How to read sentiment in AI answers without fooling yourself

Mention counts tell you whether an engine names you. They don't tell you what it says. A brand can appear in forty percent of answers and be described as "an option, though some users report issues" in most of them. That is a different outcome from being the recommended pick, and sentiment tracking exists to separate the two.

It's also an easy metric to misuse. Here is how we read it.

What a sentiment score is

In Promptwatch, each tracked response gets a sentiment score from 1 to 100, and the platform produces a weekly AI summary of the movement. The score is a model's judgment of how a response treats your brand. Treat it as a screening tool. It points you at responses to read, and it isn't a verdict.

Any automated score has blind spots. Sarcasm, hedged praise and comparisons ("good, but X is better") are hard to rate. A response that's accurate and neutral might score middling and still be great for you.

Start with the Brand Book

What counts as a mention, and what a mention means, depends on how the tool understands your brand. Promptwatch has a Brand Book whose analysis instructions decide what counts, and brand aliases to catch alternate names.

Fill that in before reading any score. If your company shares its name with a common word, or went through a rebrand, the tool can confuse mentions. We've seen sentiment dip purely because answers about a different entity with a similar name were counted in.

Read the lowest responses, not the average

An average hides the pattern. Pull the ten lowest-scoring responses for the period and read them in full. You'll typically find one of a few things:

A factual error: wrong pricing model, discontinued feature, wrong location.
A third-party complaint: an engine paraphrases a forum thread or review.
A comparison loss: you're named, then a competitor is called better for the buyer's stated need.
A mistaken identity: the answer is about a different brand.

Each has a different fix. Factual errors get fixed on your own pages first. Third-party complaints need a look at the source. Comparison losses are content and positioning issues. Mistaken identity goes back to the Brand Book.

Follow the source

A low-sentiment answer almost always has a source behind it. Use citation analytics to see which pages the engine cited for that response. If the source is a Reddit thread, our post on how Reddit threads shape AI answers covers what you can and can't do about it. If the source is your own outdated page, you have your answer and the repair is quick.

Promptwatch has Reddit, YouTube and LinkedIn citation reports and offsite mention tracking, so you can find where a negative framing originates, including pages that name your brand without linking to it.

Compare like with like

A raw sentiment number only means something against a reference. Useful comparisons:

Your sentiment this period against last period, on the same prompt panel.
Your sentiment against competitors' on the same prompts.
Your sentiment by model, since engines differ in tone.
Your sentiment by prompt type: brand, category, competitor comparison.

Brand prompts ("what is X like?") will usually score higher than competitor comparison prompts, because you're the topic. Don't compare across types. The A/B period comparison in Promptwatch is handy for before-and-after checks on a single change.

Don't chase the number

It's tempting to set a target score. We don't. Sentiment responds to what's true about you and what's written about you. Moving it without changing either is gaming, and it tends not to last.

What we do commit to is a loop: find the lowest responses, trace the cause, fix what's fixable, and re-measure. Over a quarter that shifts the number as a side effect. It's slower than a quick content push and it holds up better.

Reporting sentiment to a client or boss

Give the score a sentence of context. "Sentiment held steady this month; the two lowest responses traced to an outdated pricing page, which we've corrected." That's more useful than a chart alone, and it shows someone has read the responses.

Avoid claiming a causal link to revenue. You don't have the evidence for it. What you can show, with visitor analytics, is AI-referred visits and conversions, which is a better measure of business impact than sentiment itself.

When sentiment matters most

Some categories make sentiment more important than visibility: healthcare, finance, anything where buyers ask "is X trustworthy?". There, being named with doubt may be worse than not being named. Run brand-trust prompts ("is X legit", "X reviews", "problems with X") as a dedicated group and check them weekly.

The practical setup

Run the panel daily, read the weekly summary, and open the lowest responses every Monday. Log the cause of each in the Action Board so an owner and a due date exist. In Promptwatch that board is AI-prioritized, with a weekly Discovery agent that suggests up to five evidence-backed actions, and you can attach your own custom ones with assignees and comments.

Sentiment is a pointer. The reading, and the fix, are still yours.