Every reputation platform displays a score. Few explain how it is computed — which is precisely what makes them unusable in board meetings: a number nobody can decompose is a number nobody should decide with. This article publishes the complete HarchIQ method, exactly as it runs in production: the four components, their exact weights, their formulas, their limits. Our position is simple: methodological transparency is a condition of use, not a commercial gesture.
Why a score at all
Monitoring produces volume — hundreds of articles, thousands of mentions. A score does not replace that raw material; it compresses it for three uses: comparing brands on a common scale, tracking trajectory over time, and triggering alerts when the trajectory deviates. A good score is an honest summary: it must react to what matters, ignore noise, and — above all — signal what it does not know.
The four components and their weights
The weights encode deliberate choices. Media sentiment dominates because coverage remains the main prism of reputation. AI visibility carries a quarter because conversational engines are becoming a source of information in their own right — but no more, since its measurement is still young. Diversity and crisis share the rest: a healthy reputation is one carried by several independent voices and not crossed by active incidents.
Component 1: media sentiment
Media sentiment starts from the negative share of the flow: across all articles and mentions relevant to the brand, which fraction is classified negative by the analysis pipeline — a hybrid of lexicon and model, trained on Arabic, French and English, with dedicated Darija handling. The note applies an asymmetric penalty: the formula subtracts one and a half times the negative share, floored at zero. This asymmetry encodes an observed reality: in our markets, negative coverage marks minds harder than positive coverage reassures. A flow at twenty percent negative does not deserve eighty — it deserves seventy, and a conversation with the comms team.
Component 2: AI visibility — and the refusal to fabricate
AI visibility measures what conversational engines answer when queried about a brand. It is produced by real probes: questions asked of the engines, answers collected, classified, scored. These probes are costly and sometimes unavailable — and this is where our method departs from the industry. When no real probe is available for a brand, the component is not estimated: it is no information. The score is then recomputed on the three remaining components, renormalized over their combined weight — seventy-five points instead of one hundred. The brand sees a score honestly reduced in its methodological coverage, not a disguised number. It took us months to purge invented AI-visibility scoring from our pipelines; that purge is, in our view, one of the most consequential decisions in the product's history.
Component 3: source diversity
Diversity counts the distinct sources covering the brand over the period — each independent outlet adds points, fifteen per source, capped at one hundred. The reasoning is media-structural before it is statistical: in concentrated landscapes like ours, a brand covered by a single dominant outlet is exposed to the mood of a single editor. Diversity is not volume: three hundred reposts of one press release on aggregators are worth less, in our eyes, than three outlets doing their own reporting. The cap prevents a heavily covered brand from mechanically reaching the maximum.
Component 4: crisis exposure
The crisis component subtracts twelve points per active alert on the brand, floored at zero. An active alert is a qualified episode — not every negative article, but an identified focus: an affair, an outage, a lawsuit, a labour dispute. The component recovers as alerts close, which gives the score a convalescence memory: a brand emerging from crisis sees its score repair progressively — exactly the behaviour a board expects from a reputation indicator.
Score, grade and trend
The final note is the weighted sum, rounded to the integer. It translates into a grade: A+ from ninety, A at eighty, B at seventy, C at sixty, D at fifty, F below. The trend compares to the previous score with one point of hysteresis: more than one point of difference is required to display an arrow, up or down. Without that hysteresis the score blinks with noise and the team stops watching it — the worst fate of an indicator.
What HarchIQ is not
- It is not a financial measure: it predicts neither share price nor revenue; it describes a state of media coverage.
- It is not a moral verdict: an aggressively communicative brand can score well; a quiet, healthy brand can score average.
- It is not an oracle: it summarizes the observable flow; it does not know what has not been written yet.
- It is not a black box: every component is traceable down to the articles feeding it — the condition for an analyst to be able to challenge the score, and therefore to trust it.
Publishing our method exposes us to debate — some analysts will challenge the crisis penalty, others the AI-visibility weight. Good: a challengeable score is an improvable score. The industry alternative — proprietary, opaque, flattering scores — produces board meetings where people argue about a magic number. We prefer board meetings that argue about the real world the number summarizes.