Hardware cohort retention without pretending it is SaaS

A cold-chain team in our 2025 studio arrived with a beautiful daily-active-device plot. Gateways in rural routes buffered temperature samples for six to fourteen hours. Devices looked “dead” every afternoon and “resurrected” every night. Retention looked like a horror film. The hardware was fine.

Pick a window the physics can keep

If your uplink budget or cellular coverage cannot guarantee a same-day sample, do not measure same-day retention. For most Thailand fleets we review, a week is the smallest window that survives buffering, public holidays, and installer schedules. Some industrial valves only speak every 48 hours by design. Their cohort window is fourteen days. That is not a failure of ambition.

Firmware line is the cohort key

SaaS retention cohorts by signup week. Hardware retention should cohort by the firmware that defined the event contract at activation, then watch whether later OTAs keep that contract. Mixing 2.3 and 2.6 in one retention curve is how a sleep-mode bug gets reported as “users falling out of love with the brand.”

In IoT App Analytics, love is rarely the variable. Radios, batteries, and silent schema breaks are.

Companion app opens are a side plot

Track them. Do not let them replace device samples. A technician who never opens the app after install can still run a healthy valve. A consumer who opens the app daily on a dead sensor is not retained in any sense that should reach a board slide.

Fleet Workshop spends a full session reconciling these two plots. Device Signal Studio goes further and forces ops and product to share the firmware dimension. If you only remember one sentence from this note: retention without firmware is a story about the radio, told as if it were a story about the customer.