Customer feedback sentiment analysis: what it tells you

Sentiment analysis promises to tell you how customers feel. Run your reviews or survey comments through a tool and out comes a number: 72% positive, perhaps, or an average score of 0.4. It looks precise. On its own, it is close to useless.

This guide explains what customer feedback sentiment analysis is good for, where it misleads, and how to set it up so the result actually tells you what to do.

What sentiment analysis is

Sentiment analysis reads a piece of text and decides whether the writer feels positive, negative or neither about what they are describing. Older tools did this by counting words from a list ("great" scores up, "terrible" scores down). Modern language models read the whole sentence, so they handle negation ("not bad at all"), context and longer comments far better.

Whichever method you use, the output answers one question: how does this person feel? It does not answer about what? That second question is the one you need to act on.

What a single sentiment score hides

Imagine a hypothetical café gets this review:

Coffee is the best in town and the staff are lovely, but the wait at lunchtime is ridiculous and the tables are always sticky.

Score that review as a whole and you get something like "mixed" or "slightly positive". Average a few hundred reviews like it and you get a steady, mildly positive number that never moves. Meanwhile, two specific problems (lunchtime wait, cleanliness) sit inside it, unmeasured.

A single overall score hides:

• What people are praising. You can't protect what you can't see.
• What people are complaining about. You can't fix what you can't name.
• Changes that cancel out. If coffee gets better and service gets worse in the same month, the overall score stays flat.
• The size of each issue. A complaint in one review in fifty and a complaint in one review in five look the same in an average.

Score each mention, not each review

The fix is to attach tone to each mention of a topic, rather than to the review or reply as a whole. The café review above becomes four rows:

Now the review is useful. Across hundreds of reviews, you can see that coffee quality is consistently praised, that lunchtime wait is the most common complaint, and whether either is moving.

This is why topic and tone belong together. Topic tells you what; tone tells you how it is going. Neither is much use alone. If you haven't built a topic list yet, our guide to analysing open-ended survey responses covers how.

The four tones, and why "mixed" matters

Most tools use three labels: positive, negative, neutral. It is worth adding a fourth.

• Positive. The customer is pleased with this thing. "The bed was so comfortable."
• Negative. The customer is unhappy with it. "The bed was too soft and I woke up with a sore back."
• Mixed. Both, about the same topic. "The pool was beautiful but always too crowded to use."
• Neutral. A mention with no clear feeling. "We had breakfast in the hotel."

Mixed is where a lot of the useful detail lives. A topic with many mixed mentions is one where you are getting something right and something wrong at once: the pool itself is good, the access to it is not. That points straight at a fix.

Neutral mentions are worth keeping but not worth reporting on their own. They tell you a topic came up; they don't tell you how it went. When you calculate tone for a topic, look mainly at the balance of positive against negative.

Where sentiment analysis gets it wrong

Any reader, human or machine, gets some comments wrong. Know where the errors cluster so you can check those cases.

Sarcasm and irony. "Oh great, another forty-minute wait for a table" reads as positive to a naive tool. Modern models catch most of these but not all, especially in very short comments with no context.

Polite complaints. Some customers soften everything. "It might be nice if the Wi-Fi reached the upstairs rooms" is a complaint, gently made. A tool that only looks for strong words will call it neutral.

Comparisons. "Better than last year" is positive about the change but may still describe something that isn't good. "Not as good as your other location" is negative about this one.

Short replies. "Fine." "OK." "Could be better." These carry little signal. Treat them with caution and don't let a flood of them drag a topic's tone around.

Domain words. In some trades, words carry a meaning a general reader misses. At a summer camp, "the kids were exhausted" may be praise.

Ratings that disagree with the text. A five-star review that complains, or a one-star review that praises the staff and blames the weather. Trust the text for tone and use the rating separately.

The practical answer to all of these is the same: spot-check. Open ten mentions for each of your biggest topics every so often and confirm the tone is right. If a tool won't show you the replies behind its numbers, you can't do this, and you shouldn't trust the numbers.

Measuring tone over time

Once each mention has a tone, you can track how a topic is going. Two measures work well together:

1. Share of conversation. The percentage of all replies in a period that mention the topic. This tells you how much customers care about it, and it isn't distorted by busy and quiet months the way raw counts are.
2. Tone balance. Of the mentions of that topic, what proportion are positive and what proportion negative.

Watch for these patterns:

Avoid reporting raw mention counts as a change. If you received twice as many reviews in summer, every topic will appear to double, and none of that tells you anything about the business.

Sentiment and your ratings together

Sentiment from text and scores from questions answer different things, and they work best side by side. A rating or a recommend score question tells you how satisfied a customer is overall. Topic-level tone tells you why.

A useful exercise: take every reply from customers who gave you a low score and look at which topics carry negative tone in those replies. Those are the topics most closely tied to unhappy customers. Do the same for your highest scores to see what your happiest customers single out.

Common mistakes to avoid

• Reporting one overall sentiment percentage as the headline. It sounds informative and rarely is.
• Treating neutral as good. Absence of complaint is not praise.
• Comparing tone across very different sources without care. Public reviews skew towards strong feelings; private surveys catch the quieter middle. Look at them together, but be aware of the difference.
• Trusting tiny topics. A topic mentioned three times this month can swing from all-positive to all-negative on one reply. Wait for enough mentions before drawing conclusions.
• Never reading the replies. The numbers point you somewhere. The words tell you what is happening there.

Doing this with Userforms

Userforms gives every mention its own tone as replies and reviews arrive, and files it under a topic in a hierarchy you can rename, merge and split. Forms, website widget replies, imported reviews from Google, TripAdvisor, Yelp and Trustpilot, CSV uploads and pasted text all sit in the same library.

The dashboard shows which topics grew or shrank in share of conversation and which turned better or worse against the previous period, and every chart opens onto the replies behind it, so you can check the tone for yourself. For reviews in particular, see our guide to Google reviews analysis.

You can start for free, or see what each plan includes on the pricing page.