Background/Context: Online training has become an increasingly attractive alternative to traditional training (Laurillard 2016; Castaño-Muñoz et al. 2018). The online delivery of training allows overcoming geographical barriers and reaching higher numbers of participants in a flexible way and at reduced cost. Online training also offers higher potential for personalization to better meet the ever changing needs of learning processes (Dede et al., 2009; Bawa, 2016; Elliott, 2017). However, with little or no personalized guidance, learners with low self-regulated learning competence, little experience in online learning, or limited digital competences can be overwhelmed by the scale, diversity and flexibility offered by online training. All this translates into the low retention rates that are typically observed in scalable online courses (Jordan, 2014; 2015; Kizilcec et al., 2020) as well as in online courses delivered by post-secondary education institutions (Lee and Choi, 2011). In a context in which online training for teachers has attracted growing interest--also to face the unprecedented educational crisis induced by the COVID-19 pandemic, the need to understand what works to improve a successful participation in online training has become more compelling than ever before. While there is widespread agreement about the need to improve the instructional design and the support mechanisms in online courses in order to increase completion rates, robust evidence on what works is limited and is not specifically focused on teachers (Kizilcec et al., 2020). Purpose/Objective/Research Question: We present the results of an European Union (EU) funded policy experimentation aimed at developing a new personalized support model and evaluating its effectiveness in increasing teachers' retention in scalable online courses. The support model offers course participants with low levels of self-regulated learning competences concrete guidance and support to complete the courses they start, and also to actually start the courses they have enrolled in. Setting: The support model was tested in the context of four online courses on key 21st century competences for teachers (i.e., formative assessment, personalized learning, collaborative learning and creativity). The courses were offered to professional teachers (i.e. in-service teachers) and student teachers (i.e. future teachers who were completing their training or induction period to become teachers) during school year 2018/2019. Population/Participants/Subjects: The study involved 3,777 lower-secondary education professional and student teachers in nine European Union Member States (Austria, Estonia, Greece, Hungary, Lithuania, Malta, Portugal, Slovakia, Spain) and in Turkey, who were invited to participate following a stratified sampling design of schools and initial teacher education (ITE) organizations. Intervention/Program/Practice: The online support consisted of a package of nine personalized messages, which provided participants with concrete support and guidance to start or complete the courses. The messages were sent only to teachers "in need" of support; i.e. messages were triggered by specific teachers' characteristics (collected at baseline) and teachers' specific (in)actions on the course platform during the course delivery. For example, teachers showing poor past experience in online training or low levels of self-regulated learning online were provided guidance and received an offer to arrange 1:1 sessions for a course "walk-through" or to just answer questions. Also, teachers who had not started a course after five days, received a personal reminder offering additional support. Research Design: The identification of the effects of the personalized support relies on random assignment: control group teachers received standard access to the courses, while treatment group teachers had the possibility to receive the personalized support. Assignment to the treatment or the control group did not occur at the teacher but at organization level (school or ITE) to minimize the risk of off-line contamination. The data are analyzed by the means of three-level hierarchical linear regression models in which observations were "nested" within schools (or ITE organizations) and individuals. This modeling strategy accounts for the organization-level randomization design as well as for the fact that teachers were nested in schools and could participate in more than one of the four courses. Data Collection and Analysis: The evaluation exploits course platform-generated data on course progression and baseline survey data collected at application. The primary outcome (i.e., 'course completion among enrolled') is calculated for teachers who enrolled in a given course and distinguished teachers who completed that course from those who did not. Additional outcomes are 'course start' (computed among teachers who enrolled in a given course); and 'course completion among started' (calculated among teachers who actually started a given course). Findings/Results: The analysis shows heterogeneous impacts. A substantial and positive causal impact on course completion (+10,6 percentage points) was found for professional teachers in the pooled sample of EU Member States (average completion rate in the control group is 32%). By-course analyses show that this positive effect was a combination of support messages impacting on course completion and other messages impacting on course start. However, the tested intervention did not change course participation for Turkish professional and student teachers and it did it only partially for student teachers in EU countries (i.e., only for those who had some past experience in online training). Conclusions: The experiment shows that personalized support messages have a potential to counteract dropout in online courses. The study identifies concrete implications for professional development and initial teacher education providers, such as reaching out to teachers with poor online training experience and to "late starters" by sending personalized messages. Also, the study highlights some knowledge gaps that require further research, such as disentangling the mechanisms behind personalized support in online environments and investigating the heterogeneous impacts and the external validity of support interventions.