Internet Service Providers (ISPs) are getting involved in remediating Internet of Things (IoT) infections of end users. This endeavor runs into serious usability problems. Given that it is usually unknown what kind of device is infected, they can only provide users with very generic cleanup advice, trying to cover all device types and remediation paths. Does this advice work? To what extent do users comply with the instructions? And does more compliance lead to higher cleanup rates? This study is the first to shed light on these questions. In partnership with an ISP, we designed a randomized control experiment followed up by a user survey. We randomly assigned 177 consumers affected by malware from the Mirai family to three different groups: (i) notified via a walled garden (quarantine network), (ii) notified via email, and (iii) no immediate notification, i.e. a control group. The notification asks the user to take five steps to remediate the infection. We conducted a phone survey with 95 of these customers based on communication–human information processing theory. We model the impact of the treatment, comprehension, and motivation on the compliance rate of each customer, while controlling for differences in demographics and infected device types. We also estimate the extent to which compliance leads to successful cleanup of the infected IoT devices. While only 24% of notified users perform all five remediation steps, 92% of notified users perform at least one action. Compliance increases the probability of successful cleanup by 32%, while the presence of competing malware reduces it by 54%. We provide an empirical basis to shape ISP best practices in the fight against IoT malware.