Sentiment
Sentiment measures how favorably AI assistants describe a brand when they mention it - not whether the brand appears (that is visibility), and not how early it appears in the answer (that is position).
One of four states
Every answer that mentions a brand is classified as exactly one of four states:
| State | What it means |
|---|---|
| Praise | Described favorably, with no real criticism |
| Mixed | Both praised and criticized in the same answer |
| Neutral | Named without being evaluated either way |
| Criticism | Criticized, with no real praise |
An answer is only ever one of these. An assistant that calls a product excellent but expensive has not been "60% positive" - it has given a mixed opinion, and that is a different situation from one that says nothing either way.
Neutral is the most common reading, and that is not a failure of measurement. Most of what assistants say about brands is neither compliment nor complaint - it is a statement of what the brand is and who it is for. Traits is where that content is reported.
The numbers you see
Four percentages that always add up to 100%: the share of answers in each state. A brand that is Praise 60%, Mixed 30%, Neutral 10%, Criticism 0% is being argued about in a third of its answers. Every chart and table on the Reputation screen uses these same four numbers. The quote walls show the sentences, including criticism that arrived inside a mixed answer.
Wherever a single figure has to stand in for all four - a heatmap cell, a sort order - Silktide uses net favorability: the praise share minus the criticism share, mapped onto a 0-100 display so it reads like other metrics, with 50 as the point where praise and criticism balance. That number is always shown next to the four-state bar, because on its own it cannot tell a brand nobody evaluates from a brand everyone disagrees about. Both score 50.
Why not a single positivity score
Silktide used to grade each mention on a five-point scale from "advised against" to "clearly recommended", and average it. That was replaced because averaging destroys the finding: a brand loved by half its answers and disliked by the other half averaged out to exactly the same score as a brand nobody had an opinion about, and the second one is a marketing problem while the first is a product one.
The four states keep those apart, and the underlying judgments are far easier for a model to make consistently - "did the answer criticize this brand" has a defensible answer; "is this a +1 or a +2" does not.
The evidence
Every judged answer is recorded with the sentences that carried the judgment - up to three quoted spans per side, verbatim from the answer. Silktide discards any quote it cannot find in the stored answer text, so anything shown can be checked against the answer it came from.
Quotes that crown or condemn outright - "the best platform for X", "the worst of the five" - are flagged as superlative and shown first. They are the strongest sentences in the corpus; they never change a rate.
This is what the Reputation screen is built around. A rate tells you that you are criticized in 22% of answers; the quotes tell you what for. Mixed answers still count as Mixed in the table; their praise and criticism quotes appear on those two walls, with a note that the same answer also argued the other way.
What counts - and what does not
Every extracted mention also gets a relevance grade: how central the brand was to answering the question that was asked.
| Relevance | Counts toward sentiment? |
|---|---|
| Primary - offered as a real answer or candidate | Yes |
| Supporting - related, but not itself the answer | Yes |
| Incidental - named only in passing, not meaningfully connected to the question | No |
Incidental name-drops can still count toward visibility (the brand did appear). They do not move sentiment. Primary and supporting mentions count equally.
An answer that names the same brand several times is one observation, not several. A brand mentioned three times in one answer and praised once has been praised in one answer.
Interviews, not market answers, for the rates
Headline visibility and position use subject topics only - the competitive arenas where customers compare options. Questions that ask about a brand itself ("is Silktide legit?", "Silktide reviews") live on a brand topic and are excluded from those competitive figures so self-prompts cannot inflate share of voice.
Sentiment rates use brand topics only: the interview answers. Recommendations almost never criticize, so a praise share over "who should I use?" and one over "tell me about this brand" are not the same measurement. Mixing them made brands that had never been asked a hard question look safer than they were.
Quotes are different. A criticism is a criticism whichever kind of question provoked it. The Reputation screen therefore quotes interviewed brands from their own questions, and quotes everyone else from market answers, with that difference labeled. It never ranks those two populations in one score column.
Expect a lot of praise and little criticism
Assistants are markedly reluctant to criticize a named brand unprompted. Asked an open question like "best accessibility tools", they will typically list the ones they favor and simply omit the ones they do not, rather than say anything unfavorable. Criticism appears far more often when a question invites it.
This means a low criticism share is normal and not by itself reassuring, and one that is merely non-zero is worth reading. The quotes matter more than the percentage. In practice most criticism arrives inside otherwise favorable answers, so it shows up in the Mixed share rather than the Criticism one - which is why mixed quotes still appear on the praise and criticism walls, marked as also going the other way.
How to use it
- Prefer trends and comparisons over any single reading.
- Check the answer count before treating a large swing as definitive.
- Read the criticism quotes even when the rate is small; that is where the specifics are.
- A strong visibility number with weak sentiment points to different work than the reverse: you are being named, but not recommended.
Related
- Reputation screen - where to read this, brand by brand
- Traits screen - the specific traits behind the judgment
- Metrics at a glance - the full metric set
- Measuring visibility - cadence and assistants
- Workspaces and topics - subject vs brand topics