How is an AI visibility score calculated, and what does it mean?

Every tool follows the same basic pipeline, even when the labels differ.

1. Pick the prompts. The tool chooses a set of buyer questions, written by you or generated from your site.
2. Ask the platforms. It sends those prompts to engines such as ChatGPT, Perplexity, Claude, Gemini or Google AI Overviews.
3. Score each answer. It checks each answer on a few components.
4. Combine. It weights the components and averages across answers and platforms.

The usual components are:

• Mention: is your brand named at all?
• Citation: does the answer link to your site? (See how citation rate is calculated.)
• Accuracy: are the facts about you right, such as price, product names and location?
• Sentiment: is the description favourable, neutral or negative?

The weights are the vendor's judgement. No study shows that mentions should count twice as much as accuracy, or the reverse. That is why AI visibility tools disagree about the same brand. They ask different prompts and they combine the parts differently. The definition of an AI visibility score covers what each component hides.

Accuracy is the component most often built wrong. The tempting formula starts at 100 and subtracts points for each error. It has no denominator, so it punishes audit size instead of error rate.

Here is a hypothetical. A tool subtracts 15 points per wrong fact.

Both audits are equally accurate. The larger one scores zero. A customer who tracks more prompts watches the score collapse for buying more coverage.

The research on factual accuracy settled this long ago. FActScore (Min et al., 2023) divides supported facts by the total facts in a response. RAGAS faithfulness divides supported claims by total claims. A sound accuracy score is a rate.

A rate alone has its own flaw. The SAFE paper (Wei et al., 2024) added a recall term because precision alone does not penalise systems that say too little. The same logic applies here: in a pure average, one serious error can fade to nothing across 400 answers. Good formulas add a term for the worst single error, so a wrong acquisition claim stays visible. Two more rules matter. Count only answers that actually came back, not answers requested. Do not report accuracy for a platform with only a handful of answers.

AI platforms do not always search the web. As of October 2026, retrieval is decided per query on every major platform.

• OpenAI says ChatGPT search rewrites your question into one or more targeted queries, and the model decides when to search.
• Claude.ai web search is a user toggle with no published default.
• Perplexity's Agent API requires developers to turn web search on.
• Google says AI Overviews and AI Mode may use different models and techniques, and AI Overviews often do not trigger.

When an engine does not search, it answers from training data frozen at a cutoff. Your site changes cannot reach that answer this quarter. When it does search, access and indexing decide whether your pages are even candidates.

A single score averages both kinds of answer, usually without knowing the mix. A low score can mean your pages are blocked. It can also mean the model answered from old training data. Those need different fixes. This is also why a zero score does not mean AI doesn't know your brand.

Some things no tool can measure from the outside.

• Rank inside an answer. Google's generative AI report shows impressions only, with no position data to check a rank against.
• Influence. OpenAI's API returns more sources consulted than sources cited, so a citation count misses pages that shaped the answer.
• The real queries. Engines rewrite and fan out prompts. Your tracked prompt is a proxy for an input you cannot see.
• Full volume. In OtterlyAI's 2026 experiment, Microsoft's own report counted 42,817 citations for one site while Otterly's prompt tracker logged 3,934. A prompt set is a sample.

Citation units also differ by platform. Anthropic cites text spans while Perplexity cites whole sources, so averaging them mixes unlike things.

Use the score as a trend line inside one tool, then go straight to the parts.

• Ask how accuracy is normalised. If the vendor cannot name a denominator, bigger audits will look worse.
• Check the formula version. A score from an old formula is not comparable with a new one.
• Read each platform separately. In OtterlyAI's experiment, Microsoft Copilot produced 2,718 of one site's 3,934 tracked citations while four other engines barely cited it. A blended number would have hidden that.
• Read the wrong facts, not the total. A specific wrong price is something you can fix on your own pages.
• Check access before content. Confirm your robots.txt and firewall let each AI search crawler in, and that key facts appear in the HTML before JavaScript runs.

For a quick snapshot, the free checker on this site scores up to three buyer questions. When you report upward, explain the score's limits to stakeholders before they read it as a ranking.

Which tools give an AI visibility score? Many AI visibility trackers and audit platforms publish one, and the free checker on this site gives a snapshot score. The names matter less than the method. Ask each tool which prompts it runs, which platforms it queries, how it weights mention, citation, accuracy and sentiment, and whether accuracy is a rate. Two tools with different answers will give the same brand different scores.

Is a higher AI visibility score always better? Only within the same tool, formula and prompt set. A higher number from a different tool, or from a changed prompt list, tells you nothing about progress. A score can also rise while a serious factual error persists, if the formula averages it away. Read the per-platform scores and the list of wrong facts alongside the headline number.

Can I compare my score with a competitor's? Only if both were scored with the same prompts, platforms and formula at about the same time. Even then, brand size dominates who gets cited, so a large incumbent will usually outscore a smaller brand for reasons no page change addresses. Compare like-sized brands, and treat any leaderboard that mixes very different brands with caution.

Why did my score change when I changed nothing? Engines update models, change when they search, and return different answers to the same prompt. The tool may also have changed its prompts or formula. On the access side, a CDN or firewall can start blocking AI crawlers without any edit on your part. Check the formula version and your crawler access before reading a change as real.