Measurement

Beyond completion rates

Only 35% of organisations measure anything beyond completion. Not because the alternatives are exotic, but because nobody tells you which ones to start with. These are the seven, in order.

8 min read
In short

The seven metrics worth adding are: content-level engagement score, drop-off point, focus-adjusted time on task, tab-switch rate, replay rate, interaction latency, and assessment delta with delayed retention. All except the last are collected from ordinary browser events — no camera, no biometric data.

Completion rate answers "did they reach the end?" These seven answer "what happened on the way?"

1. Content-level engagement score

What it is: An engagement value per section, slide or segment, rather than one figure for the whole module.

Why it matters: It is the only metric on this list that produces a work order. "Module scores 3.8/5" starts a debate. "Section 4 scores 0.31 against a course average of 0.72" starts a rewrite.

How to collect it: Composite of dwell time, focus state and interaction, aggregated per content position via Intersection Observer.

Watch for: Averaging across sections destroys the signal entirely. The whole point is the variance between sections, not the mean.

2. Drop-off point

What it is: Where sessions actually end, expressed as a content position rather than a percentage.

Why it matters: This is the highest-value, lowest-effort metric available. It requires almost no analysis to interpret, and it converts directly into a prioritised fix list.

How to collect it: Record the last content position reached before session end, using visibilitychange and pagehide with a sendBeacon call.

Watch for: Distinguish abandonment from completion-at-the-end. Someone who stops at 98% finished; someone who stops at 34% did not.

3. Focus-adjusted time on task

What it is: Time spent on a section, minus the time the window did not have focus.

Why it matters: Raw time-on-task is one of the most misleading numbers in learning analytics. A module left open in a background tab during a two-hour meeting reports beautifully. Focus-adjusting it turns a vanity metric into a real one.

How to collect it: Page Visibility API plus focus/blur listeners. Accumulate time only while the document is visible and focused.

Watch for: Passive video content, where correct behaviour is stillness. Focus state still works; interaction-based signals do not.

4. Tab-switch rate

What it is: How often, and for how long, a learner navigates away from the content.

Why it matters: This is the strongest single predictor available. A 2024 study found "cyberloafing" — tab-switching away from learning content — outperformed both self-regulation and satisfaction measures as a predictor of disengagement in online courses.

How to collect it: document.visibilitychange, recording both frequency and duration.

Watch for: Duration matters more than count. Eight seconds is a reference check; eight minutes is a departure. Baseline against each learner's own pattern rather than a population average, or you will permanently flag habitual multitaskers.

5. Replay and rewind rate

What it is: Which segments people go back over.

Why it matters: It is the only signal here that positively identifies confusion rather than disengagement. A high replay rate is not a failure — it is a section that people cared enough about to re-watch and did not understand the first time. That is a precise, actionable diagnosis.

How to collect it: Media seeked events, or scroll-position reversal for text content.

Watch for: Do not conflate with disengagement. High replay plus high completion is a clarity problem. High replay plus high drop-off is a difficulty problem.

6. Interaction latency

What it is: Time between a prompt appearing and the learner responding.

Why it matters: Click counts are trivially gameable and mostly measure interface density. Latency measures whether someone was there and thinking. A latency distribution that shifts rightward through a module is a fatigue signal.

How to collect it: Timestamp prompt render, timestamp first interaction, take the difference.

Watch for: Very fast responses are as informative as very slow ones — sub-second answers to a comprehension question usually mean pattern-matching, not understanding.

7. Assessment delta with delayed retention

What it is: Pre-test to post-test improvement, re-measured at 30 days.

Why it matters: The post-test score alone is nearly meaningless — it conflates prior knowledge with what was learned. The delta isolates the learning. The 30-day re-check separates genuine retention from short-term recall, and the gap between them is usually sobering.

How to collect it: Standard assessment tooling. This is the one metric on the list your LMS probably already supports.

Watch for: Add confidence weighting if you can. A learner who is confidently wrong is a different risk from one who is uncertainly wrong, and in compliance training the first category is the one that causes incidents.

Which to start with

Do not attempt all seven. Sequence them.

Order Metric Effort Payoff
1 Drop-off point Low Immediate — a prioritised fix list
2 Focus-adjusted time on task Low Kills your most misleading existing metric
3 Content-level engagement Medium The core diagnostic
4 Tab-switch rate Low Strongest predictive signal
5 Assessment delta Medium Proves learning, not just attention
6 Replay rate Low Isolates confusion from boredom
7 Interaction latency Medium Refinement
A two-week version

Instrument one course with drop-off tracking and focus-adjusted time. Find the largest drop-off. Rewrite that section. Measure again. You now have an attributable improvement, an internal proof case, and a much easier conversation about funding the rest.

One closing note on what you are not collecting. Every metric here except assessment delta comes from ordinary browser events — visibility, focus, scroll, interaction timing. No camera, no biometric data, no emotional inference. That keeps you clear of the EU AI Act Article 5(1)(f) prohibition on inferring emotions in workplace and education contexts, and it makes the DPIA conversation considerably shorter.

The strongest predictor in the research is also the one with the least regulatory friction. That is a rare and convenient alignment, and it is worth taking.

Frequently asked questions

Do I need to replace my LMS to collect these?
No. Six of the seven are collected client-side from browser events, which means a script in the content layer rather than a change to the platform. Your LMS keeps doing what it does; the engagement instrumentation sits alongside it.
Do any of these require a webcam?
None of them. All seven are derived from browser and interaction events or from assessment data. That matters legally as well as practically — EU AI Act Article 5(1)(f) prohibits inferring emotions from biometric data in workplace and education contexts, and none of these signals involve biometric data.
Which single metric should I add first?
Drop-off point. It is the simplest to collect, the easiest to explain to stakeholders, and it produces an immediate work order: rewrite the section where people leave. Everything else is refinement.
How much data do I need before these are meaningful?
For content-level patterns, 30–50 sessions per module gives a usable signal, and the drop-off points are often obvious well before that. Individual-level inference needs considerably more, which is one of several reasons to report at content level.

See where your content loses people

Book a walkthrough and we will show you the engagement data on your own content.

Request a demo
Get Started

See what completion rates can't tell you

Find out exactly where your content works, where it fails, and what disengagement looks like before people leave.

Request a Demo