The method, before the numbers.
Anyone can publish a favourable statistic. The only way to make a transparency report worth anything is to commit to how you will measure before you know what the measurement says. So this page exists first, and the reports come after it.
There are no published reports yet. When the first one lands it will use exactly the definitions below — and if we have to change one, the change gets its own note explaining why.
Words that usually get blurred.
Most of the ambiguity in this industry lives in one word doing four jobs. Here is how we use each of them.
- Acceptance
- The receiving server took the message and returned a success code. This is what most platforms report as “delivered”. We will always call it acceptance.
- Deferral
- The receiving server asked us to try again later, usually with a reason. Counted separately from failure, because it usually is not one.
- Inbox placement
- Where an accepted message actually landed. We do not currently measure this systematically, so we do not report it. When we do, this definition gets a method attached to it.
- Latency
- Time from our acceptance of your API call to the receiving provider's response. Reported as a distribution, never as a mean alone — a mean latency hides exactly the tail you care about.
- Incident minutes
- Wall-clock minutes during which sending was degraded or unavailable, measured from first customer impact rather than from when we noticed.
Five things, monthly.
Built from anonymised, aggregated production telemetry. Reproducible from the published method, and open to being argued with — including by the providers we are measuring.
Per-provider evidence
Acceptance, deferral, and latency patterns at the major receiving providers, aggregated across all senders.
Incident timelines
Scope, cause, remedy, and whether the remedy worked. Measured from first impact.
Reputation movement
Anonymised signal changes, so the shape of a reputation problem is legible before it is yours.
Remediation outcomes
What we did, and whether it measurably helped. Including when it did not.
Method and limits
Definitions, exclusions, sample sizes, and what each figure cannot tell you.
Monthly · public · reproducible
Same cadence, same definitions, same format every time — so the reports can be compared against each other rather than read one at a time.
What we will not do.
- Publish a figure without the method that produced it.
- Report an inbox-placement number before we can show how it was measured.
- Identify customers, or publish anything that could be traced to one.
- Quietly revise a past report. Corrections get their own entry, with a date.
- Exclude a bad month. The first report we want to skip is the one that proves this is worth reading.
Get the first report.
One email a month, containing the report and nothing else. No drip sequence, no product announcements dressed up as data.
Subscribed
You are on the list. The first report goes out once there is enough production data to say something honest about.
Hold us to this page.
If a future report drifts from the method above without saying so, that is a failure worth telling us about in public.
Open core · never venture-backed · operated by Martin Business Consultants