Methodology

How Variask measures AI visibility

AI engines give a slightly different answer every time they are asked. This page explains exactly how Variask turns many answers into numbers you can trust, and how it tells a real change from normal noise.

Last updated November 5, 2026

1. The questions we ask

Every brand tracks its own fixed set of prompts: the questions its buyers ask AI, such as best, versus, alternatives, pricing and how to questions. Keeping the set fixed means changes in the numbers come from the answers, not from changing questions.

Prompts can be written by the team, picked from industry templates, or suggested from search demand. Each prompt can carry variables, such as the current year, that update on every run.

2. The engines and markets

Each prompt is asked on each AI engine the plan includes:

  • ChatGPT
  • ChatGPT app
  • Gemini
  • Claude
  • Perplexity
  • Grok
  • DeepSeek
  • Microsoft Copilot
  • Google AI Overviews
  • Google AI Mode

Prompts run as a person in a chosen country, city and language would ask them. Where an engine searches the web, the search carries the same market, so answers reflect what local buyers see.

3. Repeat answers

Because answers vary, a prompt can be asked more than once on the same engine. Every answer is kept. Visibility is worked out from all of them, so one unusual answer does not swing the score.

4. What we measure in each answer

Mention
Whether the answer names the brand, using its name and any aliases the team adds.
Position
Where the brand appears among the brands named in the answer: first, second and so on.
Share of voice
The brand's share of all brand mentions, against the competitors being tracked.
Sentiment
How the answer talks about the brand, scored from 0 (strongly negative) to 100 (strongly recommended), with 50 as neutral.
Citations
The sources the answer cites or links to, grouped by domain and page, and by type such as reviews, forums or news.
Accuracy
Whether the answer states the brand's facts correctly, checked against the facts the team sets.

5. Likely ranges

Visibility is the share of answers that name the brand. Because it comes from a sample of answers, it is an estimate, so every score carries a 95% likely range.

The range is worked out prompt by prompt. Each prompt is treated as its own group, and prompts with only a few answers borrow strength from the overall rate, so a single prompt cannot dominate the range.

6. Real change or noise

When a score moves between two periods, Variask compares the prompts asked in both periods, one by one, so adding new prompts cannot fake a change. Each move gets a verdict:

  • Real gain or drop: the move is larger than answer to answer variation at 95% confidence.
  • Likely gain or drop: the move is probably real, at 90% confidence.
  • Within normal noise: the move is the kind of variation you would expect from asking again.
  • Too few answers to judge: fewer than 12 answers on either side.

7. The AI Index

The public AI Index and citation research use the same measurement on industry question sets that Variask runs and pays for itself. Customer prompts and customer data are never used for them.

8. Limits

  • AI answers depend on the wording, market and moment, so no tool can see every answer every person gets.
  • Engines change their models and sources over time, which can move numbers for everyone at once.
  • Visibility measures how often AI names a brand for the tracked prompts, not across every possible question.