Brand sentiment in AI: how models talk about your company
Updated on
Brand sentiment in artificial intelligence describes the tone with which an AI model talks about your company in a specific answer, and Mirvel expresses it for every brand named with a label that runs from very negative to very positive, a numerical score between minus one and plus one, a summary of what is said and the topics the brand is associated with, so that you know not only whether you are named but also whether it is to recommend you or to warn about something.
Sentiment toward the brand across 506 real assessments
In this example project 90% of the assessments of the brand are positive or very positive, 10% are neutral and none is negative, with an average score of 0.64 on a scale that runs from -1 to +1.
- Very positive 19%
- Positive 71%
- Neutral 10%
- Negative 0%
- Very negative 0%
Average score: 0.64
Real data from a production project: 1,503 responses analyzed between August 6 and October 4. Brand names have been replaced with generic labels.
What does each assessment include?
Every time a brand appears in an answer, the analysis produces an assessment with four elements: a five-level label that sums up the tone, a numerical score that makes it possible to sort and average, a one-sentence summary of what the text says about that brand and the topics it links it to, such as price, customer experience or a particular specialty.
Today you can consult these assessments in the detail of each analyzed answer, where you see next to the original text exactly what is said about each brand, which is particularly useful for understanding the specific arguments the models use when they recommend you or leave you out.
Why does an absent brand have no sentiment?
Mirvel only assesses the brands that actually appear in the answer, and this rule is deliberate because earlier tests showed that, when an assessment was also requested for absent brands, the analysis tended to invent a negative score to stand for that absence, which mixed up two different phenomena that are measured with different indicators.
Absence is already reflected in visibility and in share of voice, so sentiment is kept for a different question, which is how you are described when you are named, and for that reason it should never be read as an overall reputation grade but as the tone of the occasions on which you appear.
What are the limits of this analysis?
It is an automatic reading of the text of each answer carried out always with the same method, which makes it comparable across weeks and across brands, but it is not a survey and it does not replace human judgment, so it is advisable to review the summaries of the most important cases before drawing final conclusions.
In addition, the tone reflects what the models say and not necessarily what your customers think, although in practice the two influence each other, because the models base their answers on what has been published about your brand, and for that reason poor sentiment usually points to specific content, reviews or mentions that are worth reviewing.
Frequently asked questions
- Can sentiment be negative even if my brand appears first?
- Yes, because position and sentiment are independent: a brand can be the first one cited precisely to warn about a problem, which is why Mirvel records both pieces of data separately and why it is best to read them together.
- How is the average score calculated?
- It is the average of the individual scores, which run from -1 to +1, of all the assessments of your brand in the period analyzed, and since only the answers in which you appear are assessed, the sample size is the number of answers that name you, not the total number of answers.
- What are the topics associated with the brand?
- They are the ideas or attributes the model links to a brand within an answer, such as a particular service, a product feature or a type of customer, and they help you understand why you are recommended or dismissed in each situation.