The Illusion of Sentiment Analysis
Why AI Sentiment Features Are Not Useful
AI monitoring tools are rapidly adding sentiment analysis to their dashboards. It is presented as a way to understand whether AI speaks positively or negatively about your brand. On the surface, that sounds useful. If you are trying to manage how AI represents you, then knowing whether anything negative appears feels important.
The problem is simple. AI does not have feelings. It does not hold a positive or negative view of your brand. It does not like you or dislike you. It only produces text based on patterns in the data it has been trained on. When a tool claims to measure sentiment, it is not measuring an emotion inside the model. It is measuring the output of a prompt that was designed to create the appearance of sentiment.
What Sentiment Analysis Actually Is
Traditional sentiment analysis was based on keyword scoring. If a sentence contained words like great or excellent, it was labelled positive. If it contained words like terrible or disappointing, it was labelled negative. It was crude, but at least it was grounded in the text itself.
LLM based sentiment analysis is different. It involves prompting the model to behave as if it has an opinion. For example, if you ask an LLM, “What are the good and bad things about Brand X,” it will always produce some good things and some bad things. It does this even if the brand is obscure, even if the model has no meaningful data, and even if the negative points are invented. The prompt forces the polarity. The model fills the gaps with whatever seems plausible.
This means the sentiment is not discovered. It is generated. It is a performance, not a measurement.
Why This Creates Misleading Results
LLMs do not track reality. They do not know what happened today. They do not distinguish between fact and pattern. They do not store opinions about brands. When they produce a negative point, it may be based on a hallucination, a generic criticism borrowed from similar companies, or a pattern that simply fits the structure of the question.
Despite this, sentiment analysis sells well. It looks measurable. It gives dashboards something colourful. It offers executives a simple number that feels like a KPI. It is easy for vendors to bolt onto an LLM and present as insight. But it does not reflect real user perception, brand reputation, or actual AI behaviour. It is a comfort metric, not a truth metric.
What Brands Actually Need Instead
If you want to understand how AI represents your brand, there are far more reliable questions to ask.
A simple evaluation checklist:
- Is the information factual
- Is it consistent across prompts
- Is it grounded in real sources
- Does it hallucinate
- Does it omit important details
- How does it compare to what the model says about competitors
These are measurable. Sentiment is not.
Using Waikay for ‘Sentiment Analysis’
Although Waikay does not treat sentiment analysis as a meaningful metric, it does offer a Facts feature. This feature distils what AI models say about your brand and groups those statements by topic.
When you create a topic report focused on a single idea, Waikay prompts each model to respond with what it “knows” about your brand. Those responses are then simplified into clear, single-sentence statements that you can review and mark as correct or incorrect.
Marking a fact as incorrect does not provide direct feedback to the model. Instead, it gives you and your team a structured starting point for identifying the source of a hallucination, misunderstanding, or genuinely incorrect claim about your brand. You can also filter results by model to understand why different AIs produce different inconsistencies.
In most cases, the root cause is simple: the brand has not been explicit or structured enough about what it actually offers.
The Real Illusion
AI does not love you or hate you. It predicts the next word. Sentiment analysis turns that prediction into a feeling and sells it back to you as insight.
