Sentiment in AI Answers: What We Act On and What We Ignore

Sentiment scores are the easiest part of an AI visibility dashboard to over-read. A green number feels like a clean bill of health. A red one sends a client into a panic. Neither reaction is based on much.

Here is how we use sentiment in practice. It is less about the score and more about what sits behind the lowest-scoring responses.

What a sentiment score is

In Promptwatch, each response gets a sentiment score between 1 and 100, and there is a weekly AI summary of what the responses say about the brand. That gives you a number per response and a trend across responses. It does not give you a verdict on the brand. It gives you a lens on how a particular engine phrased a particular answer on a particular day.

Treat it that way and it becomes useful. Treat it as a reputation index and you will chase noise.

Start with the bottom of the list

We ignore the average at first and sort responses from lowest to highest. The bottom ten tell us more than the mean ever will. For each one we ask:

• What was the prompt? A comparison prompt that asks about downsides will produce critical text by design.
• What did the answer actually say? Sometimes the criticism is a fair point about price or a missing feature.
• Which sources did it cite? A harsh line backed by one old forum thread is a different problem from a harsh line backed by five reviews.
• Is the claim true, outdated, or wrong?

That last question sorts almost everything into one of three piles.

True and fair

The answer says the product lacks a feature, and it does. There is nothing to fix in the AI layer. This goes to the client as product feedback, and occasionally to the content team as a prompt to write honestly about the trade-off. Honest positioning often reads better to an engine than evasive copy.

Outdated

The answer repeats something that used to be true. A pricing structure that changed, a feature that shipped, a limitation that was fixed. This is the most fixable pile. We find the sources the engine cited, check their dates, and do three things: update our own pages with clear, dated statements, request corrections where the third party allows it, and publish a page that states the current fact plainly. The Page Tracker is handy here because we can watch the outdated page and the corrective page together.

Wrong

The answer says something false. These are rarer than clients fear, and they usually trace back to a specific source, a confusion with a similarly named company, or thin coverage of the brand. We document the exact prompt and response, find the cited sources, and fix what we can fix. Where there is no source, the lack of clear, quotable facts on the client's own site is often the cause. Entities with sparse information invite engines to guess.

Do not chase a single response

AI answers vary. One unflattering response in a hundred is not a trend. We look for patterns: the same criticism appearing across prompts, engines, or weeks. A pattern means a source, and a source is something you can act on.

We also try to resist the temptation to game sentiment. Stuffing a page with superlatives does not help and can backfire if engines or readers see through it. What works is a clear answer to the objection. If buyers keep asking whether a product is hard to set up, a page that explains the setup in plain steps, with accurate timing, is better than a banner that says it is easy.

Sentiment by engine and by persona

Engines weigh sources differently, so a brand can read warmer on one engine than another. We break sentiment out by model before we blend it. Persona matters too. A prompt phrased as a budget-conscious buyer may surface price criticism that a prompt phrased as an enterprise buyer never does. Promptwatch lets us run prompts under different personas, and persona stacking combines them, which is useful for checking whether a negative theme is universal or tied to one kind of buyer.

What we put in the client report

A good sentiment section is short. Ours has four parts:

1. The trend across the period, with the caveat about variance.
2. The main theme from the weekly summary, in plain language.
3. The two or three responses we looked at by hand, quoted briefly, with the cited sources.
4. The action: what we will update, what the client needs to decide, and what we will leave alone.

We do not include a scorecard of every response. Clients stop reading after the first page of those.

What we ignore

• Small week-to-week swings with no change in sources.
• Negative sentiment on prompts that invite negativity, like "problems with X".
• Single responses with no cited source and no repeat.
• Anything we cannot trace to a page we could change or a source we could influence.

Ignoring things deliberately is part of the job. If every wobble becomes a task, the Action Board becomes a graveyard and nothing gets done.

Connecting it to the work queue

When a sentiment pattern is real, we put it where the team will see it. The Action Board in Promptwatch holds a prioritized list of GEO actions and lets us assign owners and add comments. We add a custom action for a sentiment fix just as we would for a content gap, with the evidence linked. A weekly Discovery run can also suggest evidence-backed actions, which we review rather than accept blindly.

Our recommendation

Use sentiment to find sources, not to grade the brand. Sort from the bottom, trace each response to a cited page, sort it into true, outdated, or wrong, and act only on patterns. We do this inside Promptwatch because the weekly summary, the response list, and the citation data are linked. You can see the platform at promptwatch.com.