actually describe you on the branded questions, instead of deflecting or guessing. A brand that never surfaces in category answers keeps only the perception slice — 30 points at most. The formula is applied per engine; your overall score is the average across engines.

Why not a single-shot score

Forbes is right: most AI-visibility numbers are unreliable — a single-shot score on a non-deterministic engine is mostly noise. That's why UpRadar never scores from a single shot: each prompt runs up to three times and the runs vote, so one outlier answer can't decide your score. In the weekly report, a change smaller than a single flipped answer is shown as "held steady", not as movement.

What we deliberately don't measure

• Citations beyond what engines return. Where an engine returns its sources (today: Perplexity, and ChatGPT on paid plans), UpRadar records them and shows the most-cited domains and pages in the app. Engines that return no sources give us no citation data, and citations are not a separate component of the 70/30 formula.
• Prompt-volume estimates. Some tools sell "how many people ask this per month" numbers. Those are modeled estimates, not measurements — and they say nothing about whether the engine recommends you. We score what engines answer, not how often somebody might ask.
• Revenue attribution. We show where you stand and what moved. We don't claim a score change caused a revenue change — nobody can measure that honestly from the outside.

Where the questions come from

The templates mirror the stages of a real buying conversation: discovery ("best tools"), shortlisting ("recommend one"), and diligence ("is it reputable", "alternatives", "competitors"). They are fixed per audit so scores stay comparable week to week — and on a subscription you extend them with your own.