Learning research can inform a demo’s structure, pacing and explanation. This page brings together peer-reviewed academic evidence, with original publication years and links to the authors or publishers. The studies concern educational video and multimedia learning; applying them to product demos is an editorial inference to test, rather than a measured SaaS conversion effect.
Video engagement: an observational starting point
Guo, Kim and Rubin (2014) analysed 6.9M video-viewing sessions across four edX courses. That is an observation count, not unique viewers or a performance uplift. Source: ACM Learning at Scale, DOI 10.1145/2556325.2566239.
Their observational findings associate shorter videos with greater engagement. This does not establish a universal optimal demo length or show that shortening a particular video causes a sales improvement. Application to test: organise the explanation around one task, then measure whether the intended audience can complete it.
Read the video transcript
A play button tells you someone watched. It does not tell you what they learned, or whether they will buy.
Here, a demo focuses on one task. The moving playhead illustrates viewing, without claiming a measured retention curve.
Guo and colleagues analysed 6.9 million viewing sessions across four edX courses. Shorter videos were associated with greater engagement.
This was observational research in education. Test a focused demo with your audience, and measure task success separately from viewing.
Video and learning: replacing versus supplementing instruction
Noetel et al. (2021) reported g = 0.28 for learning when video replaced existing teaching methods in higher education. Source: Review of Educational Research, public abstract.
Noetel et al. (2021) reported g = 0.80 when video was added to existing teaching. Their review included 105 studies and 7,776 students overall. Source: DOI 10.3102/0034654321990713, public abstract.
These are separate pooled comparisons, not a head-to-head test of adding versus replacing video in the same product workflow. Application to test: use a video alongside a task, guide or hands-on exercise, and assess comprehension rather than assuming a viewing event means success.
Read the video transcript
Replacing an explanation with video and adding video to an explanation are two different interventions.
In the first path, video takes the place of existing instruction. In the second, learners keep the instruction and also receive video.
A review of 105 studies reported a learning effect of g = 0.28 for replacement, and g = 0.80 for supplementation.
These are different pooled comparisons in higher education, not a direct contest between the two paths. Neither value is a percentage increase in sales.
Segmentation: retention and transfer are different outcomes
Rey et al. (2019) reported a pooled retention effect of d = 0.32 for segmented multimedia instruction, with a 95% confidence interval of 0.20–0.43. Source: Educational Psychology Review, Table 3.
Rey et al. (2019) reported a pooled transfer effect of d = 0.36, with a 95% confidence interval of 0.24–0.48. Source: DOI 10.1007/s10648-018-9456-4, Table 3.
The meta-analysis combines multimedia instruction, not SaaS demo campaigns. Its outcome-specific comparisons overlap and should not be added as independent samples. Application to test: split a workflow into meaningful stages, then assess both recall of the demonstrated steps and successful application to a new task.
Read the video transcript
A continuous explanation can be divided into meaningful stages. In this illustration, we connect an account, configure an action, and verify the result.
Each boundary offers a moment to process the previous step. The exact timing shown here is an illustration, not a tested pacing rule.
Rey and colleagues reported a standardised retention effect of d = 0.32, and a transfer effect of d = 0.36.
Retention means remembering what was taught. Transfer means applying it to a new task. Test both; neither result measures SaaS conversion.
Active recall during a video lecture
Szpunar, Khan and Schacter (2013) reported 86% vs 72% mean correct recall for the tested and nontested groups in a combined final-test analysis of the fourth lecture segment. The paper contains two experiments with 80 students overall. Source: PNAS, Discussion.
The restudy group’s corresponding mean was 66%. These scores concern a specific portion of a statistics lecture, not whole-course completion or product activation. Application to test: after explaining an important workflow, ask the learner to perform or recall the next step. The result does not imply that adding a quiz to every sales video increases revenue.
Read the video transcript
Watching a step is different from retrieving it from memory. After a short explanation, pause before revealing the next action.
In this illustrative workflow, the viewer chooses what follows a trigger. They must retrieve the next step, rather than simply replay the answer.
In a lecture study, tested and nontested groups scored 86 and 72 percent on final recall of the fourth segment. The restudy group scored 66 percent.
Those are results for a specific educational task. The workflow animation is an application to test, not evidence that a quiz will increase product sales.
Statistics calculated from the published results
These two comparisons use reported group means from the same two experiments in Szpunar et al. (2013). They are our arithmetic, not new experiments or independent findings. “pp” means percentage points; relative differences use the nontested score as the baseline. Inputs were rounded by the authors, so the calculations are approximate.
+14 pp in mean recall score for the fourth lecture segment: 86% − 72% = 14 percentage points. Relative to 72%, this is (86 − 72) ÷ 72 × 100 ≈ 19.4% higher mean score. Source: Szpunar et al. (2013), Discussion.
+22 pp on the entire final cumulative test: 90% − 68% = 22 percentage points for tested versus nontested groups. Relative to 68%, this is (90 − 68) ÷ 68 × 100 ≈ 32.4% higher mean score. Source: Szpunar et al. (2013), Discussion, all four segments.
The relative figures are descriptive differences between reported mean scores. We did not calculate new uncertainty intervals or adjust for study design. The fourth-segment and whole-test results overlap; do not count them as independent studies or add their effects together.
How to read these numbers
Standardised effect sizes are not percentages. A value such as g = 0.28 or d = 0.32 expresses a difference in standard-deviation units under the study’s estimator. It does not mean a 28% or 32% increase in learning, retention, sales or conversion. Keep g and d labelled as the original papers report them.
Also keep three outcomes separate: watching a video, learning its content, and making a purchase. Evidence for one does not automatically establish the others. A large observation count describes study scale; it does not remove selection bias or establish causality.
Read the video transcript
An effect size such as g = 0.28 describes a standardised difference. It does not mean a 28 percent improvement.
These overlapping curves are a schematic, not observed participant data. The gap between their centres represents a difference in standard deviation units.
Keep the original estimator attached to the number. Hedges g and the reported d values are not interchangeable percentage labels.
Also preserve the outcome: watching, learning, and purchasing are different measures. Cite the original study and test any application to product demos.
Selection, verification and limitations
We selected four peer-reviewed papers for relevance to explaining and demonstrating tasks. This is a curated evidence guide, not a systematic review of every study and not an original experiment. Publication details, numerical passages and their accessible locations were checked on October 20, 2026.
Guo et al. was verified through a university publication record; Noetel et al.’s numerical findings are available in the publisher’s public abstract. Rey et al.’s effect sizes were checked in Table 3 of the publicly accessible author-hosted paper, and Szpunar et al.’s comparison in the public PNAS text. The bibliography below links both DOI records and the public sources used. Some publishers may restrict full-text access.
The original years remain 2013, 2014, 2019 and 2021. Our 2026 check date does not make them 2026 experiments. We did not access private participant records, customer analytics or internal DemoBro experiments. We did not assess every risk-of-bias domain or perform a new meta-analysis.
Download the academic reference sheet
These downloads are an attributed reference list, not participant-level data. Use each finding’s original authors and study context when citing it. The copy controls include the paper title, venue, DOI and relevant evidence location.
For the practical format decision, see recorded video versus an interactive product demo. The next step is a task-specific test with your audience, rather than treating educational research as proof of a commercial result.