How VisibleAI measures AI visibility

What a scan is

A scan asks AI assistants a fixed set of questions about your category, several times each, and records every answer. We then check each answer for your business and the other brands it names.

An answer counts only when the assistant responded successfully and we could read the result. Errors, refusals and empty responses are left out. They are never scored as your business being absent.

The questions we ask

Questions are written as a customer would type them and filled in with the category you enter and, for local businesses, your location. Examples include “What is the best [category] in [city]?”, “I need a [category]. Give me 5 options.” and “Is [your business] any good?”.

• Two question sets. Local businesses with a location get a local set: a core of questions every local business receives, plus a few chosen for the type of business (for example restaurant, service or retail). Other businesses get a set written for online and service companies.
• Question types. Category discovery, shortlists, comparisons, price, trust and reviews, and your own reputation. The free scan uses 5 of them.
• Only fair questions are scored. A question that depends on something you may not offer, such as public pricing, is held out until it is confirmed rather than counted against you.
• Alternatives questions never count. When we ask for alternatives to your business, the brands named are shown as market intelligence only and never affect your score.
• Scoped to your category. Your grade is for the exact category you typed. A broader or narrower wording can give a different result.

Why each question is asked 3 times

AI assistants do not give the same answer every time. Asking each question 3 times on each assistant shows whether a mention is reliable or a one-off. How often repeats agree feeds the consistency part of the score, and reports say whether the result was stable, variable or volatile across repeats.

API results vs the consumer apps

We do not use the consumer apps. We send each question by API to the models behind ChatGPT and Gemini (through an AI gateway) and to Perplexity's own API, so every scan asks the same questions the same way. Every finding comes from these generated questions and the API answers, not from real people's conversations. The consumer apps can answer differently: they may search the web, use your location or remember past chats.

Free scans use ChatGPT only. We do not test Copilot: Microsoft has no public consumer Copilot API, and Azure OpenAI runs the same model family as ChatGPT, so it would add no new signal.

Sources and citations. Perplexity's API returns the sources behind each answer, and we record them. ChatGPT and Gemini, as we query them, return no source list, so we only record a link when the assistant writes one into its answer. We never invent, infer or back-fill a source. Assistants that return no sources are left out of citation measures rather than counted as a failure.

How the score is built

Your visibility score runs from 0 to 100 and is a weighted average of five parts, each calculated only from the questions that are scored:

• If a part has no usable evidence, it is not scored as zero. Its weight is shared among the parts that could be measured.
• Website readiness, citations, answer accuracy and how assistants identify your business are reported as separate diagnostics. They explain your result but never change the score.
• Full Reports also break the result down per assistant.

Grade bands

Every score maps to a letter grade using fixed bands:

• A80–100
• B60–79
• C40–59
• D20–39
• F0–19

The bands are the same for every business, category and plan.

The likely range

Next to your score we show a likely range, for example “likely range 34–50”. It is your score plus and minus a margin that depends on how much usable evidence the scan collected, capped at 0 and 100. It describes uncertainty, not a guarantee.

• High confidence, ±4: at least 30 usable answers from 3 assistants across 8 or more scored questions, with nearly every part of the score measurable.
• Medium confidence, ±8: at least 12 usable answers from 2 or more assistants across 4 or more scored questions, with most of the score measurable.
• Low confidence, ±15: anything less. A free scan uses one assistant, so its range is ±15.

When two scans are compared, a change smaller than the wider of their two margins is shown as stable, never as an improvement or a decline.

Methodology versions

Every scan records the version of the question set and sampling design it used (for example launch_v2_20q_3r: 20 questions, 3 repeats; local scans record launch_v3_local_intent plus the business type), along with the score version (visibility-score-v2.1) and grade bands (grade-bands-v1.0). A re-scan reuses the question set of the scan it repeats.

We only show a change between two scans when they measured the same thing: the same business or domain, the same methodology and score version, the same number of repeats and assistants, and at least 60% of scored questions in common. Otherwise the report says the scans are not comparable instead of printing a misleading change. (Trend rules: scan-trends-v1.0.)

Scans made before our current scoring was introduced keep the score they were given and are never re-scored.

What we don't claim

• No guarantees. Nobody can buy a place in AI answers, and we cannot promise you will appear. We show what is within your control.
• Not the app you use. API results can differ from what you see in the ChatGPT, Gemini or Perplexity apps.
• No invented sources. Citations come only from sources an assistant returned or links written in its answer.
• No certainty about why. When we suggest why an assistant behaved differently, it is labelled as a possible, unverified explanation.
• Small moves are noise. A change inside the likely range is not progress or a setback.