Skip to main content

​
What sentiment tracking answers

Visibility tells you how often AI models mention your brand; sentiment tells you what they say when they do. Responses with a determinable stance toward your brand get a score on a 100-point scale; thin name-only mentions are left unscored. How scoring works is covered in what is a sentiment score; this article is about the workflow: reading the trend, spotting a problem, and finding the exact responses behind it. That last step is the point of the page. An average score of 62 is not actionable. The three responses that call your product “overpriced with slow support” are. You’ll find the page in the sidebar under Monitoring → Sentiment.

​
Reading the trend

The top of the page shows your daily average sentiment score over the selected date range. Next to it, a summary card shows the current score out of 100, the change versus the previous period, a score-position bar, and a breakdown by prompt type (Organic, Competitor Comparison, Brand Specific). The Promptwatch Sentiment page with the trend chart and current-score summary cards. Two reading habits keep this honest:
  • Judge the level once, the movement always: a stable 55 is a fact about your brand; a drop from 70 to 55 in two weeks is an event with a cause. Investigate movement, not levels.
  • Check the volume behind the line: on days with few brand mentions, one harsh response swings the daily average. A one-day dip is noise; a multi-day slide is signal.
Below the chart, an AI Insights card can show a weekly digest of recent sentiment, with a View actions button that opens Sentiment items on the action board. The card covers the last week regardless of the date-range filter, and hides when there is no recent summary.

​
Slicing the data

The filter bar narrows the page. Models, types, topics, and date range apply to the chart and both tables; score ranges apply only to the tables, not the trend chart:
  • Sentiment score ranges: < 30, 30 - 50, 50 - 75, and 75 - 100. Selecting < 30 is the fastest route to problem responses.
  • Models: sentiment often diverges per platform, because each model reads different sources. A score that’s fine on ChatGPT and poor on Perplexity points at the review sites and forums Perplexity favors.
  • Prompt types and topics: comparison prompts usually score lower than informational ones, since answers weigh you against competitors. Filtering by topic shows which themes carry the negativity.
  • Date range: scope the analysis to a release, a campaign, or a PR incident.

​
Finding the responses dragging you down

Below the chart, two tabs hold the drill-down:
  1. Open the By Prompt tab, already sorted lowest score first. Each row shows a prompt with its average score as a bar, its type, brand mentions, topics, and response count, so you see which questions consistently produce negative answers. Expand a row to see its responses.
  2. Switch to the Responses tab for individual responses, each with its own score, model, and date. Apply the < 30 score filter (add 30 - 50 for the next band).
The Promptwatch Sentiment Responses table filtered to negative scores and sorted lowest first.
  1. Click a response to read it in full. The response view shows what the model actually wrote and which sources it cited.
The cited sources are usually the answer to “why is this negative?”. AI models don’t hold opinions; they synthesize what they read. A negative response almost always traces back to a specific review page, forum thread, or outdated article, and that source, not the AI answer, is what you can act on. Both tables export to CSV when you need to hand the list to someone else.

​
Making it a routine

Sentiment earns a slot in your weekly review: check the trend against last week, and if it moved, run the drill-down, filter negative, sort ascending, read the new responses, note the sources. Over time you build a short list of the external pages shaping your brand’s tone in AI answers, which is a far more concrete asset than the score itself.