For decision-makers
One value per period, with interval and change: it reads well in a meeting and stands up to questions.
AI Trust Score
One number. Six reasons.
Six signals observed in AI assistants’ answers, in one value from 0 to 100: how often they name you, recommend you, cite your sources, speak well of you, prefer you to competitors and stay consistent. Code calculates it, never a model.
Why a number
Six measures. One shared reference.
Before choosing, a customer can ask ChatGPT, Gemini, Perplexity, Copilot or Claude and read a ready-made answer. Mentions, recommendations, sources, tone and competitors move together: following them one by one is slow, presenting them to decision-makers is hard.
The score gives the team a shared reference, with a published method and the details always one click away. It does not replace the measures: it orders them and shows which signal to work on first.
For marketing
The dimensions that weigh most and those that hold back, with the answers behind them.
For content writers
The cited sources and the statements to clarify, linked to Actions.
The six dimensions
Six signals. Six declared weights.
Each dimension is a ratio of observed counts, from 0 to 100, with its own sample. The weights add up to 100 and change only with a new, published version of the method.
Authority in version 1 measures the presence of your sources among those cited; the quality of the source type will come with a later version. Preference compares mentions in pairs: 50 means parity, whatever the number of competitors.
| Dimension and weight | What it observes | On what base |
|---|---|---|
| Visibility · 25% | How often answers name you. | Complete, reviewed answers, averaged by question and assistant. |
| Recommendations · 20% | How often they recommend you for the need in the question. | The same answers as Visibility. |
| Authority · 15% | How much the answers that show sources cite your pages. | Only answers in which the citation can be assessed. |
| Reputation · 15% | The tone of the answers when they name you. | Positive, neutral (counted as half) and negative mentions. |
| Preference · 15% | Where you stand against each declared competitor. | Pairwise comparisons in the same answers. |
| Consistency · 10% | How stable your presence stays across assistants and runs. | Only where you appear: absence is already in Visibility. |
How it is calculated
No number comes from a model.
The language model reads an answer and classifies it: does it name the brand, recommend it, cite one of its pages, with what tone, next to which competitors. A deterministic judge checks every classification against the text and a person reviews it. The score is calculated by code, from those counts only.
The total is the sum of weight times value for the measured dimensions, divided by the sum of their weights. Same data, same result, on every machine.
Panel questions
The same questions, to the same assistants, in every cycle.
Verified classification
An extractor classifies, a deterministic judge checks, a person reviews.
Six ratios
Each dimension with its numerator, its denominator and its weight.
Total with interval
A weighted average, always shown with coverage and a 95% interval.
Sample and coverage
Below the threshold, no invented zero.
A dimension is measured only with enough observations. Below the threshold it stays excluded and declared, with the reason: it is not worth zero and does not quietly lower the total.
At least 30 observations
For each dimension, plus a minimum of questions or competitors: a single case does not decide a value.
Declared coverage
The sum of the measured weights. The total appears only if Visibility is measured and coverage reaches at least 70%.
The weight of what is missing
Next to the value, the range the total would have if the excluded dimensions were worth 0 or 100.
Uncertainty
Every value says how much it can move.
A score without an interval promises a precision the sample does not have. That is why the 95% credible interval always appears next to the value, with the declared precision.
The interval comes from 10,000 Monte Carlo draws with a fixed seed, resampling questions and assistants: the same panel always gives the same interval. If the half-width exceeds 15 points, the value is declared “beyond the precision threshold”.
The same sample value with two different samples: the narrow interval is compared, the wide one is declared as beyond the 15-point threshold set by the method.
| Made-up example | Value and 95% interval | How to read it |
|---|---|---|
| Large panel | 62, between 58 and 66: half-width of 4 points. | Precision within the threshold: the value enters comparisons. |
| Reduced panel | 62, between 44 and 80: half-width of 18 points. | Beyond the 15-point threshold: the value is published and declared as such. |
| Same panel, again | The same interval, at every new calculation. | The fixed seed makes the Monte Carlo draws repeatable. |
Comparisons over time
Compared only under the same method.
Every score carries the fingerprint of the method that produced it: version, weights, thresholds, statistics. Two values line up only with the same fingerprint; a method change breaks the series instead of inventing a change.
A difference, not a cause
The comparison says how much the score changed. Attributing an effect requires experiments with a control group.
By assistant
The same calculation on a single assistant’s answers: smaller bases, wider intervals, declared.
Next to competitors
Competitors’ six dimensions on the same questions, where the data allow.
From number to action
What lifts it. What holds it back.
The dimensions that contribute most and those that weigh least are read in points, with their base. From there you open the answer behind them and prepare the right action: a page to clarify, a source to win, an FAQ to write.
A sample reading for a fictional brand: the dimension that supports the score, the one that holds it back, the answer to open and the action that follows.
Score factors · Northstar Audio, fictional brand
- Supports it
- Visibility: the brand is named in most valid answers.
- Holds it back
- Authority: among the cited sources, the brand’s pages are rare.
- Answer to open
- Those that cite a magazine and a forum, but not the official product page.
- Action
- Make the spec sheet complete and citable, then repeat the same questions.
What it is not
An indicator, not a verdict.
It is a proprietary indicator with a published method. It summarises what the observed answers say about you, over one period, for one market and one language.
It does not read models’ minds
It observes answers; nobody measures what a model “thinks”.
It is not a verdict on the company
Nor a sales forecast, nor a rating of its products.
It is not a generated number
No value comes from a language model: only reviewed counts.
Frequently asked questions
Questions about the AI Trust Score.
Why the AI Trust Score uses six dimensions, what it really measures, why a dimension can be missing and how it differs from the Visibility Index.
Why six dimensions instead of a single indicator?
Because being named, being recommended, being cited as a source, tone, comparison with competitors and stability change for different reasons. The value from 0 to 100 summarises them, and each one stays readable with its own base.
Does it measure what a model thinks of my brand?
No. Nobody can measure what a model “thinks”. The score summarises signals observable in answers: mentions, recommendations, sources, tone, comparisons and stability.
Can the score rise with certainty?
No. You can work on content and sources and measure what changes, with the same method. Nobody can promise an increase.
Why can a dimension be missing?
Because it did not reach the minimum sample. It stays excluded and declared; if coverage drops below 70% of the weights, the total is not calculated.
How is it different from the Visibility index?
The Visibility index summarises presence across five channels: AI, search, social, press and reputation. The AI Trust Score looks only at AI answers and reads six dimensions in them. They are two distinct methods, published separately.
Which assistants are included in the calculation?
Those in your panel, qualified for the period and declared next to the value. Every combination of question and assistant carries the same weight.
Related paths
After “AI Trust Score”, the next step.
Each feature answers a specific question. From here you move on to related features, to the methodology behind their measures and to help for using them in everyday work.
Explore every intelligence tool
AI visibility, AI Trust Score, search, reputation, competitors, actions and execution: a feature for every question about your brand.
ReadFrom question to action
How TrustInsider works in six steps: brand, questions, quote, evidence, page-by-page intervention and a check on the same questions.
ReadA transparent method, not a black box
How every TrustInsider measure is built: sampling of AI answers, sources and rights, index calculation and declared limits.
ReadFind the right answer
Help with access, credits, collections, integrations and data: guides, FAQs, the site assistant and support requests from your workspace.
Read