Sentiment analysis has been part of the comms toolkit for more than a decade. It has come a long way, and it is genuinely useful. It also tends to carry more weight in reports than it can reasonably support on its own.
Sentiment analysis is not broken. What has changed is its role. As a standalone metric it tells you less than it used to, and a single positive or negative percentage rarely gives a comms team enough to act on.
Here is what it does well today, where vendors still oversell it, and what to use alongside it.
What sentiment analysis actually does well
For text, in the major languages, it works. Most social listening and media monitoring platforms run on language models, and the accuracy of sentiment analysis on English content is strong. Some vernacular languages such as Mandarin has caught up well over the past few years.
The value is speed. Sentiment analysis lets a practitioner look at a piece of news, a post, or an article and decide in a split second whether it needs attention. For a team processing thousands of items a day, that first cut matters.
So the base capability is real. The question is what buyers assume it covers.
Four things sentiment analysis still cannot read
The accuracy of sentiment analysis holds up in a clean English demo. Here is where it tends to slip in the real world.
- Local language and context. In English and Mandarin, accuracy is strong. In Tamil, Hindi, Bahasa Malaysia, Bahasa Indonesia, Vietnamese, or Thai, it drops off. Buzzwords, local lingo, and dialect still trip up models that look sharp in a demo. Singapore alone has four major languages, and the region adds many more.
- Content that is not clean text. On visual platforms like Instagram, a video can cover one issue while the caption says something unrelated. Your brand might appear in the footage or on a logo while the text gives no signal, which makes the content hard to even pull in, let alone score.
- Ambiguous signals. A thumbs up can mean agreement or disagreement depending on the post. An emoji shifts meaning by context, and a laughing reaction on an unfortunate incident is not the same as one on a joke. Machines still struggle to read these the way a person would.
- Coordinated behaviour. Sentiment cannot tell you whether the posts driving a spike are genuine. We have seen coordinated activity across lookalike accounts pushing narratives that do not reflect authentic voices, and it becomes very visible around social and political issues. If you are making decisions on media sentiment tracking, you need to know the data is real before you act on the score.
The pace of AI development may close some of these gaps in the next year or two. They are not closed today. This is also where the work is heading. At Truescope we are building capability to detect coordinated information campaigns across similar accounts on X, Facebook, TikTok, and Instagram, so teams can judge whether a signal is worth trusting.
Sentiment is one metric, not the whole picture
The bigger shift is this. Sentiment is one dimensional, and there is a wealth of information around every media item that most teams ignore.
Look at the metadata. A post with more shares than likes tells you something different from one with more comments than shares. It signals how people are engaging, advocating, or pushing back. Velocity of engagement matters even more, because that is what tells you whether a story is about to go viral and whether you need to alert your client now rather than tomorrow.
Sentiment is part of that metadata, not a substitute for it. Brand sentiment monitoring gets far more useful when the score sits alongside context, volume, reach, and velocity rather than standing in for them.
Benchmark against your own data, not the industry
One more thing we coach clients on. Your thresholds should be calculated from your own historical data, not borrowed from a peer or an industry average.
A thousand comments means very little to a large consumer brand and a great deal to a niche one. Finance and logistics do not live on the same scale. So the right approach is to take three to six months of your actual data, plot the distribution, and set your own bands. As a rudimentary example, the bottom of the curve is low traction, the middle is medium, and the top decile is very high. Then map those bands to the issues that are actually relevant to you.
This matters most in the public sector, where sentiment is scrutinised closely. A minister will see a report, compare it to another agency's numbers, and ask why one story is high traction and another is not. If your benchmarks are grounded in your own data and a proper method, you can answer that with confidence.
Give leaders something they can act on
Here is the test I apply to any sentiment report headed for the board. If you tell your CEO that ninety percent of people are negative right now, what do they actually do with that?
Ninety percent negative is not a decision. The job is to bridge the sentiment marker to something a leader can act on.
Say a shared public service breaks down and most of the commentary is negative. Do not stop at the percentage. Ask what sits underneath it. What proportion is calling for a boycott of the service. Is there a narrative forming around fares. Are people questioning leadership directly. Those are the questions a CEO can make a decision on. AI and good analysis should be doing that bridging work, moving from a raw score to an insight that connects to what the leadership team cares about.
The question to ask a vendor
Most buyers ask a vendor demoing sentiment analysis whether it is accurate. That is the wrong question. Almost everyone runs on a language model, and on text the variation between them is not large.
The better question is whether you can train the model to your own organisation. Can you feed it examples of what positive and negative look like in your context. Can you personalise it and improve it over weeks with machine learning, so it ends up sharper than any off the shelf model a competitor is using. That is where the real difference in AI sentiment analysis shows up, and it is what separates a tool that scores generic sentiment from one that understands yours.
There is also a level of nuance worth asking about. Mixed sentiment, at the article level, captures content that holds both positive and negative views within the same passage. Most vendors file that as neutral, which flattens a polarising story into a non event. At the sentence level, you can catch a positive article that mentions your brand negatively, which is exactly the kind of thing you want to be alerted to without drowning in every negative story that has nothing to do with you.
How to make sentiment earn its place
Sentiment analysis is useful. It is just insufficient on its own. Treat it as one input among several, benchmark it against your own history, bridge it to decisions your leaders can act on, and push your vendor on customisation rather than accuracy. Do that, and the score starts earning its place in the report again.
If you want to work out what belongs in your measurement setup, and what to use alongside sentiment, our team can help.
