AI Visibility Score: How It's Calculated

An AI visibility score combines how often you're mentioned, cited and recommended across a fixed set of prompts and engines, adjusted for position and sentiment, into one number, usually out of 100. There's no industry standard: every tool calculates it differently. A useful score is transparent, meaning you can see the prompts, the runs, the weights and the formula, and reproducible from the raw answers.

Key takeaways

• The inputs are mention rate, citation rate, recommendation rate, position and sentiment.
• Weights are a choice; publish them.
• A score hides detail. Always show the component rates beside it.

What goes into an AI visibility score?

All are calculated from the same log: one row per prompt × engine × run.

How is the score calculated, step by step?

Here is the formula used in MentionX's Learn materials. It's one reasonable approach, shown so you can copy, change or challenge it.

1. Compute the rates for each engine: M, C, R from yes/no columns.
2. Compute position: for each mention, score 1 / position (first = 1.0, second = 0.5, third = 0.33...). Average over mentions to get P.
3. Compute sentiment: S = mentions that are neutral/positive and accurate ÷ all mentions.
4. Combine per engine:

Engine score = 100 × (0.35·M + 0.20·C + 0.30·R + 0.15·P) × S

1. Combine across engines with a simple average, or weights you justify (see below).

The weights (0.35, 0.20, 0.30, 0.15) reflect a judgment that being named and recommended influences buyers more than a link. They're not derived from data. Change them if your business disagrees, then keep them fixed.

Worked example

Illustrative numbers only. On Perplexity, across 90 answers (30 prompts × 3 runs):

Score = 100 × (0.35×0.50 + 0.20×0.30 + 0.30×0.20 + 0.15×0.45) × 0.89 = 100 × (0.175 + 0.06 + 0.06 + 0.0675) × 0.89 = 100 × 0.3625 × 0.89 ≈ 32

Should engines be weighted?

It's tempting to weight engines by "market share", but reliable, current usage figures by engine and by your audience are hard to get. Options:

• Equal weights (simplest, easy to explain).
• Weights from your own data, such as AI referral sessions by source in GA4 (guide), labelled as a proxy.
• Weights from customer surveys asking which assistants buyers use.

Whatever you choose, state it in the report. Don't import usage stats from a vendor blog without checking their source.

What are the limits of a single score?

• It hides trade-offs. Your score can stay flat while citations rise and recommendations fall.
• It depends on the prompt set. Add easy prompts and the score rises without any real change.
• It varies run to run. Small samples produce noisy scores; report ranges or include run counts.
• It doesn't measure revenue. Pair it with pipeline or sign-ups from AI referrals.

What should you ask any vendor about their score?

1. What prompts are used, and can I see and edit them?
2. How many runs per prompt per engine?
3. Which engines and modes (e.g. ChatGPT with search on or off)?
4. What's the exact formula and weights?
5. How is sentiment judged, by a model or a person, and how is it checked?
6. Can I export the raw answers?
7. How are brand-name variants matched?

If a vendor can't answer these, treat the score as a marketing number.

How do you report it responsibly?

• Show the score, the component rates and the share of voice together.
• Include a method box: prompts, engines, runs, dates, formula version.
• Show the trend over at least three months before drawing conclusions.
• Note any change to prompts or method on the chart.

Can you calculate it in a spreadsheet?

Yes. With one row per prompt × engine × run and columns mentioned, cited, recommended (1/0), position (blank if not mentioned) and accurate_nonnegative (1/0):

Add an engine column and use a pivot table or FILTER to calculate per engine. Keep the formula version in a cell at the top of the sheet so old reports stay explainable.

How do you compare scores over time?

Only compare scores calculated with the same prompt set version, engines, run count and formula. When you change any of these, start a new series or recalculate the old months from the raw log. A score that silently changes method is worse than no score.

How do you raise the score?

Work on the components, not the number:

• M low: off-page coverage and brand clarity (AI visibility).
• C low: answer-first pages and crawler access (robots.txt for AI).
• R low: clearer positioning on who you're for.
• S low: fix outdated facts across your site and listings.

Sources

• Aggarwal et al., GEO: Generative Engine Optimization, which proposes visibility metrics such as position-adjusted word count for generated answers
• Google Search Central, AI features and your website