A camera can record a shopper’s brow movement. A model can label what it sees. If you need to know whether someone understood the price or chose the product, give them a task that lets you check that decision. Use facial coding only where the movement can add something to what you learn from that task.
Orbital’s October 2025 company presentation lists facial coding alongside eye tracking, physiological measures and customer research. The evidence below concerns how people interpret faces and how they respond to advertising. A proposed Jamaica test puts those distinctions to work.
Changing a face in an ad tests how viewers respond
In a paper made available on 15 October 2026, Maria Giulia Trupia, Martina Cossu and Zachary Estes describe six preregistered studies of advertising models’ pupils. The authors’ abstract reports more favourable viewer responses to models with smaller pupils, with a weaker effect when the model smiled. Some experiments asked viewers to make choices with incentives. The article is a journal pre-proof.
The feature being changed was on the face in the ad. The viewer then gave a rating or made a choice. That setup answers a different question from recording the viewer’s own face and trying to tell how they felt. A brand team can use the first setup to compare versions of an ad. It needs its own evidence before using the second to infer a customer’s experience.
Give each claim the name of what you measured
The Facial Action Coding System describes visible facial movements. In a 2019 review, Lisa Feldman Barrett and colleagues show why people, settings and cultures matter when you interpret a movement. Faces carry social information, and a given movement can have several meanings. Record the movement before deciding which interpretation to test. Read the review and its correction.
In Study 1, Goel and colleagues used 7,852 pairs of observations to compare models with people’s ratings of a pictured face and a described situation. A model using just the face fit those ratings less well overall. Adding the face to a model of the situation helped with some emotion categories and weakened the fit for others. The ratings came from observers. Read the study, Figures 2 and 3.
A separate study used pictures taken as tennis players won or lost a point. Of the 15 observers who saw both face and body, 53.3% said they had relied on the face. The group that saw only the faces did not reliably distinguish winning from losing by how positive the face looked. The researchers show why a report should distinguish what a person says they relied on from what the test itself measured. Aviezer et al., Experiment 1. Princeton research record.
Preston and Page drew about 3.8 million frames of viewers watching video ads from a database of over 70,000 commercial studies. The authors, who were affiliated with facial-coding vendors, report that some patterns appeared across regions. The count describes frames, which can include several moments from one viewer. Recurring movements are useful to document without treating the viewer’s inner state as known. Read the published study abstract.
Naidoo and Hamelin paired ad footage with webcam recordings from about 250 university students. In a peak-timing comparison, the authors report that the viewer’s strongest moment fell within a second of the actor’s top 5% software-scored peak in roughly 58%–65% of cases, against a reported 33% chance baseline. They also asked viewers to mark when they felt most emotional and whether they would comment or like the ad. A brand deciding whether an ad sells a product still needs a purchase measure. Naidoo and Hamelin, sections 3.2 and 4.3.
Design the test around where the shopper will see the offer
Suppose a team wants to know whether shoppers in Jamaica understand a price and a condition printed on a package. Decide which Caribbean shoppers the package is intended to serve and where they will see it. Recruiting one group from a store near the office may miss people who encounter the offer in a different setting. If Brown’s Town is part of the intended market, recruit there for a reason you can write down.
Show the package under conditions that match how people will read it. Ask each shopper to explain the price and condition before you prompt them with an answer. If an analyst records a brow movement while the shopper reads the price, keep the movement and the shopper’s explanation in separate fields. Ask the person what happened at that point rather than supplying a label such as confusion.
Different sites in the Caribbean can also require different terms or languages in a task. Check the words with local staff and use the language the participant expects when reading the offer. Record where and how the task took place. Without that information, a difference between sites could be a change in the research setup rather than a difference among shoppers.
This applies Adrian Dunkley’s People First, Problem Second, AI Last method. First decide what the shopper needs to understand. Once you have a way to test that, decide whether footage helps the research team diagnose a mistake.
Make the vendor show what the score can support
Ask whether the model detects a visible movement, predicts how another person would label a clip or estimates the participant’s answer. Request a test on recordings from people the model did not encounter during training. Keep recordings of the same person together when the team separates training and evaluation files, or the evaluation may be easier than a new recording from a new shopper.
Goel’s team retained 268 of 303 people recruited for its fourth US study, about 88.4%. The researchers excluded cases based on attention checks, rating quality and a test for unusual responses. The result describes which ratings made it into that analysis. For your own study, keep separate counts for people who completed the task and those whose footage you could score. Goel et al., Study 4 methods.
Agree on what to do with a blocked camera view and a person leaving the frame. The report should mark unusable footage and show how often it occurs. Tell participants who can see their recordings and how long the team will keep them. Publish only authorised aggregate results. A response to a package gives you no basis for assigning a personal trait to the shopper.
Audit the next research brief
- Pick 5 claims in a report. For each one, write down the task and the evidence the report uses to support the claim. Keep identifiable clips inside the approved study system.
- Check whether a colleague can trace each claim to a movement, a model label, the participant’s answer or a recorded choice. Mark any claim the colleague cannot trace.
- Test the price and offer condition with a small group before recruiting for the larger study. Ask each person to state what they understood, then change the task if a prompt gives away the answer.
- Compare any recommendation that relies on video with what the team would recommend from the task log and shopper responses. Keep the footage only if the difference justifies the extra work and the team can explain how it protects the participant.
An Orbital research assistant could retrieve the task, its measured result and the study note supporting a recommendation. The same audit gives a researcher a reference for checking whether the assistant retrieved the right record. The checks above are proposed for a Caribbean brand team. Any effect on review time or accuracy would have to be measured separately.
Questions that come up when you commission facial coding
Can I use facial coding to select clips for an interview?
Use an agreed feature such as a visible brow movement to find a clip. Show the participant the full moment and ask what they were doing or thinking. Record their answer in their own terms and let them correct how the clip is described.
Can I use a still-image model to score a moving ad?
Treat video as a separate test. Show how the tool handles changes in framing, faces moving out of view and moments where the cut hides the start of a response. Ask whether the training labels cover those conditions before you interpret a trend across time.
What if a shopper agrees to the test but declines recording?
Offer an unrecorded version when the study allows it. Keep that shopper in the comprehension or choice results and mark their video as unavailable. That preserves what they contributed and helps the researcher see whether permission to record changes who appears in the facial-coding sample.
What would justify paying for the facial data?
Decide what better work would look like before you pay. A team that uses video to find interview moments could test whether two reviewers using the video agree on more useful clips than reviewers using the task log alone. Include the extra review and storage cost.
Sources and research status
This is a narrative research explainer by Adrian Dunkley. The test described here is a proposal for Jamaica. The two archive photographs illustrate retail settings. The model-fit chart uses published category comparisons and reports no Orbital experiment or result.
See Orbital’s guide to testing attention, preference and purchase and its research-assistant service brief. Related work from Adrian Dunkley, StarApple AI and the Caribbean AI Association.