trustinsider
English
Analyse your brand

AI Trust Score method

How the AI Trust Score is built, step by step.

The AI Trust Score is an indicator from zero to one hundred of how AI assistants name, recommend and describe a brand in observed answers. Here is the brand visibility analysis methodology: weights, thresholds, minimum sample and how the rating is read.

The “AI Trust Score” panel in the Northstar Audio Overview: the rating in words with its icon, the number with its 95% interval, the sample, the personal comment and the scale with the rating thresholds.
Demo data · Northstar Audio

From answer to number

Four steps: the model never calculates the number

The score comes from counts of classifications that have already been checked. The language model reads and classifies, a deterministic judge verifies against the text, a person confirms and the code calculates. Every step leaves a trace you can reopen.

The panel questions stay the same, to the same assistants, in the same market and the same language, for the whole period. Refusals, incomplete answers and errors stay out of the count and are counted separately. Same data, same result, on any machine.

Read how we choose the panel questions
The path of an observed answer to the AI Trust Score

Four steps: the panel answers are collected, the model classifies them, a judge and your team verify them, and the code calculates six ratios and a total with its interval.

  1. Collect

    The assistants’ answers to the panel questions, in every cycle.

  2. Classify

    Mention, recommendation, citation, tone and competitors named in each answer.

  3. Verification

    A deterministic judge checks against the text and your team confirms.

  4. Calculate

    Six ratios and a total, always with their interval and their base.

Six ratios, six weights

What each dimension counts and how much it weighs

Each dimension is a ratio between observed counts, from zero to one hundred. The weights are declared, add up to one hundred and change only with a new version of the method; next to each one is the base it is calculated on.

Preference compares you with each declared competitor, one pair at a time: fifty means parity, whatever the number of competitors. Consistency measures stability only where you appear: your absence is already in Visibility and is not counted twice.

See the six dimensions in the platform
The six dimensions of the current method, with unit and minimum sample

For each dimension of the current method: its weight in the total, the unit it is counted on and the minimum sample below which the dimension stays excluded and declared.

DimensionWeightUnit of measureMinimum sample
Visibility25%The question-by-assistant cell, on complete answersAt least 3 cells and 30 observations
Recommendations20%The same cell, on all complete answersAt least 3 cells and 30 observations
Authority15%The cell in which a source citation can be assessedAt least 3 cells and 30 observations
Reputation15%The cell with mentions in a positive, neutral or negative toneAt least 3 cells and 30 observations
Preference15%The declared competitor, compared in pairsAt least 1 competitor and 30 observations
Consistency10%The question in which the answers name you at least onceAt least 3 questions and 30 observations

A word above the number

Five ratings, read on the answers and not on the company

Above the number appears a large, coloured word: LEADER, STRONG, GOOD, LIMITED or MINIMAL. It stays in English in every language while its meaning is translated, and it describes what the assistants say about you.

The number stays next to the word, small and neutral, with its interval and its base. Word and icon always appear together, because three greens side by side are not enough for people who struggle to tell colours apart. The thresholds are provisional: they are recalibrated on the real distribution after three months of measurements, with a review planned for January 23, 2027.

The rating scale with the current thresholds and the meaning of each band

Five adjacent bands from zero to one hundred, from the lowest to the highest, each with its name, its colour, its icon and what the observed answers say; set apart, the absence of a rating.

  1. MINIMAL

    In the observed answers AI assistants rarely name you or present you unfavourably: caution is needed, and work on the sources they read.

  2. LIMITED

    In the observed answers AI assistants name you little or seldom recommend you: your presence needs checking and strengthening.

  3. GOOD

    In the observed answers AI assistants know you and present you correctly: you are a reliable presence.

  4. STRONG

    In the observed answers AI assistants present you often and favourably: you are a solid choice for people who ask.

  5. LEADER

    In the observed answers AI assistants name you, recommend you and cite your sources consistently: you are a reference for people who ask.

Not rated

The sample or the coverage is not enough for a rating: you read the reason, never a band word or a number.

Method figures version soglie-giudizio-2026-10-v1, provisional, in force since October 23, 2026

Value and interval

On the threshold, the lower band applies

The rating looks at the whole number you see and at the 95% interval that comes with it, the zone where the true value falls with high probability given the sample. Three cases, one rule for each.

If the interval lies entirely within one band, the rating is that band. If it crosses one threshold, the band of the lower end applies, with “close to” the band above: a rating never promises more than the sample supports. If it crosses more than one, you read Not rated.

Three intervals from the demo cycle read with the same rule

Three items from the same demo cycle: one interval within a band, one straddling a threshold and one straddling two thresholds, with the rating the rule derives from each.

CaseValue and intervalRating readWhy
Within one band: AI Trust Score50, between 46 and 54GOODBoth ends of the interval sit in the same band.
One threshold: Authority25, between 18 and 39MINIMAL close to LIMITEDThe interval crosses the threshold of 25: the band of the lower end applies.
Several thresholds: Visibility65, between 53 and 72Not ratedThe interval touches several bands: the estimate is too uncertain for a word.

When the data is not enough

A short sample never becomes a zero

A dimension without enough observations stays out of the total and says so, with the reason. The total appears only if Visibility is measured and the coverage reaches the minimum; otherwise you read Not rated, with no number.

A dimension is measured with at least 30 observations and its minimum units. Coverage is the sum of the weights of the measured dimensions: the total appears from 70% upwards, renormalised on the measured dimensions only, and next to it is the interval it would have if the excluded ones were worth zero or one hundred.

The four checks that decide whether a total appears

First the observations of each dimension are counted, then the measured weights are added up, then Visibility and coverage are checked; finally what was left out, and why, is declared.

  1. Count the observations

    Each dimension comes out as measured or insufficient, never zero.

  2. Add up the weights

    Coverage is the sum of the weights of the measured dimensions.

  3. Check the minimum

    Visibility measured and coverage from 70% upwards, otherwise Not rated.

  4. Declare

    Excluded dimensions, their reason and the coverage interval next to the total.

A demo cycle

From a demo brand’s cycle to its rating

Here is a cycle calculated by the code on declared demo data: six dimensions, their intervals, the total and the word that results. The numbers belong to no real company and serve only to read the method.

In this cycle the AI Trust Score is 50, between 46 and 54, and reads GOOD. The strongest dimension is Reputation, 69, STRONG close to LEADER; the lowest is Authority, 25, MINIMAL close to LIMITED: it is the first point to work on. The example’s intervals come from 2000 draws, as in the demo.

Open the demo and compare the dimensions
The six dimensions and the total of the Northstar Audio demo cycle

For each dimension of the demo cycle and for the total: the weight, the value shown, the interval at the declared level and the rating the rule derives, all calculated by the code.

DimensionWeightValue95% intervalRating read
Visibility25%65between 53 and 72Not rated
Recommendations20%27between 20 and 40Not rated
Authority15%25between 18 and 39MINIMAL close to LIMITED
Reputation15%69between 57 and 74STRONG close to LEADER
Preference15%60between 52 and 67GOOD close to STRONG
Consistency10%51between 37 and 67Not rated
AI Trust Scorecoverage 100%50between 46 and 54GOOD
The “Six dimensions” panel in the Northstar Audio Overview: for each dimension the rating with its icon, the base, the comment and the bar with the value, on the same cycle as the table.
Demo data · Northstar Audio

Words from your facts

Next to the rating, a comment written on your data

Every rating and every dimension has a comment. It is not the generic meaning of the band: it is composed from your facts, and a sentence appears only if the fact it tells is really there in the cycle.

When a language model writes the comment, it receives placeholders instead of the numbers and the rating word, and the code places the values. A sentence with a misplaced digit, a rating different from the calculated one or a promise is dropped, and the comment composed from your data remains.

Discover the periodic reports
The personal comment of the demo cycle, sentence by sentence

The comment on the total of the demo cycle split into the facts that make it up: each sentence appears only because its fact is there, and the system placed the figures and the rating.

ReportsSample document

Periodic report · Northstar Audio, demo brand

Rating
GOOD, 50 out of 100, between 46 and 54.
Strongest dimension
Your strongest figure is Reputation: the tone is favourable in 69% of 93 mentions, counting neutral ones as half.
Dimension to strengthen
Your lowest figure is Authority: one of your sources is cited in 27 of 108 answers with assessable sources.
Position among competitors
On Visibility, you are ahead of the only measured competitor.
Compared with the previous cycle
Compared with the previous comparable cycle, the score is up 0.4 points.
First useful action
The first useful action concerns Recommendations: at its maximum it would add 14.6 points to the total.
Towards the next rating
The next rating threshold is 5 points away.
The system places the rating and every figure; no number passes through a model.

Every change has a name

Method, weights and thresholds: three declared versions

The method and the weights form the score’s fingerprint: changing them starts a new series. The rating thresholds have their own version, outside the fingerprint, because recalibrating them changes the word read but no number.

Two scores are compared only with the same fingerprint: the same method, weights, interval level and draws. With a different fingerprint the series stops. The rating is not stored: it is derived each time from the saved value and interval, with the threshold version used to read it alongside.

The current versions of the method, the weights and the thresholds

The calculation method, the dimension weights and the rating thresholds, each with its version, its status and its planned review, read from where the calculation and the rating are defined.

WhatVersionStatusPlanned review
Calculation methodai-trust-score-v1In force, in the fingerprintOnly with a new version
Dimension weightspesi-2026-10-v1Confirmed on October 21, 2026On January 21, 2027
Rating thresholdssoglie-giudizio-2026-10-v1Provisional, outside the fingerprintOn January 23, 2027

What it does not say

Four wrong readings, and their correct version

An indicator with a published method reads well only if you know what it does not measure. Here are the most frequent misunderstandings, next to the reading the method supports.

The AI Trust Score does not read what a model thinks, it is not a forecast and it is not a certificate issued by third parties. It sums up the answers observed in a period, for one brand, one market, one language and one panel of questions.

Read the TrustInsider methodology
Four misunderstandings about the AI Trust Score and the reading the method supports

For the rating, the change, the sample and the comparison over time: on the left the most frequent wrong reading, on the right the reading the published method supports.

Wrong readingCorrect interpretation
The judgementA high rating means the company does good work.The rating describes how the assistants present you in the observed answers.
The changeThe score went up thanks to the latest intervention.The comparison tells a difference between two cycles; the cause is proven with a control group.
The sampleAn excluded dimension counts as zero in the total.An excluded dimension has no value, and the coverage declares it.
Comparison over timeTwo scores can always be lined up side by side.Only with the same fingerprint of method, weights and draws.

Frequently asked questions

Questions about the AI Trust Score method

Who calculates the score, why the rating word stays in English, what happens on the threshold, what Not rated means and how a brand is compared with its competitors.

Who calculates the AI Trust Score, a person or a model?

The code, starting from counts of verified and confirmed classifications. The language model classifies the answers, but no number in the score passes through it.

Why does the rating word stay in English?

LEADER, STRONG, GOOD, LIMITED and MINIMAL are proper names of the scale, the same in every language: a report shared across countries reads the same way. The meaning of each band, however, is translated.

My value is above the STRONG threshold, but I read GOOD: why?

Because the lower end of the interval falls below the threshold: the lower band applies, with the note “close to STRONG”. With more answers in the next cycle the interval narrows.

What does Not rated mean?

That the period’s data is not enough for a word: a short sample, coverage below the minimum or an estimate straddling several thresholds. It is not a low mark, and the reason is always shown alongside.

Can I compare my score with a competitor’s?

Yes, on the same panel answers and with the same calculation read from the competitor’s side. The competitor’s Reputation is not observed, because tone is classified only for your brand, and its coverage declares it.

Which AI visibility metrics track success over time?

The six dimensions, read cycle after cycle with the same fingerprint: how often you are named, how often you are recommended, how often your sources are cited, in which tone, how you stand against competitors and how stable your presence is. A change tells how much they moved, not why.

How do you measure brand presence in AI search recommendations?

With the same questions put to the same assistants in every cycle: Visibility counts the answers that name you, Recommendations those that recommend you, even when the question does not mention your name. Each with its base and its interval.

Will the rating thresholds change?

They are recalibrated on the real distribution after three months of measurements. The threshold version is written next to every rating, and numbers already stored do not change.

Does a real brand’s rating appear on public pages?

On product and method pages the rating appears only on the demo brand and in declared examples. A real brand’s LIMITED or MINIMAL never appears in public.

Related paths

The methodology states how we collect, measure and interpret answers. From here you read the other parts of the methodology and the features that apply it to your brand’s data.

From question to action

How TrustInsider works in six steps: brand, questions, quote, evidence, page-by-page intervention and a check on the same questions.

Read

Find the right answer

Help with access, credits, collections, integrations and data: guides, FAQs, the site assistant and support requests from your workspace.

Read
Apply the method to your brand