Firebase’s push‑notification system can mislead teams if they rely only on dry‑run checks. In practice, a token that passes a dry‑run may become invalid minutes later, causing a real send to fail. The author’s team began tracking every actual send result in January 2025, discovering that 500,000 messages were sent but only 350,000 reached devices—a 70 % delivery rate.

This gap skewed engagement metrics and wasted server time on 150,000 dead tokens, costing 20 hours of processing per month. By parsing FCM responses, the team classified failures into invalid token, temporary, quota, and other categories. Permanent errors triggered immediate database cleanup, while temporary ones were retried with exponential back‑off, achieving a 97 % success rate on retries.

Automation of token removal after each campaign lifted the overall delivery rate from 70 % to 86 % over six months, saving roughly 10 hours of server time monthly. The approach shows that measuring ground‑truth delivery, not just predictions, is essential for accurate analytics and cost control in high‑volume push services. Teams adopting this methodology report clearer KPI visibility and reduced operational overhead.