How AI classifies your brand mentions, and where it still goes wrong
Modern listening tools hand each collected mention to a language model with a job description: decide the sentiment toward this specific brand, categorise the topic, judge severity, and flag whether a response would help. Understanding the pipeline helps you trust its outputs appropriately.
What the model does well
Context. "Great, third price rise this year" reads as the sarcasm it is. "The install was late but the engineer was brilliant" comes back mixed, service-positive, logistics-negative. Category sorting, is this about price, product, service, or general banter, lands reliably, and at a per-mention cost measured in fractions of a penny, full-coverage reading becomes economic for the first time.
Where it stumbles
- Whose sentiment: posts about a sponsorship can be furious at the club and neutral about the sponsor; tuning for the client entity matters
- Deep irony and in-jokes: community-specific humour still misleads models occasionally
- Ambiguous severity: distinguishing a grumble from a brewing story requires judgement about the author's reach, not just the words
- Manipulated input: text written to game classifiers, which well-designed systems constrain and validate against
The human layer
Sensible programmes keep people reviewing all negative classifications and a sample of everything, feeding corrections back into the prompt. The division of labour is clean: machines read everything so humans judge the things that matter. Full automation without review drifts; full manual reading does not scale. The pairing works.
MyView watches every mention of your brand and tells you what needs a reply.
Book a demo