Strategy & Growth

Online Course Completion Rates: What the Research Actually Shows

Course-completion statistics travel without their definitions. Here is what the 6.5% and 12.6% MOOC figures count, why half of registrants never start, what four reminder experiments found, and why no independent study backs the "cohorts finish at 90%" line.

oToK Team14 min read
Online Course Completion Rates: What the Research Actually Shows

Not legal advice: The research below covers free open university courses (MOOCs) and one software vendor's platform data. It is not a sample of paid Israeli courses, and we found no public Israeli completion data. The funnel example uses made-up numbers; replace them with your own.

Search for online-course completion and you meet the same few numbers: only 5% to 15% of students finish, cohort courses complete at 80% to 90%, and a well-timed reminder can lift completion by almost a third. They are repeated as if they described any course. They do not. The first comes from free university courses, the second from founders and software vendors, and the third from one early experiment that larger tests did not reproduce.

We went back to the papers: Katy Jordan's two surveys of completion rates, Justin Reich and José Ruipérez-Valiente's 2019 analysis of 12.67 million HarvardX and MITx registrations, a 44-course study of what registrants actually do, four experiments on planning and scheduling prompts, a comparison of government nudge trials with published ones, and the one public vendor dataset on cohort courses. The short version: completion in open courses is genuinely low, most of the loss happens before anyone really starts, reminders help less than the headlines say, and no independent study tells a paid Israeli course what to expect. What you can do is count your own funnel with one stated definition and change one thing at a time.

Key takeaways

  • Jordan's famous 6.5% is a median over 39 courses from Coursera, edX and Udacity (a list dated 22 July 2013), counting certificates as a share of total enrolment. On an active-student basis the median was 9.8% (33 courses). Her 2015 update, covering 129 courses that reported completion, found a median of 12.6%. "Completion" meant a certificate in 93 of those 129 courses.
  • Much of the loss is people who never start. Across 44 Coursera courses, on average 54.4% of registrants watched at least one lecture video, and in 16 of the 44 courses most never watched one (Evans, Baker and Dee). In the HarvardX and MITx data, 52% of registrants never entered the courseware.
  • Completion did not improve over six years of MOOC investment. Among learners who registered and viewed a course, the completion rate went from 4.95% in 2013-14 to 3.13% in 2017-18; among those who said they intended to complete, from 11.51% to 15.42% (Reich and Ruipérez-Valiente, supplementary table).
  • The "planning prompts raise completion by 29%" result is from three HarvardX courses. In an experiment that assigned 269,169 students across 247 courses to several interventions or a control, plan-making prompts did not improve completion for committed English-fluent students (Kizilcec and co-authors, PNAS 2020). A scheduling nudge tested in an 18,043-learner experiment had no near-term effect and weakly negative longer-term effects.
  • Nudge results shrink as studies get bigger. Two US nudge units' 126 trials, reaching 23 million people, averaged 1.4 percentage points, against 8.7 in published academic papers (DellaVigna and Linos). Those trials measured sign-ups and similar actions, not course completion.
  • No independent study backs "cohort courses complete at 80-90%". The 96% for altMBA is self-reported by a Maven co-founder, with no definition or data. The one public dataset, from the vendor Ruzuku, shows 53.4% for scheduled cohorts against 42.9% for on-demand drip courses, is not peer reviewed, and does not adjust for price, subject or length.

What is the real completion rate of online courses?

It depends on which study you pick, and every number answers a different question. The most-quoted source is Katy Jordan's "Initial trends in enrolment and completion of massive open online courses" (IRRODL, 2014). She gathered public figures for Coursera, edX and Udacity courses from a list of 279 courses dated 22 July 2013. Completion was known for only 39 of them. Counting the students who met the certificate criteria as a share of everyone enrolled, the rates ran from 0.9% to 36.1% with a median of 6.5%. The abstract says "average"; the body reports medians.

Jordan flags the denominator herself. Many people enrol without ever starting, so she also computed completion as a share of "active" students (those who logged in, attempted a quiz or viewed a video). For the 33 courses with that data the median was 9.8%, with a range from 1.4% to 50.1%. She also notes that 67.4% of eligible courses had released no data at all, so the courses that published may not be typical.

Her 2015 follow-up, "Massive Open Online Course Completion Rates Revisited", used 221 courses, with completion data for 129 of them. The median completion rate, defined as the share of enrolled students who completed, was 12.6% (range 0.7% to 52.1%). She reports that enrolments had fallen while completion rates had risen. That is why quoting 6.5% today is quoting a 2013 sample.

Typical completion rate by study and definition (medians for Jordan, averages for Evans et al.)
Jordan 2015: share of enrolled who completed (129 courses)
12.6%
Jordan 2014: share of active students who got a certificate (33 courses)
9.8%
Evans et al.: certificate rate among registrants who watched a video (40 courses with certificates)
8.5%
Jordan 2014: certificate share of total enrolment (39 courses)
6.5%
Evans et al.: certificate share of all registrants (average, 40 courses with certificates)
5%
Source: Jordan, IRRODL 2014 and 2015; Evans, Baker and Dee, Journal of Higher Education 2016 (online appendix). Definitions differ by bar.

These bars are not one series. They mix medians and averages, different course sets and different denominators. Read them as a map of how much the definition moves the answer, not as a trend.

Why does "completion rate" mean five different things?

Because there is no standard definition. Of the 129 courses in Jordan's 2015 sample that reported completion, 93 defined it as earning a certificate; the rest used "completed course" (14), "passed course" (10), "completed assignments" (6) and a handful of one-off definitions such as taking the final exam. Her warning, from the 2014 paper, is that comparing MOOC completion with campus-course completion is inappropriate.

The denominator matters as much as the definition. Brent Evans, Rachel Baker and Thomas Dee studied 44 Coursera courses, more than 2.1 million student observations in all. On average 54.4% of registrants watched at least one lecture video, and in 16 of the 44 courses the majority never watched a single one. Only about 10% watched more than 80% of the lectures, and certificates averaged around 5% of registrants. The authors write that using all registrants as the denominator to produce a completion rate is "very misleading"; even conditional on watching one video, the average certificate rate rose only from 5% to 8.5%.

Reich and Ruipérez-Valiente's 2019 Science analysis of every HarvardX and MITx course on edX through May 2018 (565 course runs, 261 courses, 12.67 million registrations from 5.63 million learners) arrives at the same place: 52% of registrants never enter the courseware. So a completion rate computed on registrants is partly a measure of how many people merely signed up.

54.4%

of registrants watched at least one lecture video, on average across 44 Coursera courses. In 16 of the 44 courses, most registrants never watched one.

Source: Evans, Baker and Dee, "Persistence Patterns in Massive Open Online Courses", Journal of Higher Education, 2016

A free course that anyone can join with one click collects curious browsers. A paid course with a start date collects people who paid. The two denominators are different populations, so their completion rates cannot be compared directly.

Has completion improved as courses got better?

Not in the largest dataset we could verify. Reich and Ruipérez-Valiente write that "six years of investment in course development and learning research has not produced meaningful improvements in these figures." Their supplementary table is more nuanced than the headline. Among learners who registered and viewed the course, the completion rate drifted down from 4.95% (2013-14) to 3.13% (2017-18). Among those who said at registration that they intended to complete, it rose from 11.51% to 15.42%, after peaking at 17.74% in 2014-15. Learners who paid for the verified certificate track completed at roughly 44% to 56%, and the paper notes that completion stays low even among people who pay.

Two readings follow. First, "unchanged" is a simplification: intent-to-complete learners improved a little while the broad group slipped, probably because the pool of casual registrants grew. Second, paying does not make people finish. Even a self-selected group that put money down completed about half the time.

Completion rate by learner group, HarvardX and MITx coursesRegistered and viewed the courseSaid they intended to complete

2013-14

Registered and viewed the course: 4.95%
Said they intended to complete: 11.51%

2014-15

Registered and viewed the course: 5.85%
Said they intended to complete: 17.74%

2015-16

Registered and viewed the course: 4.11%
Said they intended to complete: 16.44%

2016-17

Registered and viewed the course: 3.98%
Said they intended to complete: 16.13%

2017-18

Registered and viewed the course: 3.13%
Said they intended to complete: 15.42%
Source: Reich and Ruipérez-Valiente, Science 2019, supplementary table. Cohort years as labelled in the table.

These rates use participants and intenders as the denominator, a different base from Jordan's enrolment-based figures. Do not mix the two.

When do learners actually leave?

Early. The most robust finding across the studies is not a smooth decay curve but a cliff at the start. Evans and colleagues found many students "watch one or two videos and then disengage." Jordan's 2015 paper reports that the first two weeks are critical for engagement. In one 12-week psychology MOOC analysed by Marius Kloft and colleagues, 11,607 participants were active in the first week and 3,861 remained until the last week, about a third. The authors found the dropout probability "especially high in the first two weeks," then again at the end of the course.

We would not draw a universal week-by-week curve from this. The week-level shapes come from specific courses, and a single course's curve is not a law. What several independent datasets share is that the largest single loss happens before or during the first week or two, and that a large share of registrants never begin.

For a creator, that moves the focus. If a big part of your loss happens before the first lesson or the first live session, then the invitation, the confirmation, the access email and the first reminder are the course, as far as completion goes. That is where a fix is cheapest to test and where there is the most to win.

33%

of one MOOC's first-week participants (3,861 of 11,607) were still active in the last week of a 12-week course (our arithmetic from the paper's counts). One course, not a benchmark.

Source: Kloft, Stiehler, Zheng and Pinkwart, EMNLP 2014 MOOC workshop

Dropout in that paper means no activity in that week or any later week. A learner who skips a week and returns is counted as still active.

Do reminders and planning prompts raise completion?

The first result is encouraging and widely quoted. In a natural field experiment in three HarvardX courses, Michael Yeomans and Justin Reich (LAK 2017) found that a planning prompt, which asked learners to make a plan for when they would study, increased course completion by 29% and payment for certificates by 40%. Note that 29% is a relative increase on a low base, not 29 percentage points, and the study was small compared with what followed.

What followed was much less impressive. Kizilcec and eight co-authors (PNAS, 2020) assigned 269,169 students in 247 courses to several interventions or a control at Harvard, MIT and Stanford. They report that long-term planning prompts "did not improve course completion for committed English-fluent students," and that a social-accountability intervention did not improve completion in non-individualist countries. One intervention, helping learners connect the course to their own values, raised completion by about 2.79 percentage points, but only for learners in less-developed countries, in courses with a global achievement gap. The authors warn that deploying interventions widely can cut their average effectiveness by an order of magnitude.

Four randomized trials of planning prompts on openHPI and openSAP courses (Andor and colleagues, 2018) reported an overall null effect of planning prompts on completion across four courses, with one course showing a 19% gain. And a scheduling nudge, tested in a randomized experiment with 18,043 learners in which a random subset could choose when to watch the first lecture, "had no effects on near-term engagement and weakly significant negative effects on longer-term course engagement, persistence, and performance" (Baker, Evans and Dee, AERA Open 2016).

StudyWhat was testedResult
Yeomans and Reich, LAK 2017 (3 HarvardX courses)Planning promptCompletion +29% (relative), certificate payment +40%
Kizilcec et al., PNAS 2020 (269,169 students, 247 courses)Plan-making, value relevance, social accountabilityNo completion gain from plan-making for committed English-fluent learners; value relevance +2.79 pp in a specific setting
Andor et al., 2018 (openHPI and openSAP, 4 courses)Planning prompt, four randomized trialsOverall null; +19% in one course
Baker, Evans and Dee, AERA Open 2016 (18,043 learners)Choose when to watch the first lectureNo near-term effect; weakly negative longer term

Every row is about open, mostly free courses. None tests a reminder that a paying customer receives from the business they bought from, which is a different relationship.

Why do nudge results shrink at scale?

Because small early studies overstate, and because tests that work in one place do not travel. Stefano DellaVigna and Elizabeth Linos (Econometrica, 2022) assembled 126 randomized trials from two US nudge units, reaching 23 million people. The average effect was 1.4 percentage points, an 8.0% increase, "sizable and highly statistically significant," but far below the 8.7 percentage points average in nudge papers published in academic journals. In the published version, they find publication bias explains about 70% of that gap, with differences in the nudges themselves explaining most of the rest. The outcomes there were administrative take-up, not course completion.

Education-specific nudge trials show the same pattern. A pair of FAFSA campaigns reached over 800,000 students and found "no impacts on aid receipt or college enrollment overall or for any subgroups" (Bird and co-authors, Journal of Economic Behavior and Organization, 2021), and the authors discuss why strategies that work locally may be hard to scale. Philip Oreopoulos and Uros Petronijevic ran five years of online and text-message nudges for nearly 25,000 students at three campuses and found that none of the interventions significantly improved academic outcomes (NBER working paper 26059, 2019). Those were on-campus students, not online learners.

The wider literature is argued, not settled. A 2022 PNAS meta-analysis of more than 200 studies found an overall nudge effect of Cohen's d = 0.45 and a moderate publication bias toward positive results. A reanalysis in the same journal concluded that once publication bias is corrected, no evidence for nudge effectiveness remains. Its authors later said the result is undecided rather than evidence against nudging, and critics argue that averaging very different nudges makes both numbers hard to interpret. A fair summary: reminders probably do something, probably less than the first paper you read, and the effect depends on the exact message, audience and outcome.

Average nudge effect, percentage points
Published in academic journals
8.7 pp
Nudge-unit trials (126 trials, 23 million people)
1.4 pp
Source: DellaVigna and Linos, Econometrica 2022 (126 nudge-unit trials; academic-journal average reported alongside)

Both bars measure behaviors such as sign-ups and applications. Treat them as a caution about published effect sizes in general, not as a forecast for your reminders.

Do cohort courses really complete at 80-90%?

We could not find an independent source. The figure traces back to operators describing their own programs. Wes Kao, a co-founder of Maven who also co-founded the altMBA with Seth Godin, wrote in a 2021 essay on a16z's Future site that "we saw a 96 percent completion rate" for the altMBA and that Harvard's case-method courses' completion rate "rose to 85 percent." The essay gives no definition of completion, no data source and no denominator. Other vendor pages repeat 85% to 90% for cohorts and 10% to 20% for self-paced, without a study we could open.

The one public dataset with a stated method is from the course platform Ruzuku. Its author analysed 27,156 published courses and 1,316,272 student enrolments (data copied 20 August 2026) and counted an enrolment as completed when the platform recorded a completion date for it. Scheduled cohort courses (calendar-based access; the platform does not classify whether delivery is live) showed 53.4% completion, on-demand drip courses 42.9% and open-access self-paced courses 41.9%. Courses with a lesson discussion showed 50.9% against 37.3% without, and courses with assessments 59.1% against 42.4% without.

The author states the limits: a single database snapshot, observational data that does not establish causation, no adjustment for price, course length, subject or school, no peer review, and the author co-founded and runs the company that sells the compared features. We would add that the platform's "completed" flag has no common standard, so it is not comparable with a MOOC certificate rate. Read as a pattern, it says scheduled cohorts finished more often than open self-paced courses, by about 11.5 points, in one vendor's data. It does not support 80% to 90%.

Completion by course format, one platform's data (vendor analysis, not peer reviewed)
Scheduled cohort courses
53.4%
On-demand drip courses
42.9%
Open-access self-paced courses
41.9%
Source: Ruzuku, cohort vs self-paced completion study, data copied 20 August 2026 (27,156 courses, 1,316,272 enrolments)

Selection explains part of any gap. Learners who commit to a start date and a group are not the same people who click "enrol" on a free self-paced course, and small creators publish their good numbers.

What should a course creator count?

Start with one written definition and four or five counts from your own course, because that is the only evidence that describes your audience. Pick a definition before the cohort starts (for example, attended or completed the last live session, or passed the final task) and keep it for every cohort. Then count each step against the same registrants.

Here is the shape, with invented numbers for a cohort of 200 registrants. The point is the order of magnitude of each loss, not the values.

  • Read it two ways: 60 of 200 is 30% of registrants, but 60 of 130 is 46% of the people who started. The first measures the invitation and the first week; the second measures the course.
  • The biggest single drop is 70 people between registering and starting, which is where the research says the loss usually concentrates. A change there (a faster access email, a reminder the day before session one) is worth more than a change in week six.
  • Be realistic about testing. To detect a lift from a 30% to a 33% completion rate with 80% power at the usual 5% significance level, you would need about 3,760 learners per group (our calculation). A cohort of 200 cannot prove that a nudge worked. Compare cohorts only as a rough signal, change one thing per cohort and judge by large, structural effects.
  • Keep the denominator honest when you report numbers to yourself or to buyers: say whether you mean registrants, starters or people who paid.
Hypothetical funnel for one cohort of 200 registrants (invented numbers)
Registered
200
Started lesson 1 or attended session 1
130
Reached lesson 2 or session 2
90
Finished by the written definition
60
Source: Illustrative only, not data from any study or customer

Do not take a completion rate from a MOOC study or a vendor blog as a target for your course. Your own previous cohort is the benchmark that matters.

How does oToK help you measure and fix the first weeks?

oToK is not a course-hosting platform, and it does not track lesson-by-lesson progress inside a video course. It manages what surrounds the course: registration, cohorts, live-session attendance, the messages around them and the money. If your lessons live on another platform, a completion event can reach oToK through an inbound webhook or the API and then trigger an automation like any other event.

For the stretch the research points at, the registration and the first sessions, these are the pieces a course creator uses:

  • Product cycles: each cohort has its own dates, price and number of seats; a full cycle shows a "Full" badge, and each cycle keeps its own roster, waitlist and attendance
  • Registration, cancellation, attendance and waitlist triggers, so confirmations, reminders and no-show follow-ups run without anyone remembering to send them
  • One event automation can target a whole product, so the same flow covers every future cohort instead of being rebuilt per cycle
  • A connected Zoom account registers contacts into the session and syncs attendance back to the event
  • WhatsApp, SMS or email campaigns to an event's attendees, by attendance status, with exclusions by event and status
  • Automations hold over Shabbat and holidays and resume by rule, which matters when a cohort runs through the Hebrew calendar
  • Reports can be sliced by cycle, so you can compare the funnel of this cohort with the last one

None of this raises completion by itself. It lets you count each step with one definition, reach people at the moments where the research says most of the loss happens, and compare one cohort with the next.

Share this article

Frequently asked questions

What is the average completion rate of online courses?
There is no single average. For free open university courses, Katy Jordan found a median of 6.5% in 2014 (39 courses, certificates as a share of enrolment) and 12.6% in 2015 (129 courses). A 44-course study found certificates averaged around 5% of registrants. For paid creator courses we found no independent public figure, only platform and vendor numbers.
Is the 5-15% MOOC completion rate accurate?
It is a loose summary of research on free courses from 2012 to 2015, and the exact figure depends on the definition and denominator. Jordan's medians were 6.5% and 12.6% of enrolled students; her active-student median was 9.8%. A 2019 analysis of HarvardX and MITx data put completion among those who registered and viewed at 3.13% to 5.85% depending on the year. It is not a benchmark for a paid course.
Do cohort-based courses have higher completion than self-paced courses?
Probably somewhat, but the size is unproven. One vendor's data (Ruzuku, 27,156 courses) shows 53.4% for scheduled cohorts against 41.9% for open self-paced courses, with no adjustment for price or subject and no peer review. The 85% to 96% figures come from operators describing their own programs without a definition or data. Selection of learners explains part of any gap.
Do reminder emails or WhatsApp messages increase course completion?
The evidence is mixed. One small planning-prompt study in three HarvardX courses found completion up 29% (relative), but an experiment spanning 269,169 students and 247 courses found no completion gain from planning prompts for committed English-fluent learners, and four later randomized trials found an overall null. We found no peer-reviewed study of WhatsApp or SMS reminders on online-course completion, so the effect on your course has to be measured.
When do students drop out of an online course?
Mostly at the start. In 44 Coursera courses, on average only 54.4% of registrants watched even one lecture, and in one 12-week MOOC about a third of first-week participants were still active in the last week. Jordan reports the first two weeks as critical. The exact week-by-week curve varies by course, so measure yours.
Does paying for a course make people finish it?
Not by itself. In HarvardX and MITx courses, learners on the paid verified track completed at roughly 44% to 56% across five yearly cohorts. That is much higher than the broad group, but it means about half of people who paid still did not finish. Payment selects motivated learners; it does not guarantee completion.
How big must a test be to prove a reminder works?
Larger than most course cohorts. Detecting a rise from a 30% to a 33% completion rate with 80% power at 5% significance needs about 3,760 learners in each group, by the standard two-proportion formula. With a few hundred learners, treat a before-and-after difference as a hint, not proof, and prefer changes with large expected effects.
Is there Israeli data on online-course or webinar completion?
We looked for public figures from Israeli statistics and education bodies and found none we could verify for short online courses or webinars. Everything above comes from international research and one vendor's platform data, so use your own cohorts as the Israeli benchmark.
Can oToK track how far learners get through a video course?
Not inside the lessons. oToK manages registration, product cycles, live-session attendance, automations, messages, payments and reports around the course. If your video platform can send an event, an inbound webhook or the API can bring lesson completion into oToK and use it as an automation trigger.

Ready to upgrade your customer communication?

Get started with oToK — connecting a WhatsApp number takes minutes.

No credit card required — start right away