The line we don’t cross

Undoing one industry word, one honest distinction at a time.

VM

Vadim Matyushkin Behavioral Scientist & Sociologist · ICDS

PublishedNovember 24, 2025

The industry’s default word is sentiment, attached to a default method called sentiment analysis, and in ordinary use it slides very easily into a synonym for emotion — as if counting text told you what someone felt. We did not start by rejecting the word outright. We started by shrinking it.

01

Shrinking the word

The unit we count is Expression, nothing more claimed yet: whether a mention carries any evaluative charge at all. No polarity attached to that word yet, no intensity, no verdict about feeling. Just: charged, or not.

02

The neutral bin nobody looks inside

A standard sentiment tool then sorts everything into three bins — positive, negative, neutral — and moves on. The first two bins get all the attention. The third quietly absorbs almost everything else, and almost nobody asks what is actually sitting inside it. In our own corpus, that third bin is enormous: of 273 million brand mentions collected over two years, 96.5% carried a neutral tone. Only about 10 million mentions — roughly one in twenty-seven — carried any positive or negative charge at all. If “neutral” is treated as one thing, the overwhelming majority of the dataset is being described by a label that explains almost nothing about it.

03

Not one thing

Because once you actually look inside that bin, it is obviously not one thing. A product listing with a price and a spec sheet is neutral. A post where half the replies praise a brand and the other half tear it apart, in exactly equal measure, is also neutral, by the arithmetic a standard tool applies. But these are opposite situations. One is silence — nobody had a strong opinion worth expressing. The other is a real argument that happens to cancel out on paper. A brand peak that is 2% expressive and a brand peak that is 40% expressive but perfectly balanced would both be reported, by conventional sentiment scoring, as “neutral overall” — and anyone reading only that label would have no way of knowing they were looking at two entirely different events.

04

Two axes, not one

So “neutral” had to be taken apart into the two questions it was quietly answering at once. Expressive Intensity asks the first one: of everything said about a brand in this peak, what share carried any evaluative charge — positive or negative — at all? This is the silence question. Balance of Expression asks the second, computed only among the mentions that already cleared the first bar: among the people who did have something evaluative to say, how lopsided was it — did one side dominate, or did the field genuinely split? This is the argument question. A peak can score low on the first and extreme on the second — quiet, but nearly one-sided among the few who spoke. Or high on the first and near-zero on the second — loud, and genuinely torn. Standard sentiment analysis collapses both possibilities into the same single neutral score. We refused the collapse because the two situations call for opposite editorial and investment decisions, and no brand manager should have to guess which one they are looking at.

05

A question of jurisdiction

A parallel confusion sits one level up, and it is less about vocabulary than about jurisdiction. Some practitioners look at the same media field and describe themselves as reading perceived risk — as though tallying online behavior toward a brand told you what people privately believe their own exposure to be. We take the more modest position that the best instrument science has for finding out what is actually on someone’s mind remains asking them directly — imperfectly, since self-report carries its own distortions, but still the least-bad tool available for that particular job, and one we are not in a position to improve on from a social listening dashboard.

06

Behavior, not perception

What social listening honestly gives us is not perception. It is behavior: realized incidents, how visibly they were discussed, how the tone shifted once the story broke. We name that Behavioral Risk Index precisely so it cannot quietly borrow the authority of a survey it did not run. And the relationship between the two is not the simple one-way shadow it might first appear to be. A real incident can create perceived risk that did not exist in anyone’s mind beforehand — behavior as the cause. Or an already-anxious stakeholder group, primed by something outside our data entirely, can overreact behaviorally to a minor incident that would have registered as nothing anywhere else — perception as the cause, behavior as its visible symptom. The index cannot tell, by itself, which direction the arrow runs in any given case. It can only insist on staying honestly on its own side of that question until the evidence says otherwise.

07

Three moves, one discipline

Three moves, one discipline running through all of them: shrink the word to what it can actually prove, then look inside whatever bin absorbed the leftover complexity, and only rebuild distinctions where the data can genuinely support them. Sentiment became Expression, so it would stop implying access to feeling it never had. Neutral became two separate axes, so a silent peak would stop being confused with a torn one. Perceived risk stayed with the surveys that can actually ask about it, so behavior would stop quietly standing in for belief. None of it is stubbornness for its own sake. It is simply refusing to let a measurement’s receipts claim to add up to more than what was actually paid.