AI Answer Sentiment: What a 1 to 100 Score Tells You and What It Doesn't

Getting mentioned is half of AI visibility. The other half is what the answer says once it names you. "A solid option for small teams" and "a tool some reviewers find limited" both count as a mention. They do very different things to a buyer.

Sentiment scoring exists to separate them. This post explains what a per-response score is, how to read a trend, and where the number can lead you astray.

What the score is

In Promptwatch, each tracked response gets a sentiment score from 1 to 100 for the brand in question, and a weekly AI summary condenses the pattern. The score applies per response, so you can see a distribution instead of a single average. That is useful because averages hide the bad answers that matter most.

A score tells you the tone about a brand in one answer. It does not tell you whether the claims are true. That is a separate check, and one a human has to make.

Read the distribution, not the average

An average of 70 can come from many answers at 70, or from half at 95 and half at 45. Those are different situations. The second one means some prompts produce glowing answers and others quietly warn people off, and the second group is where your work is.

Look at the distribution and then at the low-scoring tail. For each low-scoring response, ask three questions.

What claim drove the low score? Price, a missing feature, support, reliability, fit for a segment. Write the claim down in plain words.

Is it accurate? If yes, the fix is on the product or messaging side, not the content side. If no, you have a misinformation problem to trace.

Where did the engine get it? Check the cited sources for that response. A single outdated review or forum thread often explains a whole cluster of negative answers.

Trace the source

This is the step that turns sentiment from a score into a task. Pull the citations for the low-scoring responses and look for repetition. If the same third-party page appears in most of them, the page is your lead.

Promptwatch's citation analytics help: page-level and domain-level citations, offsite mentions, and Reddit and YouTube reports. The Page Tracker can follow the specific offending page so you see when it changes, and whether the answer follows.

Common patterns:

• An outdated comparison page. The facts about your product are old. Contact the author with current, verifiable information.
• A forum thread. Engines cite them, and the discussion may be fair. Participate honestly if you can add something real, and do not astroturf.
• Your own old content. Sometimes the negative claim comes from your site, for example an archived page describing a limitation you have since fixed. Update or retire it.

Beware of false precision

A score from 1 to 100 looks exact. It is a model's judgment of tone, and tone is fuzzy. Treat differences of a few points as noise. Treat a sustained decline across many responses, or a cluster of responses under a clear threshold, as a real signal.

Also keep in mind what a Brand Book does here. Promptwatch lets you define analysis instructions that decide what counts as a mention, and with aliases for your brand names. If your brand is a common word, or has a similar-named competitor, loose rules can pull in the wrong responses and distort sentiment. Tighten the rules before you trust the trend.

Compare with competitors, carefully

You can look at sentiment for competitors on the same prompts. A comparison is informative when the prompts are identical and the set of competitors is fixed. It is less so across different prompts, because question type affects tone. Comparison prompts often yield cooler language for every brand than recommendation prompts do.

If you share competitor sentiment outside your team, say that the scoring is a model's reading of tone, not a customer survey. People do not always hear that distinction.

What to do about negative sentiment

Work in this order.

1. Verify the claim. Is it true, outdated or false?
2. If true, tell product. The best fix for a true complaint is not content.
3. If outdated or false, find the source and request a correction with evidence.
4. Publish a clear, factual page on the topic from your side. Engines need something accurate to cite. Content gap analysis can show which questions you do not yet answer.
5. Re-measure after a few weeks, and log the date of each change.

Do not try to flood the web with positive copy. It rarely works, and it creates other problems.

Where this fits in a report

Sentiment sits next to Visibility Score, Share of Voice and Position in Answer. Reading them as four views of one panel is better than treating sentiment as a stand-alone vanity score. A brand with high visibility and falling sentiment has a different problem from one with low visibility and good sentiment.

The weekly AI summary is useful for people who do not have time to read responses. For anyone accountable for the number, read at least a sample of the low-scoring answers yourself every week. Ten minutes is enough.

The recommendation

If sentiment matters to you, and in most categories it should, pick tooling that scores each response, links the score to the cited sources, and lets you define what counts as a mention. Promptwatch covers all three, with the sentiment trend, the weekly summary and the citation view in one place. Start with the free Explore plan at promptwatch.com to see how your brand is described on ChatGPT, then widen to the other engines on a paid plan.