Measurement protocol

How we measure

Every figure in the dashboard comes from somewhere. This page says where: per measurement the source, the rhythm, the formula, and what is deliberately left out.

Set on 20 August 2026

Anyone selling measurability should be able to show their own measurement. This is not a summary but the actual method; if it changes, this page changes with it and the date above says when.

What counts as a mention

Everything starts here. Every other figure on this page is calculated over mentions that passed this filter, so this one governs the rest.

The rule	A mention counts if it is about this organisation. Not: a company with the same name, a quote lifted from another context, or the same text syndicated across twenty sites.
How it is decided	A language model assesses every mention with the organisation knowledge profile as context: related people, connected organisations and topics. Without that profile a namesake cannot be told apart from a real mention.
What happens to the rest	What falls away is kept and stays visible, with the reason attached. Filtered out is not thrown away, because a judgement has to be checkable.
How much falls away	That differs per organisation and per period, so we quote no fixed figure here. The share the filter holds back is shown per organisation in the dashboard.

Sentiment

Sentiment is determined per individual mention, not per brand. The brand figure is a sum of those, and that sum is what most questions are about.

The scale	Five classes, from very negative to very positive, alongside a score from 0.0 to 1.0. Below 0.3 a mention counts as outspokenly negative.
What is assessed	The mention in full: for a video or a podcast that means what is shown and said, not only the accompanying text.
The sum	The brand figure over a period is the arithmetic mean of the individual scores in that period. The label beside it is the class that occurs most often.
Every mention weighs the same	We do not weight by reach. Reach cannot be established reliably for most sources, and an estimate that disappears into a formula as if it were a fact makes the end figure impossible to follow. The consequence: a widely shared post moves the average no more than a small one. That is a choice, and it is stated here so you can check it.

Brand figure over a period

period figure = sum of scores / number of mentions in the period

score: the value between 0.0 and 1.0 assigned to each individual mention
number of mentions: everything that passed the relevance filter in that period
label: the class occurring most often in that period, so the mode and not the mean

Twenty mentions with scores summing to 11.6 give 11.6 / 20 = 0.58.

This section covers only the answers AI assistants give to questions that do not contain your brand, such as someone asking which provider they should use for something. There the question is whether you appear at all, which is a different thing from counting how often you were written about. For the other channels we calculate no share; there we count and weigh mentions.

What it covers	Questions without a brand name in them. Ask about your name and you will almost always get an answer about you; that says nothing. The question that counts is whether you appear in the answer to what your customer actually asks.
Visibility, the headline figure	The proportion of measured answers in which your organisation appears, over a thirty-day window, with the previous period beside it. This is a presence rate, not your portion of all brands named.
Share of mentions	Your mentions divided by all brand mentions in the same answers. That is what the term share of voice has always denoted, and it is a lower figure than visibility as soon as several providers are named per answer.
Prominence	Where you are named among the other brands in that same answer, and how many brands are named in total. Those two stay side by side rather than being divided, because an answer listing twelve names is a different thing from one naming three.
How that position is established	A language model lists every provider the answer substantively names, in order of prominence. So it is a judgement, not a count of word positions. Platform names, source citations, product categories and place names do not count, and the list stops at fifteen.
No composite figure	We do not fold share and prominence into one number. That would require a weight we would have invented, and then we would be publishing an assumption as if it were a measurement.
What we do not do here	We calculate no share across news media, social channels or reviews. That would require a denominator that does not exist: the number of times a topic was written about is not a figure anyone knows.

Visibility

visibility = answers naming you / all measured answers

The denominator is fixed: every round we put the same questions to the same platforms.
So a rise means you took a place someone else lost, not that the topic happened to be busy.
Mind the name. This is presence, not a share of attention.

Seven questions across eight platforms is 56 answers per round. Over four rounds 224; appear in 54 answers and visibility is 24 percent.

Share of mentions

share = your mentions / all brand mentions in the same answers

The figure share of voice has always denoted. It comes out of the same measurement, because for every answer we record each brand named with its name and position, plus whether it is the client, a tracked competitor, or a provider not yet known.
Limit: at most fifteen brands per answer. If an answer names more, the denominator is truncated.
Always lower than visibility as soon as more than one brand is named per answer.

Appearing in four of ten answers that together name 45 brands: visibility 40 percent, share 4 / 45 = 9 percent.

Prominence

prominence = your position in the list, with the number of brands named beside it

Two separate figures, not a fraction: first of four and third of twelve would otherwise give the same result.
The order comes from a judgement by a language model, which lists every provider named with the most central brand first. Platform names, source citations, product categories and place names do not count; the list stops at fifteen.
If you are not in that list but you are in the text, we establish your position from where your name first appears.

Named second among five providers is recorded as 2 of 5. Over a period we average both figures across the same answers, so the reading is: position 1.4 on average out of 4.8 brands named, in 40 percent of answers.

For questions that do contain your brand name we measure something else: sentiment, recurring subjects, and which other brands appear in the same answer.

Being named and being cited are two things

This belongs with the previous section and applies to the same AI answers. An assistant can name you without using your site, and use your site without naming you. So we count them separately.

Count 1: mention	Does your name appear in the answer, and where among the other brands named.
Count 2: citation	Does your own domain appear among the sources the assistant cites for that answer, how often, and what share of all cited sources that is. Four out of twelve is not the same as one out of twelve.
Named and cited	The strongest position: the model knows you and uses your own words.
Named, not cited	The model knows you, but someone else supplies the story. This is where most can be gained, and the first question is whose page holds that place.
Cited, not named	Your page feeds the answer while another name sits on top of it. A silent contribution that stays invisible without this count.

Citation share per answer

citation share = sources on your domain / all cited sources

Calculated per answer, then averaged over the thirty-day window.
An assistant citing no sources does not enter the denominator; those answers are counted separately.

Twelve cited sources of which four are yours: 4 / 12 = 33 percent.

How often we query an assistant

The same model does not give exactly the same answer to the same question every time. That is a property, not a fault, and it determines how you have to measure.

The rhythm	One query per question per platform, weekly, across eight AI platforms.
Where the stability comes from	From breadth, not from repetition. How many questions run depends on the plan; in practice that is five to seven containing the brand name, across eight platforms. So forty to fifty-six observations per round, and variation in a single question disappears into that.
What that means	At the level of one question it is a single observation and the answer may differ next week. At the level of your brand it is a set, and that only moves when something actually changes.

How we count search positions

A position no longer means what it meant ten years ago. The results page is full of blocks, and those take up room.

The numbering	We count sequentially and de-duplicate, with blocks taking their own place. A local results block in ninth place is position nine.
Why this way	Because it has to match what you see when you search yourself. That is the only check anyone ever performs, and a numbering that deviates from it is useless in a conversation.
What comes with each position	The sentiment of that specific URL, so it is visible whether position four is a glowing review or a damaging article.

What is deliberately left out

No advertising value. Putting a euro figure on editorial coverage is a sum the profession itself abandoned, and we are not reviving it.
No reach estimates inside the formulas. Where reach cannot be measured reliably, it does not go in.
No composite total with a hidden weight. Two figures you can check beat one you have to believe.

Something not adding up, or want to know how a specific figure in your dashboard came about? Email info@reptor.ai and we will walk through it.