Every report said “delivered”. Gmail disagreed.

The situation

Green reports, silent failure.

An AI companion app on a monthly subscription, with 66,000 contacts. Campaigns were scheduled, onboarding was live, re-engagement was live. Sends went out every day and every report came back green.

The reports measured whether emails were sent. Nobody was measuring whether they arrived, whether any sequence responded to what a user did, or what the users falling out of the funnel were worth. We were brought in for six weeks to audit the whole system, rebuild it and hand it back documented.

What we found

Five problems no dashboard showed.

1. 1

   Gmail had stopped trusting the domain

   User-reported spam had spiked repeatedly since January, peaking around 28% in early February. SPF, DKIM and DMARC needed correcting. Gmail bounces sat at 56.3%; only 43.7% was delivered.

2. 2

   A third of the list was invalid

   32% of the 66,000 contacts were invalid addresses the system had been emailing for months, poisoning the sender reputation the important campaigns depended on.

3. 3

   One trigger queued 174,000 sends

   A single misconfigured entry condition on the re-engagement campaign had queued around 174,000 sends against a 66,000-contact base. It appeared in no campaign report.

4. 4

   Onboarding ran on a timer

   Emails fired on elapsed days, whatever the user had done. In the month before, it sent 76,266 emails and 13.3% of new users reached companion creation, the product’s activation moment.

5. 5

   No step checked behaviour

   No sequence checked for an open, a click, a product event or a return, so an engaged user and a dead address were treated identically.

What we built

Six moves, in the order they shipped.

1. Authentication rebuild

   SPF, DKIM and DMARC corrected, so every receiving provider could verify the domain before volume came back.

2. List cleanup

   The invalid third suppressed. Sends stopped going to addresses that could only bounce.

3. Controlled ramp

   Volume brought back on a schedule the domain could sustain, instead of re-flagging it on day one.

4. Behavioural onboarding

   Onboarding rebuilt around what the user does: opened, created a companion, sent a message, came back. Each branch gets a different next email, or exits.

5. Five lifecycle flows

   Onboarding, re-engagement, gate-hit, paywall conversion and broadcast. Re-engagement rebuilt on a correct entry condition, clearing the queue.

6. Measurement and handoff

   A benchmark for every flow, reported daily in Slack, plus a launch-readiness roadmap and the SOPs the internal team now runs the system from.

Results

What changed, and where each figure comes from.

Later, the product’s paywall changed. A reading of onboarding after that change shows 43.4% activation on 2,590 sends: a different stage of the product, so it is not a re-measurement of the rebuild.