Video Survey Responses: How to Analyze What Nobody Watches

Video survey responses pile up after fieldwork and few get watched. What a video answer carries, why a sentiment score falls short, and how to code them.

The concept test is a normal online survey: screener, exposure, ratings, purchase intent. Near the end, one question asks the respondent to turn on the camera. They explain, in their own words, why they gave that score. Fieldwork closes on Friday. On Monday the platform shows the closed questions as tables and the video survey responses as a folder of files. The tables go into the report that week. The folder waits for someone with a quiet afternoon.

Key Takeaways

• Forsta's guidance for video open ends expects usable videos from 5 to 10 percent of survey completes. On a 1,000-complete study, that is 50 to 100 videos.
• The same guidance sets a goal of about a dozen clips for the presentation, so most of what respondents record is never meant to be analyzed.
• A 2019 scientific review found that facial movements do not map reliably onto emotion categories. A sentiment label read off a face is a weak code.
• Transcription with a confidence score and a timestamp for each segment lets every code point back to the second it came from.
• Coding video answers is qualitative analysis at the volume of a quantitative study. It needs a codebook, a reviewer and a written rule for what counts.

What a video open end is, and what it was sold to do

A video open end is a survey question that asks the respondent to record a short video answer instead of typing one. It sits inside an otherwise quantitative questionnaire. There is no moderator. The respondent reads the question, records, and moves on to the next screen.

Suzy describes the format as a way to capture tone, expression and body language from hundreds or even thousands of respondents at once. QuestionPro launched its own video question in February 2026. Its announcement says the platform transcribes the video and analyzes sentiment and emotional signals, with the results shown next to the quantitative data. The pitch is the same in both cases: the closed questions say what, the video says why.

That pitch is fair. The trouble starts at the next step. A video only explains a score if someone turns it into something the analysis can count.

How the folder of videos gets handled today

PERSONAL EXPERIENCE In the studies I have worked alongside, the video question is almost always the last thing anyone opens. The tables are due first. When someone finally gets to the folder, it is usually the most junior person on the team. They watch at double speed with a spreadsheet open and mark the files that might work in the deck. The rest get a glance, then nothing.

Forsta's own best practice guide for video open ends is candid about the goal. It says participation tends to be modest, with quality videos from 5 to 10 percent of survey completes. It tells researchers to aim for "a dozen snippets" to illustrate the survey data in a presentation. A representative set of videos is explicitly not the goal.

So the video question is designed as a data source and treated as a clip library. Respondents spent three to five extra minutes on the interview, which is the length increase Forsta reports for those who take part. Most of that time ends up as files nobody plays.

The folder grows with the sample. The clip goal does not.

A bar chart in videos. Using the rate in Forsta's best practice guide for video open ends, usable video answers come from 5 to 10 percent of survey completes. For a study of 500 completes that is 25 to 50 videos, for 1,000 completes 50 to 100 videos, and for 2,000 completes 100 to 200 videos. Each bar shows the 5 percent value in solid blue and the extension to 10 percent in a lighter dashed segment. A dashed orange vertical line marks about a dozen clips, the goal the same guide sets for the presentation, and it stays in the same place for every study size. The study sizes are illustrative and the video counts are arithmetic on the published rate. A closing line says watching every video stops scaling long before the sample does. None of the numbers is a Cassi.ai measurement.

At the published participation rate, the number of video answers grows with the sample while the number of clips that reach the deck stays at about twelve.

What a video answer carries that the transcript drops

A transcript keeps the words. The recording keeps more, and the extra is exactly what the format was bought for. Consider a respondent who says the pack is "fine" while turning it over twice, looking for the price.

Forsta's guide says visual and audio cues convey emotion in a way text cannot. The last row is where that needs a limit. Lisa Feldman Barrett and four co-authors reviewed the evidence in Psychological Science in the Public Interest. They found substantial variation in how people express emotion across cultures and situations, and even within a single situation. One configuration of facial movements often signals something other than an emotion. A scowl at a phone camera can be concentration.

Why a sentiment score is not an analysis

UNIQUE INSIGHT The open question sits in the survey to explain a closed one. A respondent gives the concept a 4 out of 10 and is asked why. The brand needs a reason, such as too expensive for the size, did not believe the claim, or already uses a competitor. Sentiment returns a direction, and the 4 out of 10 already said negative.

QuestionPro's announcement promises no manual coding and no post-fieldwork delay. For a tracker that needs a quick read on mood, that can be enough. A concept test, a positioning study or a pack test needs more. There, a positive or negative label per video is the one thing the analysis already knew. The reasons are still in the folder.

The volume problem is real, though, and it grows with the sample size. At Forsta's rate, a 1,000-complete study produces 50 to 100 videos. A 2,000-complete study produces up to 200. Nobody codes 200 videos by watching them at double speed in the week the report is due.

What transcription with confidence makes possible

Speech-to-text transcription is the automatic conversion of recorded speech into written text. A confidence score is the number the transcription engine attaches to each segment to say how sure it is of the words. Diarization is the step that works out who is speaking when there is more than one voice.

Microsoft's documentation for Azure AI Video Indexer shows what a usable output looks like. Each transcript line comes back with its own confidence score, start time and end time. The published example line carries a confidence of 0.8879. The open research model WhisperX goes further on timing. Its authors report word-level timestamps for long recordings and a twelvefold speedup from batched processing.

Two things follow for the analyst. First, low-confidence segments can go to a person who listens to those few seconds, instead of someone listening to everything. Second, every code assigned later can carry the timestamp of the words behind it. A client who doubts a theme can then see the exact moment in the exact video.

Coding video answers is qualitative work at quantitative volume

A codebook is the written list of codes for a question, each with a definition, an example and a rule for when it applies. Coding is assigning those codes to each answer. Both are routine for typed answers. The difference with video is that the evidence for a code may be in the frame and not in the words.

The methodology has to be on paper before anyone opens a file. This is the sequence we recommend for a video question in a quantitative study.

1. Write the codebook before fieldwork closes, from the closed question the video is meant to explain.
2. Transcribe every video and keep the confidence score and timestamps for each segment.
3. Send every segment below an agreed confidence threshold to a person who listens to it.
4. Code each answer against the codebook, keeping the timestamp of the passage or frame behind each code.
5. Mark separately what was said, what was shown on camera, and what was inferred.
6. Have a second person code a sample, and compare the two before coding the rest.
7. Cross the codes with the closed questions, so the reasons can be read by score, segment and market.

From a video answer to a code you can check

A flow in four steps, each with an exception to the right. Step one, transcribe every video and keep the confidence score and the start and end time of each segment. If a segment falls below the agreed confidence, a person listens to those seconds and corrects the text. Step two, code against the codebook, with each code carrying the timestamp of the words or frame that justify it. If no code fits the answer, flag it, because the codebook may be missing a reason. Step three, separate what the respondent said, what the camera showed and what the analyst concluded. If an emotion is read off the face, record it as inferred, never as said. Step four, a second coder codes a sample before the rest is coded. If the two coders disagree, rewrite the code definition, then continue. A closing band says to cross the codes with the closed questions, so the reasons behind each score can be read by segment and market. The figure contains no measured numbers.

The person reviewing the work listens only where the transcript is unsure, and every code can be traced back to the second it came from.

Step 7 is where the video stops being a clip library. Once each answer carries codes, the reasons behind a 4 out of 10 can be cut by the same banner as the rest of the study. For example, the analyst can show which reasons dominate among rejecters in one region, with a clip behind each. The chain from code back to evidence is the one we described for interviews in qualitative coding without losing the evidence. And the coded reasons belong next to the scores they explain, in the same dashboard built from the decision.

Who answers for the folder nobody opened

The proposal promised the voice of the consumer. The client approved the extra cost of the video question and the extra minutes per interview. Then the debrief shows twelve clips and a sentiment chart. The insight lead is the one who has to explain what happened to the other eighty videos.

The respondents pay a cost too. A study in methods, data, analyses found that fewer respondents said they would answer by video than by text. The main reasons they gave for refusing to record were privacy and data security concerns. The people who did record crossed that line. Leaving their answers unwatched spends that trust and returns nothing for it.

Where Cassi.ai comes in

@Cassi.ai is a software engineering company specialized in the pains of market research, innovation and insights. It was founded by Cassiano Albuquerque after twenty years in market research. We automate research projects with or without language models.

Video Lab is a Cassi.ai product that analyzes ethnographic and in-home test videos and returns behavioral KPIs, with a timestamp and a frame as evidence for every conclusion.

Our qualitative work starts from material in whatever form the client sends it: video, audio or transcript. When the material is media, the transcription happens inside the system. Video Lab is that work applied to footage. It takes in the full batch of videos and separates what each respondent said from what the camera showed. It codes each answer with the timestamp behind it. Then it hands back the coded base, the deck and the dashboard together.

The analyst still owns the codebook and the final reading. What changes is that the whole folder gets analyzed. The dozen clips for the deck are then chosen from coded answers, not from whatever someone happened to watch first. A general chat assistant does not do this on its own, for the reasons in why a general chatbot is not a research platform. More in the cassiai.com portfolio.

FAQ

What are video survey responses?

Short videos that respondents record inside an online survey to answer an open question in their own words. Suzy and QuestionPro both offer the format inside quantitative questionnaires.

How many respondents actually record a video answer?

Forsta's best practice guide expects usable videos from 5 to 10 percent of survey completes. It says an engaging topic or an extra incentive can raise that.

Is sentiment analysis enough for video open ends?

It gives a direction, not a reason. The closed question usually shows the direction already. A 2019 review of facial expression research also warns against reading emotion straight off the face.

How do you analyze video survey responses at scale?

Transcribe with confidence scores and timestamps, and send low-confidence passages to a person. Then code against a written codebook and cross the codes with the closed questions.

Should the transcript replace the video?

No. Objects, setting and hesitation live in the recording. Keep the timestamp on every code so the analyst can return to the moment, as tools like Azure AI Video Indexer allow.

Before the next video question goes into a survey

UNIQUE INSIGHT A video question that nobody codes is a clip request with a sample attached. The respondents who recorded gave the study the richest answer in the questionnaire. That answer only counts as data when every video is read against the same codebook.

If your last study with a video question ended with a dozen clips in the deck, count the videos nobody opened. That number is the size of the analysis the study paid for and did not get.

About the author

Cassiano Albuquerque is the founder of @Cassi.ai. He spent twenty years in market research operations before building software for research teams. To talk about how the video answers in your studies are coded, contact him on LinkedIn.

Published by Cassi.ai. Read the full article at https://www.cassiai.com/blog/video-survey-responses-analysis.