Sentiment is not enough: the four scores that actually describe a mention

Sentiment analysis is the feature every monitoring tool leads with, and the one practitioners trust least. Not because the classifiers are bad — modern language models read tone far better than the keyword-polarity systems of a decade ago — but because sentiment answers the least interesting question about a mention. Is it negative? Fine. Negative and what?

A one-star review from a throwaway account and a skeptical feature in the trade press are both "negative." One is noise. The other can change a quarter. A scoring system that can't tell them apart isn't measuring reputation; it's measuring mood, averaged until it means nothing.

Here's the fuller picture we think a mention deserves.

Credibility — should anyone believe this?

Credibility asks about the source and the claim together: is this a sourced piece from an outlet with editorial standards, a personal opinion clearly framed as one, or an unsourced claim on an aggregator that reprints anything? Low-credibility negativity often deserves no public response at all — responding elevates it. High-credibility negativity deserves a response even when it's polite and small. The score changes the playbook, which is the entire point of scoring.

Influence — how many people will hear it?

Influence is about the author, independent of the message: audience size, yes, but more importantly engagement — a hundred thousand inactive followers matter less than ten thousand people who actually read, reply and repost. Influence converts "someone said X" into "an audience of roughly this size and composition just heard X," which is the version of the fact a PR team can act on.

Virality — is this going somewhere?

Virality is the forward-looking score: given the early engagement pattern, the platform mechanics and who's amplifying, how likely is this mention to be much bigger tomorrow? It's the most valuable score to get right, because it is the one that gives you time. A mention flagged as high-virality at two hundred views means you can prepare while the story is still small. The same mention discovered at two hundred thousand views means writing a statement under pressure.

Related but distinct: mainstream likelihood — the chance an online conversation crosses into traditional press coverage. Online stories that never reach the press can often be waited out. The ones that reach the press were usually visible online first.

Sentiment — last, deliberately

Sentiment still matters; it's the emotional direction all the other scores modify. But notice what happens to it in context: it becomes an input to judgment instead of a verdict. Negative + low credibility + low influence + low virality = log it and move on. Negative + credible source + rising virality = the top item in this morning's briefing, with a recommended action attached.

No score without a why

One more rule we build by: a number you cannot question is a number you cannot defend. When a PR lead tells a client "we're not responding to this," they need a better justification than "the tool scored it 23." Every score Clairwire produces comes with a plain-language rationale — credibility is low because the outlet has no sourcing history and the claim cites no evidence — so the human can check the reasoning, disagree with it, and stay the actual decision-maker.

That is also the honest way to use AI in this domain. Language models are genuinely good at reading tone, weighing sources and spotting patterns across thousands of items — and genuinely fallible at all three in ways that show up without warning. Rationales keep the fallibility inspectable. Scores without reasons ask for faith; scores with reasons ask for review. Only one of those belongs in professional work.