# How MentionX Turns AI Answers Into Evidence: Methodology and Findings

> The method, dataset and first findings from 6,813 AI answers and 36,944 citations: how stable AI visibility is, what engines cite and when AI is negative.

Source: https://mentionx.ai/learn/research/ai-answer-methodology-and-dataset  
Author: Dipuck Jones, founder of MentionX  
Updated: 6 November 2026  
Level: Practitioner

**MentionX sends defined prompts to ten AI engines, stores every completed answer, and calculates each KPI from the structured record of that answer.** This page publishes the method, the dataset and the first findings from 6,813 AI engine answers: how stable AI visibility really is, what engines cite, and what they say when they are negative.

**Key findings**

- **0.9%** of citations on category questions came from the brand's own website. Engines rely on what others publish.
- **25 points** is how far a brand's mention rate moved between runs of identical prompts on the same engines.
- **33.1%** of statements AI engines made about brands were negative, and 45.1% of those were about fit or price.
- **0.05%** was a young brand's mention rate on category questions, while its blended rate looked steady at 33–36%.

## How a scan works

AI engines do not publish what they say about a brand, and they do not say the same thing twice. A MentionX scan treats each answer as a piece of evidence: it asks the same questions a buyer would ask, records exactly what came back, and turns those records into numbers that can be checked.

1. **Prompts:** buyer-style questions are generated from the brand's website and market, then fixed as a set.
2. **Engines:** every prompt is sent to each selected engine, and each completed answer is stored in full.
3. **Reading:** each answer is read for brand mentions, ranked shortlists, cited sources and statements about the brand.
4. **Scoring:** KPIs are calculated per scan from those records, with branded and category prompts kept apart.

## The dataset

The research dataset holds 6,813 completed AI engine answers to 2,311 prompts, with 36,944 citations captured from those answers. Every answer is stored with its engine, its prompt and its date, so any figure on this page can be traced back to the answers that produced it.

The answers span ten engines and answer surfaces: ChatGPT, Gemini, Google AI Mode, Google AI Overviews, Perplexity, Microsoft Copilot, Claude, Grok, DeepSeek and Meta AI. The brands scanned cover B2B software categories such as CRM, workflow software, real estate technology and AI visibility tools.

The dataset is repetitive by design. A large share of scans are repeat runs of identical prompts, which is what lets us measure how much AI answers change from one run to the next. Failed runs are excluded from every count, and the prompt count includes scans that did not complete.

## One visibility score, three signals

Being named is not the same as being chosen, and being chosen is not the same as being trusted. The visibility score blends all three: mention rate (M) at 40%, normalised shortlist position (P) at 35% and citation share (C) at 25%.

```
V = 0.40 × M + 0.35 × P + 0.25 × C
P = 1 − (average position − 1) ÷ lowest position in the scan
```

P is zero when the brand is never ranked. For an illustrative brand named in 30% of answers, at an average position of 3 in a scan whose lowest position is 10, holding 20% of brand-linked citations: P = 0.80 and V = 0.12 + 0.28 + 0.05 = 0.45.

A brand counts once per answer however many times it is named. Because citation share only counts citations linked to a tracked brand, it can swing widely in scans where few answers cite any tracked brand.

## Every KPI, defined

- **Mention rate:** answers that name the brand ÷ completed answers.
- **Recommendation rate:** answers that place the brand in a ranked shortlist ÷ completed answers.
- **First-choice rate:** answers where the brand is position 1 of a shortlist ÷ completed answers.
- **Average position:** mean shortlist position across the answers where the brand is ranked.
- **Share of voice:** the brand's mentions ÷ mentions of all tracked brands.
- **Citation share:** citations tied to the brand ÷ all citations tied to any tracked brand.
- **Mention-recommend gap:** mention rate minus recommendation rate, never below zero.
- **Visibility score:** the weighted blend of mention rate, position and citation share.

## Finding: a blended score can hide the real picture

We tested a young marketing software brand with the same 20 prompts across 3,384 AI answers. Its overall mention rate held steady between 33% and 36%. Split by prompt type, the picture changes: prompts that name the brand were answered with a mention 99.5% of the time and placed it in a ranked shortlist in 56.8% of answers. Category prompts that do not name it mentioned the brand in 0.05% of answers.

The blended number moved so little because it mostly measured the share of prompts that already contained the brand name. Report category questions on their own: they are where buyers discover options, and they are the honest baseline to improve against.

## Finding: the same question gets different answers

We ran a large business software brand through the same category prompts on ChatGPT and Gemini every one to three days. The category mention rate ranged from 33% to 58% across identical runs, and moved by up to 25 points between two runs two days apart. Of the prompt and engine pairs, 25% always mentioned the brand, 33% never did and 42% flipped between runs.

Broad, high-intent questions mentioned the brand every time. Narrow questions about specific jobs rarely or never did. One scan is one sample: read trends across repeated scans on the same prompt set, never from two adjacent points.

## Finding: what AI engines cite

On buyer questions that do not name the brand, engines cited 27,007 sources. The brand's own website made up 0.9% of them, about 1 in every 108 citations. 8.8% were Google grounding links that hide the real source; tech media (2.3%), social and publishing (1.3%), review sites (1.3%), video (1.2%) and community sites (1.2%) each carried more weight than the brand itself. The remaining 81.7% came from other domains such as vendor sites, niche publishers and blogs.

When a buyer asks about a brand by name, the pattern reverses and engines cite the brand's own pages heavily. Owned content explains who you are. Third-party coverage decides whether you are recommended.

## Finding: what AI says when it is negative

Across 2,455 statements AI engines made about scanned brands, 33.1% were negative. The most common types were a concern or drawback (29.5%), losing a comparison to a rival (24.8%) and fit for a narrow use case only (20.4%). Use-case fit (31.2%) and pricing and value (13.9%) account for 45.1% of all negatives, and 15.3% were doubt rather than criticism: knowledge gaps, unverified claims and hedged answers.

The negative share ran from 25.0% on Google AI Mode to 42.6% on DeepSeek, with ChatGPT at 37.5%, Perplexity at 33.8%, Gemini at 32.1% and Microsoft Copilot at 29.8%. A brand that only reads one engine sees only part of its reputation.

## When two scans can be compared

Prompts are tailored to each brand and stored as a fixed set. We compare scans on the same prompt set and the same list of engines, and category questions only with category questions. We never mix scans built on different prompt sets, scans with different engine lists, or branded and category questions in one number.

## What the method does not claim

- **Answers vary by user:** AI answers can differ by user, location and time. Results describe these scans, not every person's experience.
- **Generated prompts:** prompts are written by MentionX's prompt generator, not sampled from real buyer searches.
- **Scope:** results describe the scans in this dataset, not the whole market.
- **No causality:** nothing here shows that a specific action changed a score, or explains why answers vary.
- **Automated reading:** mentions, rankings, citations and statements are read from answer text by a language model.

The general method, including how ranges are calculated, is on the [methodology page](https://mentionx.ai/learn/methodology). Standards for every statistic on Learn are in the [editorial policy](https://mentionx.ai/learn/editorial-policy).

## Frequently asked questions

### How is the MentionX visibility score calculated?

It weighs mention rate at 40%, normalised shortlist position at 35% and citation share at 25%. Position counts as zero when the brand is never ranked.

### What counts as a mention?

An answer that names the brand. A brand counts once per answer, however many times it is named.

### Why do AI answers change between runs?

Engines generate each answer fresh and may retrieve different sources each time. In our data, the same prompts on the same engines moved a brand's mention rate by up to 25 points between runs.

### Why report category questions separately?

Questions that name a brand are answered with a mention almost every time. Mixing them in makes a score look stable while hiding how often a brand is found when buyers do not already know it.

### When are two scans comparable?

Only when they used the same prompt set and the same list of engines.

### Do MentionX results show what every user sees?

No. Answers can differ by user, location and time, so results describe the scans in this dataset, not every person's experience.

---

Written and reviewed by Dipuck Jones for MentionX Learn. The HTML version of this page: https://mentionx.ai/learn/research/ai-answer-methodology-and-dataset