A sending dashboard can report excellent delivery while recipients say they never saw the message. The disagreement often comes from different definitions rather than faulty arithmetic. Before changing infrastructure or content, decide which event the investigation needs to explain: a server accepting responsibility, a message appearing in a particular folder, or a person completing an intended action.
Define the event before calculating the rate
SMTP acceptance occurs when the receiving system returns the relevant final success response after message data. It establishes a transfer of responsibility, not an assurance that a person saw the message. An application saying “sent” might mean only that its submission service accepted the request. Identify the hop represented by each event. RFC 5321 sections 3.3 and 6.1.
For a submission-cohort acceptance metric, count each message-and-recipient submission once in the denominator and count it in the numerator if it reaches acceptance at the named hop within the window. Report individual retry-attempt success separately; do not mix those units. State how you treat retries, unresolved deferrals and multiple recipients. Counting every retry as a new campaign recipient inflates the denominator. Counting an unresolved deferral as a permanent failure changes the question. Keep those states separately visible until the observation window closes.
Measure placement with an explicit sampling boundary
Inbox placement requires evidence from a receiving mailbox or another clearly described observation method. A sender-side 250 response cannot tell you which folder was selected. Seed accounts can provide controlled observations, but they are a sample with their own history, settings and engagement. Do not present a small seed panel as a census of the production audience.
Record provider, mailbox type, receiving time, observed folder, sample selection and missing observations. Compare like-for-like samples before attributing a change to a DNS modification. If the panel changed between measurements, show that change alongside the percentage. An absent observation should remain unknown, not be silently counted as inbox or spam.
Treat engagement as an imperfect behavioral signal
An image request is not the same event as a person reading. A missing image request does not prove that a message was ignored. Privacy features and recipient content settings can affect remote-content loading; Apple's documentation describes private loading and cases where remote content is withheld pending permission. Apple Mail remote-content guidance.
Clicks also need an event definition. Deduplicate repeated clicks when reporting unique recipients, disclose any filtering and distinguish the link request from a completed business action. Security scanning and automated clients are reasons to investigate suspicious patterns rather than declare every click human. Use approved application-side completion events when the actual goal is a purchase, booking or confirmed account action.
Work through a bounded example
In a hypothetical hand-worked example, there are 1,000 attempted recipient submissions and 950 accepted submissions, giving 95% acceptance. A separate panel observes inbox placement in 15 of 20 mailboxes, giving 75% within that panel. Of the 950 accepted message-and-recipient submissions in that cohort, 60 have at least one recorded link request within the same observation window. Counting each submission at most once gives 60/950, or about 6.32%. One recipient receiving two messages can contribute two submissions, so this is not a unique-person rate. These are different denominators and cannot be substituted for one another.
The arithmetic uses invented teaching inputs: 950/1,000 for acceptance, 15/20 for sampled placement, and 60/950 for accepted submissions with a recorded link request, deduplicated per message-and-recipient submission in the same window. No script, mail delivery or mailbox inspection is claimed. It is not a customer benchmark or a provider delivery test. Its useful lesson is that 95% acceptance and 75% sampled placement can coexist without contradiction. Nor does the click figure prove that 60 distinct humans read the entire message.
Choose the next investigation from the failed event
Investigate transport responses when acceptance deteriorates. Inspect receiving observations and sending reputation when acceptance is stable but placement changes. Review audience expectations, content and the completion journey when messages arrive but intended actions decline. Keep authentication evidence in the picture without treating it as the answer to every metric.
Close the report with the event, denominator, time window, sample limits and next action. A narrower conclusion supported by observations is more actionable than a single unexplained deliverability percentage.
Choose a metric that answers the question
- Transport: name the accepting hop and recipient-submission cohort.
- Placement: report observed folders and missing cases for the selected panel.
- Engagement: distinguish remote requests from identified human actions.
- Completion: use the approved application event tied to the business goal.
- Comparison: retain denominators, retries, observation window and scope before claiming a change.