Reading AI sentiment without absolute thresholds

Is 71 good? Somebody asks a version of that most weeks, about a sentiment score on a slide, and honestly nobody in the room can say. Not the vendor. Not the analyst who built the slide.

That is not a data problem. It is a threshold problem. AI sentiment has no absolute scale on which a number becomes good, and pretending otherwise is how sentiment reporting becomes decoration.

There are two honest ways to read it, and neither is a magic number.

Direction: your trend against your own history

The first reading is movement. Is your positive-sentiment share rising or falling against your own recent weeks?

Your baseline is not neutral ground. It reflects your category, the controversy your industry attracts, and the way models happen to talk about companies like yours. No universal threshold normalizes any of that away, which is why the level on its own tells you so little.

Reading against your own trailing window strips all of it out. Whatever your baseline happens to be, direction against it means something, and it is the first thing worth knowing.

Window choice is part of the honesty. Pick one long enough to smooth weekly chatter and short enough to catch a real shift, then read it the same way period after period. The failure mode is cherry-picking whichever comparison flatters this month, and it is easy to fall into.

The trend also survives model updates better than any absolute number. An update that shifts everybody's baseline shifts yours and your rivals' together, so the gap between you holds steadier than the level does.

Distance: your gap on the same prompts

The second reading is position. On the same prompt basket, do assistants speak more warmly about you or about your rivals?

Your prompt basket is the set of questions replayed in the engines to measure you, generated from what your buyers search for. Asking every brand the same questions in the same period is what makes a gap fair, rather than an artifact of who wrote the list.

The gap puts a bad week in proportion. A falling positive share during an industry-wide rough patch reads differently when every rival fell further, and a rising share means little if the category rose faster. Together they tell a story neither tells alone.

This is why the recipe behind those prompts matters here. They come from first-party demand, and they go to every brand identically. A gap measured on a hand-picked list inherits that list's bias twice over, once in your favor and once against whoever got left off it.

One more habit keeps it honest. Read the gap engine by engine before you blend it. Assistants differ in how warmly they discuss whole categories, and an aggregate can hide one engine's problem inside another's enthusiasm.

See also: how the prompt basket gets built, and why hand-picked lists lie →

A sentiment number alone is decoration

So the tools refuse to hand you one on its own. One check carries the trend, another the competitor gap, both per engine.

Both land beside citation rate, how often an AI answer names one of your pages as a source, and share of voice, your slice of that answer space against the rivals you track. Sentiment arrives with them or it does not arrive, because the rule lives in the analysis rather than with whoever is building slides at eleven at night.

The restraint that refuses a threshold also refuses false precision. Sentiment classification is model-read language, so Quattr treats it as directional evidence rather than an instrument reading, and the card says so.

It is also why sentiment never raises an alarm alone. A move sends you to citations and share on the same prompts. Three metrics failing together means something different from one twitching alone.

Reporting inherits the shape: a line and a gap, never a lone percentage in a big font, with the scope note beside the chart so nobody mistakes the denominator.

See also: citations, mentions and share of voice, measured honestly →

Sentiment is half a bigger question

In the Quattr Method, sentiment is the trust half of R4 · Represented. That is the fourth question: does AI describe your brand the way you would describe it yourself, per your positioning, and recommend it with confidence?

The other half is description accuracy, which sits on the roadmap rather than in the product. So R4 ships flagged partially instrumented until it lands, and every page touching it says so.

For sentiment the summary fits in a line. Watch the trend, mind the gap, distrust the threshold.

Teams who adopt trend and gap stop arguing about whether 71 is good. They notice instead when the trend breaks or the gap closes, the only two sentiment events that ever deserved a meeting.

And when somebody insists on a threshold anyway, hand them the two questions that replace it. Better or worse than our own last quarter? Better or worse than our rivals on our prompts? Every decision a threshold pretended to enable lives inside those two answers.

See also: the fourth question, and why it ships flagged partially instrumented →

The workflow that does this: AI platform scorecard →