AI sentiment monitoring tools for brands: how we track tone in AI answers for clients
How we run AI sentiment monitoring for brands: scoring tone in ChatGPT and Perplexity answers, tracing it to sources, and what we report to clients.
When a client asks about AI sentiment, they are usually reacting to something they saw: an assistant described them as expensive, or dated, or confused them with a competitor. They want to know whether that is a one-off or a pattern, and what to do about it. That is a measurement problem first and a content problem second.
This is how we handle it.
What we mean by AI sentiment
We separate two things. Sentiment of AI answers is the tone in which assistants describe a brand when asked across a set of prompts. Sentiment of sources is the tone of the social posts, reviews and articles those answers draw on. Social listening covers the second. It does not cover the first, because generated answers are not posts. We need to measure both, and the first needs its own tooling.
Our setup
We run client programs on Promptwatch. Sentiment there is a per-response score from 1 to 100, with a weekly AI summary. The per-response detail matters. A single blended number hides what we need, which is the actual sentence an assistant used to describe the client.
Before we trust any score, we configure the Brand Book. Its analysis instructions decide what counts as a mention. Without that, a client with a common name or overlapping product lines gets a pile of irrelevant answers scored as if they were about them. We also set brand aliases and check a sample of answers by hand in week one.
The panel is built by topic, persona and country, and prompts are tagged by type: organic, brand and competitor comparison. Brand and comparison prompts are where sentiment is most visible ("is X worth it", "X vs Y"), so we weight our attention there. Prompts run daily per model, country and language, and we review trends over weeks using the A/B period comparison. Single days are noisy because answers vary between runs.
Real-UI monitoring covers ChatGPT (Search), Google AI Overviews, Google AI Mode, Microsoft Copilot, Gemini, Perplexity, Amazon Alexa and Meta AI. Claude, DeepSeek, Mistral and Grok are tracked via API, and we label them that way in reports.
Tracing tone to sources
A score tells you something is off. It does not tell you why. For that we use the citation layer. Promptwatch shows which pages and domains each response cites, with Reddit, YouTube and LinkedIn reports and offsite mentions (the brand named in third-party pages without a link). When a negative description keeps appearing, we usually find one of three causes:
- An outdated page on the client's site, such as old pricing or a retired feature.
- A third-party page that describes the brand inaccurately.
- A thin spot. No page answers the question, so the model fills in from whatever it finds.
Each cause has a different owner and a different fix, and a score alone would not separate them.
Fixing what we can
For outdated or thin pages, Content Refresh proposes diff-style updates to existing pages, and content gap analysis lists prompts where coverage is weak. For third-party sources, we look at whether a correction request, an updated listing or a contributed article is realistic. We do not pay for mentions, and we never try to write fake reviews. If a source is unfair, we say so to the client and decide whether it is worth engaging.
The Action Board keeps these as tasks with owners. Its Discovery agent suggests up to five evidence-backed actions per weekly run.
What we report
Monthly, we send:
- Sentiment trend by engine and by prompt type.
- The five most negative responses, quoted in full.
- Cited sources behind them.
- What changed and what we did.
- Competitor sentiment on the same prompts, from the competitor comparison views.
PDF reports can be scheduled from any dashboard, so the method and the numbers stay together.
Other approaches we see
PR teams often start from media-monitoring suites that have added AI answers, such as Meltwater, Muck Rack or Sprinklr. That makes sense when communications already lives there. These suites start from media and social data. They are less likely to show crawler logs or conversions from AI referrals. Social listening tools with AI add-ons, like Brand24, work the same way.
What it cannot do
Sentiment tracking cannot make an assistant say something kind. It cannot show what every user sees, and a positive average can mask one damaging prompt. We tell clients to read the underlying answers, not only the chart. If you are setting this up, Promptwatch is where we would begin, and the free Explore tier covers ChatGPT with 10 prompts for a first look.