How to measure brand sentiment

Collect public posts

Start with the posts people write about the brand: public posts that match the brand's name and keyword rules, with exclusion terms for names that are also common words. Senti reads X, Reddit, TikTok and Instagram, in English, Spanish, Portuguese, Chinese, German, Arabic, Dutch and Swedish, and a new report reaches back up to 90 days. Keep each platform separate: they differ in who posts and in tone, and a slow platform should not hold back the others. Count what you found, not only what you kept.

Label each post

Label every post on its own: is it about the brand, and how does it feel about it? Posts that are not about the brand, or are spam, leave every number. Senti asks all its questions about a post in one labeling call (relevance, sentiment in five steps, topic and intent) and records the model version with each label, so a number can always be traced back to the labels behind it.

The 0 to 10 scale

Five steps run from very negative (0) to very positive (4). The score is the mean of those steps on a 0 to 10 scale: each step is worth 2.5 points, so a very positive post scores 10 and a neutral one 5. A change is stated in the same points out of 10.

Sampling

A small brand can be measured in full; a large one has to be sampled. A platform whose first page returns fewer than 100 posts is labeled in full; otherwise at most 50 posts per week and platform are labeled, chosen evenly across the week rather than only the newest. Because platforms are sampled at different rates, figures across platforms weight each platform by its share of the posts found, and a report says how many posts it labeled out of how many it found.

Intervals

A score from a few dozen posts moves a lot by chance. Give every figure an interval: Senti shows 95 percent intervals (1.96 standard errors either side), and they widen as the posts get fewer. When two intervals overlap widely, the two numbers are not shown to differ.

Compare periods

To say sentiment changed, compare periods, not single points. Senti works in weeks: a week counts once it has enough posts, nothing is tested until 6 weeks do, and each block of weeks compared needs at least 60 posts with a sentiment. The test starts from the first 3 weeks against the last 3; when it confirms a change between other blocks, the report names them A and B, always with their real dates. The latest week alone is an observation, never a tested change.

What a confirmed change means

A confirmed change is one a significance test accepts: across every pair of blocks it could have compared, the chance of a difference this large from week-to-week noise alone is below 5 percent. It says sentiment moved between two dated periods. It does not say why: a shift in which platforms the posts came from can move a score while sentiment on every platform holds, so a report splits the change into platform mix and change within platforms, and the posts behind it are where the reasons are. Not confirmed means the data does not show a change, not that nothing happened. The methodology has every rule.

A worked example