Sentiment scores are a triage tool. They are not a brand health grade. A model that labels posts positive, neutral or negative will be wrong on sarcasm, mixed language, and screenshots of something that happened elsewhere.
The failure mode is executive reporting that says “sentiment is 72% positive” while the negative 8% is a concentrated service failure among high-reach accounts. Average sentiment improved. The risk did not.
Use sentiment as a filter: what should a human read today? Then add topics (is it price or outage?), source (news vs a single community), and momentum (is this new?). Share of voice without those cuts is just a louder vanity metric.
In mixed-language markets, code-switching and context-heavy jokes will trip models trained mostly on a single variety of English. Plan for analyst review. Do not automate a public response from a score.
Brand254 shows sentiment next to volume and themes for this reason. The number is a starting point for investigation, not the conclusion of one.