The token-health checklist from week two is now a Monday ritual. We found 14% of the “active” Android audience had not accepted a message in ninety days. Finance finally stopped asking why volume was down while “reach” looked healthy.
Reviews
What people actually kept after the cohort closed
We ask alumni which artefacts they still open. The notes below mention specific modules from this studio — not a generic “great course” template.
Quiet-hours mapping for Bangkok versus Phuket was the piece I did not know I needed. I still argue with Marek’s stance on retry budgets — our engineers think his backoff diagram is too conservative — but tap-through on evening sends stopped looking random.
★★★★☆
Solid on delivery vs display. The cohort-chart week assumed more SQL comfort than our growth intern had. We paired her with an analyst and she finished. Would not send a complete beginner alone.
We came for attribution windows and left arguing about OEM battery restrictions. That was useful. The Lab will not write SQL for you; if your warehouse is a mess, fix that before week five or you will stall.
Case study
Beauty retailer, Bangkok: collapsing “delivered” into three truths
A mid-size beauty group sent Cloudqueuegrid a vendor export that showed 92% delivery on a new-product drop. After the Delivery Intelligence Lab critique, the same file split into accept (vendor), display (OS), and tap (app). Display sat closer to 71% on a popular Android OEM once battery restrictions were labelled.
The team did not change creative in the first month. They changed retry policy and stopped counting collapsed notifications as impressions. Attributed add-to-cart from push, using a six-hour window taught in week six, fell — then stabilised at a number finance would sign.
Limitation they still live with: web push was out of scope for that cohort, so desktop subscribers remain on a separate, weaker report.
Case study
Regional bank: queue latency that looked like “user fatigue”
Fraud-alert push was arriving 9–14 minutes after the triggering event. Product assumed users were ignoring the alerts. Queue Throughput Diagnostics (run as a private room after two alumni recommended it) showed the delay sitting in an internal worker, not in APNs.
Once latency became a product metric with an owner, tap-through on fraud alerts rose without a copy rewrite. The bank still declines to publish exact figures; they allowed us to describe the mechanism.
Nalinee’s note in the debrief: “If your alert is late, no amount of emoji testing will rescue trust.”