What is measured sentiment?

Sentiment answers one question, for every AI answer that names you (or names a competitor you track): how favorably does this answer portray it to the person asking, on a scale of 0 to 100. 0 means clearly discouraged, 50 means neutral or simply listed among other options, 100 means strongly recommended.

It's the SAME scale for your brand and for every competitor you track — without a shared scale, a "you vs. them" comparison wouldn't mean anything.

What the score covers

Each scored answer receives:

  • An overall score, on the scale above.
  • A per-aspect score, but only for aspects the answer actually discusses for that entity. The six aspects are fixed and the same for every brand: Quality, Price, Support, Ease of use, Reliability, Offering. An aspect the answer doesn't mention does not appear in the score — a score of 50 on price means "price is discussed, the verdict is mixed," never "price isn't mentioned."

Your brand description is given to the judge as context, so it knows which company it is scoring. It never makes the score stricter: your brand is judged on the same scale as its competitors.

Every answer counts, not a sample

Every run is scored — up to 3 calls per model on the default panel — not just the first. A model's own noise is around 15% from one answer to the next: judging on a single run would throw away two-thirds of the sample and make aspects and competitor comparisons rarely displayable.

A missing score is never a neutral score

  • An entity not named in an answer has no score for that answer.
  • An entity named but whose text couldn't be analyzed (answer too long, or an unusual spelling) has no score either.
  • In both cases the missing score is null, never 50. A 50 is always an expressed mixed verdict, never an empty box.

The dedicated sentiment screens show coverage ("scored on N of M answers that name you") and, below 3 scored answers, say "sample too small" instead of showing a number that would look more precise than it is. The same threshold applies to what the Mentova assistant tells you; the CSV export delivers the raw numbers with their sample size (n), for you to judge.

Average of tracked competitors

The measurement computes the average of the competitors you track (the ones you haven't hidden): it's a comparison point, not a ranking. It appears as a second series on the sentiment and citation-rate charts, and in the history CSV export.

  • A tracked competitor that's never named counts as 0% mention rate — it was measured, it's simply absent from the answers.
  • A tracked competitor qualifies for sentiment with at least 3 scored answers; otherwise it doesn't enter the average.
  • At least two tracked brands must qualify for an "average" to exist. With just one, the series carries that brand's name instead of the word "average" (and the export leaves the "average" column empty).
  • With no tracked competitors at all, there's no average — never a zero average that would suggest something was measured when nothing was.

The method change is shown, never smoothed over

The measurement has changed twice, and each change carries a method number:

  • Method 1 (the first version) only scored your brand, only on the first of the three runs, on a scale relative to your stated positioning.
  • Method 2 scores the brand and its competitors on a shared scale, on every run.
  • Method 3 (since October 20, 2026) keeps exactly the scale and instructions of method 2 but changes the model that scores, because its provider is retiring the previous one. Even with identical instructions, two judges don't give the same scores: we measured a 4-to-5-point average gap, against 1 to 2 points when the same judge scores twice. So it's treated as a new method, not as a continuation.

These methods are never blended or compared to each other: no variation (percentage, alert, report summary) is ever calculated across the change, and the PDF report flags when its period spans it. A vertical marker on the curve flags the day the measurement changed. If you see a sudden sentiment drop on a given day, check whether it's a method change before assuming it's a real shift in tone.

No retroactive scoring

This measurement doesn't rescore the last 90 days of answers: scoring them today would produce scores dated in the past but based on today's judgment, which would be misleading. The new measurements therefore start with your next analysis. Until a recent answer has been scored under the current measurement, the dedicated screens say so plainly instead of showing an empty chart with no explanation.

Common pitfall

Never compare a point before and a point after a method change (the export's "Sentiment Method" column shows it) as if it were a real shift in how AI talks about you. Also only compare measurements backed by a sufficient number of answers: a few points of difference on 3 answers isn't a signal, it's noise.