How MentionX Measures AI Visibility: Methodology

MentionX measures AI visibility by running defined sets of prompts across AI engines, repeating runs to account for variation, and scoring each answer for whether a brand is mentioned, cited or recommended. Results are reported as rates with ranges, alongside the sample size and dates. This page explains the method behind every MentionX statistic on Learn, and its limits.

What is measured?

How are prompts chosen?

Prompt sets are built around a topic and the stages of a buying journey: category, comparison, problem-led and brand questions. Prompts are written the way real users ask, and leading prompts are avoided. For published research, the prompt set, or a representative sample, is listed with the results.

Which engines are covered?

Engines can behave differently depending on whether they're used through their consumer interface or an API, and on the model version. Each study states which was used.

How are answers scored?

A mention counts when the brand name, or a clear variant, appears. A citation counts when a brand-owned URL is shown as a source. A recommendation counts when the answer suggests the brand as a choice for the user's need, not just in a list of options. Ambiguous cases are flagged for review.

Why use repeat runs?

The same prompt can return different answers each time. Running prompts more than once shows how stable a result is. A brand mentioned in 3 of 3 runs is more reliably visible than one mentioned in 1 of 3.

How are ranges calculated?

Rates are proportions, so we report confidence intervals using the Wilson score interval, a standard method for binomial proportions described in the NIST Engineering Statistics Handbook. Small samples produce wide intervals; we show them rather than hide them.

What are the limits?

• Personalisation: answers can vary by location, sign-in status, history and settings. Scans use consistent conditions and don't represent every user.
• Change over time: models and retrieval systems are updated, sometimes without notice. Each result is dated.
• Sample, not census: a prompt set samples possible questions; it doesn't cover every way people ask.
• No search volume for prompts: AI platforms don't publish prompt volumes, so prompts aren't weighted by popularity unless stated.
• Correlation, not causation: visibility changes after an action may have other causes.

How is published research reported?

Each MentionX study states:

• [ ] Topic, prompt count and prompt examples
• [ ] Engines, access method and dates
• [ ] Run count and location
• [ ] Metrics, with intervals
• [ ] Known limitations

How does this relate to our editorial standards?

All statistics on Learn follow our editorial policy. Numbers not produced with this method are cited to their original source or labelled illustrative.