Metrics & Calculations

Net Sentiment Score distills the entire conversation into one number by taking positive posts, subtracting negative ones, and expressing the result as a percentage of everything said — driven purely by language, not by likes or views. A reading above +20 signals a healthy conversation, anything from 0 to +20 is lukewarm, and below -20 means detractors are winning the room.

This view breaks the conversation into positive, negative, neutral, and mixed shares, each expressed as a percentage of the total, alongside the average tone score. A conversation dominated by neutral posts usually points to factual or news-driven chatter rather than opinion, while a heavy mixed share means people are praising and criticizing the brand in the same post.

Share of Voice measures how much of the total conversation a given brand, topic, channel, or region accounts for. Owning more of the conversation is only an asset when it's paired with positive sentiment — high share alongside negative NSS marks a lightning rod, not a win, so it should always be read side by side with sentiment.

The advocacy-to-complaint ratio looks only at people who took a clear side, and measures what fraction were recommending or defending the brand versus complaining. A ratio above 0.6 indicates a genuine fan base, while below 0.35 signals that the most vocal customers are unhappy ones.

Brand Health Index is the executive-level scorecard, blending sentiment, the inverse of controversy, the inverse of toxicity, advocacy, and the inverse of bot share. It resolves to a letter grade — A/A+ for strong standing, B for stable, C for fragile, D/F for a brand actively being damaged — though the component breakdown should always be checked before acting on the headline grade.

The topic metrics table lines up every topic in the ontology side by side with its volume, share of voice, sentiment, controversy, spike z-score, momentum, and severity-weighted impact. It's built to be sorted by impact rather than volume, since a small billing complaint can outrank a much larger but harmless marketing conversation.

Severity-weighted impact ranks topics by the damage they're actually doing, multiplying volume by negative share and by the topic's inherent severity, rather than by sheer loudness. It's the metric to use when deciding what to fix first, since two topics with identical volume can differ threefold in real impact.

The overview KPI cards are the ten headline numbers on the landing dashboard, each carrying its current value, its change versus the prior period, a twelve-bucket trend line, and a traffic-light status. Red means act today, amber means investigate this week, and green means hold course.

The volume and sentiment series plots posts bucketed by hour, day, or ISO-week alongside the per-bucket sentiment score, with gaps filled so quiet periods show up as zero rather than vanishing from the chart. Shape matters more than height — a sustained plateau of negative chatter is a bigger problem than a single loud spike.

The rolling average is a trailing simple moving average computed over a chosen window of buckets, smoothing out day-to-day noise from the volume line. Use the smoothed line to judge the trend and the raw line to spot individual events.

Spike z-score measures how unusual the latest period is relative to that same topic's own recent history. A reading of 2 is notable, 3 is high, and 4.5 or above is critical — and because it's self-referential, a small topic doubling in size is flagged just as reliably as a large one.

Anomalies are the buckets where the z-score crosses a defined threshold, automatically classified and paired with a written explanation of what happened. A spike carrying negative sentiment reads as an incident; a spike carrying positive sentiment reads as a campaign that's working.

Momentum tracks whether a topic is growing or fading, expressed as the percentage change between the current and prior period. Above +50% a story is accelerating; below -30% it's burning out and may not need a response at all.

Burst episodes are contiguous periods where the conversation ran far above its normal rate, merged into named episodes with a start, end, and intensity weight. That weight reveals which historical flare-up was truly the biggest, not just which single day looked tallest.

Controversy Index looks only at the people who picked a side. A reading of 0.5 means opinion is evenly divided; past 0.65 the topic is working against the brand, while under 0.35 the topic is actually helping. All-neutral chatter returns 0.

Polarization asks whether people have split into two opposing camps with nobody in the middle, built from the shape of the tone distribution. Near 1.0 describes the hardest communications problem there is; near 0 means people broadly agree, whatever the mood.

The statistical engine underneath polarization: a test of whether the tone distribution has two separate humps. The threshold is 0.555 — cross it and the data supports two distinct camps.

Measures how deserted the middle ground is between a conversation's two loudest camps. A dip of 1.0 means literally no one holds a moderate view.

Consensus measures agreement itself, independent of whether that agreement is happy or angry. A score of 1.0 means the audience is saying essentially one thing; below 0.4 the conversation is too fragmented for one response to land with everyone.

Heat combines how many people are talking with how strongly they feel, so a small furious thread and a large but indifferent one are never confused. Above 60 the topic is actively burning; under 25 it's background noise.

Translates heat, polarization, consensus, and toxicity into a plain-English label. Inflammatory — two camps plus hostility and volume — is the label that means escalate now.

The bubble chart that shows which topics are both large and divisive at a glance. Top-right bubbles — high volume, high controversy — are the priorities.

Ranks the individual posts doing the most damage. Responding to the top five of them typically moves the overall metric more than a single broadcast statement.

Finds the words and phrases that reliably show up when the conversation turns against the brand. High-lift keywords are effectively the vocabulary of your critics.

Rolls seven weighted signals into one score for how close a conversation is to becoming a reputational incident. Bands run from low (under 25) through guarded, elevated, high, up to severe (80+) — but the same score can call for very different responses depending on which component is driving it.

Measures how fast anger is accumulating relative to that conversation's own normal baseline. It's the earliest warning signal in the group, typically moving hours before total volume gives any indication that something is building.

Splits Crisis Risk Index into its seven underlying sub-scores. Risk driven by bot amplification calls for evidence-gathering; the same score driven by influencer amplification calls for direct outreach.

Projects three plausible near-term futures from the current index and its momentum. Plan operationally against the escalation scenario while budgeting resources against the more likely persistent one.

Distills an author's real weight in a conversation into a single number, blending authority, reach, and volume. Above 70 is a genuine mover of opinion; below 20 is a normal customer talking.

Measures credibility rather than raw size. A focused account with 20,000 followers can outrank a scattergun account with 500,000 — and that focused voice usually matters more to your specific category.

Estimated impressions: real view counts where the platform reports one, otherwise a conservative 10% of followers. Reach is exposure, not approval — a high-reach negative post is a warning sign, not a win.

Sorts every voice into standard influencer size bands. Nano and micro accounts are usually where authenticity lives; macro and verified-media accounts are the ones that turn a story into an escalation.

Ranks everyone by influence score and shows voice share. If the top five accounts control more than a third of total voice share, you're looking at a conversation being led, not one that emerged naturally.

Isolates the authors with the strongest public stance in either direction. Both lists stay short by design: engage your top detractors directly, and hand your top advocates something worth amplifying.

Groups content that's saying essentially the same thing in nearly the same words. A cluster of 30 near-identical posts is one message repeated 30 times, not 30 independent opinions.

Scores how suspiciously synchronized a group of posts is. Near 1.0 means the posts landed within seconds of one another — a timing pattern no organic audience produces.

Checks whether account names in a cluster look machine-generated. Once the score crosses 0.6, the handles were almost certainly created in bulk.

Buckets authors by how recently their accounts were created. A young cohort carrying a large, angry share of the conversation is the classic astroturf signature.

The composite read on whether a cluster of posts is an organized campaign. 0.70+ is treated as confirmed — confirmed clusters should be stripped out of sentiment reporting.

Answers how much of a conversation you can trust as real people expressing real views. Above 90 the data is clean; below 55 the conversation is being actively manufactured.

Lets the audience's actual story-lines surface on their own rather than starting from a keyword list. Each cluster represents a genuinely separate message, so a single response rarely covers more than one.

Shows whether a story-line is running critical, supportive, neutral, or mixed. A mixed reading means the same narrative is being cited by both sides — correct the facts before picking a side.

Catches narratives that are still small but picking up speed and recency. Cheap to address today, considerably more expensive if left until next week.

Projects every post into two dimensions so similar posts land near one another. Tight, separated islands indicate genuinely distinct conversations; one large smear means the topic is really just one blurred discussion.

Counts how often two subjects show up together in the same post. Strong co-occurrence tells you which complaints travel together, so resolving one often quiets the other.

Checks whether two topics appear together more often than chance alone would predict. A lift above 2 signals a real association; near 1 the pairing is coincidental.

How connected a topic is to the rest of the conversation. High-degree topics act as hubs nearly every complaint gets routed through.

Identifies which topic bridges otherwise separate parts of the conversation — where a narrow service complaint tends to turn into a broader brand narrative.

Groups topics that are habitually discussed together. These clusters often make better reporting categories than the org chart does.

Lines up every tracked brand against the same measures, separating share of voice from share of positive voice — exposing the brand that talks the loudest but owns none of the actual goodwill.

Plots every brand on a four-quadrant map. Landing in the lightning rod quadrant — loud but disliked — is the most dangerous place to be, though usually the fastest to escape once the driving topic is fixed.

Pairs a brand against each rival directly, surfacing the specific topics where the advantage is widest — effectively writing the messaging brief.

Reveals where the conversation is concentrated and where the mood departs from the national norm. A large negative sentiment gap in one region usually signals a local issue, not a brand-wide one.

Multiplies a region's share of volume by how far its sentiment falls below the national average — built specifically to guide where field teams get sent.

Compares sentiment across language communities. Diverging sentiment usually means the audiences are hearing genuinely different messages.

Measures how badly a legacy, engagement-only tool would misreport this dataset. Above +40 points, engagement-based reporting is worse than useless.

The statistical proof of the product thesis: does popularity actually track approval? Near zero means engagement carries no information about sentiment.

Flags popular, furious posts that an engagement-only tool would have counted as brand wins — the most persuasive exhibit in any executive readout.

Projects sentiment or volume over the next N periods with a confidence band. Use the direction and band width, not the exact number — social data is noisy and the band is honest about that.

The 95% confidence band around the forecast, which widens the further ahead you look. If it spans both positive and negative sentiment, shorten the horizon rather than acting on the number.

How wrong the model was on average, backtested against data it hadn't seen. Under 15% is trustworthy, 15-30% is directional, above 30% is a talking point only.

A plain rating of how much weight to put on a forecast. Low confidence almost always means not enough history yet.

A saved condition that watches one metric and raises an alert when breached. Scope rules to a topic wherever possible — global thresholds either miss local problems or fire constantly.

Prevents a single ongoing problem from generating the same alert every refresh. Shorten the cooldown to be re-notified sooner rather than creating a duplicate rule.

The workflow state of an alert as the team works it. Acknowledge means someone owns it; resolve means the underlying condition is gone, not just that the shift ended.