You are standing behind the glass after the debate. One person in the room says the candidate seemed off in the middle. Another one loved the closing. A third spent the whole second half looking at their phone.
Everybody is describing something real. Nobody can tell you which ten seconds actually did it.
That is the whole problem. A speech is not one thing. It is a few hundred moments in a row, and a handful of them are doing almost all the work. The trick is finding those moments instead of arguing about them.
Why the Room Disagrees
Speech teams write for applause lines and clips that will play well on cable. Debate prep goes deep on policy and on not making mistakes. All of that matters and none of it is wasted.
But here is the thing about people. They feel a speech before they think about it. One phrase that lands wrong can quietly ruin the next 30 seconds, even when the argument that follows is a good one. A flash of real frustration can look like passion to one voter and like losing your cool to another, in the same second.
You will never see that split in a single favorability number. A number tells you where you ended up. It does not tell you where you went sideways.
What the Timeline Actually Shows
Here is how it works in plain terms. Opt-in panelists watch your speech or debate clip on their own devices. The webcam reads their expressions frame by frame using the Facial Action Coding System, which is the standard built by Ekman and Friesen in 1978 and updated in 2002. It maps small muscle movements in the face to emotions.
Our platform reports seven of them: happiness, sadness, anger, surprise, disgust, fear, and confusion.
You get a graph. Time along the bottom, emotion strength going up. Then you lay the transcript over it. Now you can point at a sentence, and sometimes at a single word, and say that is where it turned.
I want to be careful about what that does and does not mean, because this is where people in my industry tend to oversell.
What This Can and Cannot Tell You
What the research supports. Höfling and Alpers recorded the faces of 219 people watching video commercials and found that facial expressions significantly predicted what those people reported about the ad, the emotion, and the brand. Published in Frontiers in Neuroscience, Volume 17, 2023.
Read that closely. The study showed faces line up with what people say. It did not show that faces beat what people say. The real advantage is not that the face is smarter than the survey. It is that the face comes with a timestamp. A survey gives you one score for the whole speech. A face gives you second 40 and second 95.
What it cannot do. It cannot tell you someone is lying. It cannot read boredom, or contempt, or skepticism, because those are not among the seven we report. If you hear a vendor promise those, ask which emotion list they are working from.
The hardest question about all of this. In 2019 a team led by Lisa Feldman Barrett published a 68 page review in Psychological Science in the Public Interest asking whether you can read what a person feels from their face. Their answer was mostly no. The same feeling shows up differently on different faces, and the same facial movement can mean different things depending on the person and the moment.
I bring that up on purpose, because it is the strongest criticism of my own industry and you should hear it from me rather than from a competitor. Here is how I think about it. Barrett and her coauthors are right that you cannot look at one face for one second and announce what that person feels inside. But that is not what a campaign needs. You are not diagnosing a voter. You are comparing two cuts of a speech across a couple hundred people and asking which one held them longer. That is a comparison between groups, and comparisons are where this tool actually earns its keep.
Where it gets soft. Automatic facial coding is more accurate on big clear expressions than on small everyday ones. Büdenbender and colleagues documented this in PLOS ONE in 2023. So trust the cliffs and go easy on the ripples. A sharp spike means something. A tiny wobble might just be someone shifting in their chair.
That is still plenty. You do not need a lie detector. You need to know which ten seconds to fix.
Testing Debate Answers Before the Debate
Let me walk through how this would go. This one is made up, but the shape of it is real.
Say your candidate has a soft spot on healthcare and the debate is nine days out. You write three different 90 second answers. One leans on policy. One tells a story. One turns it back on the opponent's record.
Build a panel that looks like the voters you need. Show all three in random order so nobody is just reacting to whichever came first. Watch what the faces do.
Maybe the policy answer tests fine on the questions afterward but happiness slides and stays down through the middle. Maybe the story answer holds people right up until the moment it swings back to policy, and confusion climbs at exactly that seam. Maybe the pivot answer keeps everyone wide awake but spikes anger in a group you were trying to reassure.
Now you have something to work with. Take the story opening, cut the seam that caused the confusion, land it faster. Or decide that for this question, keeping people awake matters more than keeping everyone comfortable. Either way it is a decision instead of an argument.
Reacting Faster After the Event
If you line up your panel ahead of time, you can run this during the live debate. People watch from home with the camera on. You have the timeline within about an hour of the closing statement.
That is early enough to matter. It shapes what your surrogates say tonight, what you post, and what your candidate leads with on the morning shows. If the closing statement lifted happiness among the voters you need, that clip goes everywhere. If a specific attack from the other side spiked fear or anger, you now know which soundbite to answer and which one to let go.
You can also see how different groups reacted to the same second. Maybe the joke landed with younger voters and did nothing with older ones. That tells you where to run the clip and where to lead with something else instead.
Where Trust Is Built and Lost
Trust is not one big thing. It gets built in moments and it gets lost in moments. A candidate who looks down while making a promise. A pause that reads as calculating instead of thinking. A line that is technically true and somehow lands cold.
You will not get a verdict on any of that. What you will get is a signal that something happened right there. If happiness drops and confusion climbs during one answer, you have a problem with that answer.
Sometimes the fix is small. Cut the jargon. Add one real detail. Let the candidate say "I" instead of talking about themselves in the third person. Sometimes the problem is the position itself and you have a bigger decision to make. Either way you are working from something you can point at.
Segment or You Are Guessing
The average across everyone is interesting. The breakdown is what you act on.
Run the same speech past strong partisans, soft partisans, and true independents. The lines that light up your base can leave the middle cold. The lines that reassure the middle can read as weak to your core.
You cannot win everyone, and you should stop trying. What you can do is see the trade clearly and pick on purpose. If this speech is about turnout, you want to see energy in your base and you can live with independents staying flat. If you are fighting for the middle, you want attention held and negative reactions low with persuadables, even if your core is a little bored by it.
That choice gets made either way. The only question is whether you make it with evidence or with vibes.
The So What
A speech is a sequence of moments, and a few of them are carrying the whole thing. You can find those moments before the event, fix them, and check that the fix worked. After the event, you can know within the hour which clip to push.
Next step: Take the last speech your candidate gave and one clip you argued about internally. Run a panel on it. Find the second things turned. That one test will change how your team talks about the next draft.
Try EmotionTrac: See which ten seconds moved people. Start your first panel test at politics.emotiontrac.com.
Sources
- Höfling, T.T.A. and Alpers, G.W. (2023). Automatic facial coding predicts self-report of emotion, advertisement and brand effects elicited by video commercials. Frontiers in Neuroscience, 17:1125983. https://doi.org/10.3389/fnins.2023.1125983 (PMID 37205049, open access)
- Ekman, P. and Friesen, W.V. (1978). Facial Action Coding System. Revised edition Ekman, Friesen and Hager (2002).
- Büdenbender, B., Höfling, T.T., Gerdes, A.B. and Alpers, G.W. (2023). Training machine learning algorithms for automatic facial coding: The role of emotional facial expressions' prototypicality. PLOS ONE, 18(2): e0281309. https://doi.org/10.1371/journal.pone.0281309
- Barrett, L.F., Adolphs, R., Marsella, S., Martinez, A.M. and Pollak, S.D. (2019). Emotional expressions reconsidered: Challenges to inferring emotion from human facial movements. Psychological Science in the Public Interest, 20(1), 1-68. https://doi.org/10.1177/1529100619832930
- EmotionTrac FAQ, seven captured emotions. https://creative.emotiontrac.com/resources/faqs/
- EmotionTrac Politics. https://politics.emotiontrac.com/