Email open rates aren’t dead. They’re pixel loads.

Your email dashboard says someone opened the campaign. What it actually knows is narrower: something requested an image. That something might be a reader’s email app, a privacy service or a security scanner—and a reader who blocks images might never register at all.

In a post about open-rate measurement, Mailgun makes the case for keeping the signal while dropping the certainty attached to its name. The company also describes a reporting mismatch that produced an hourly unique open rate above 100%, without anyone opening more messages than existed.

What happened

Mailgun’s argument starts with the tracking mechanism. When open tracking is enabled, a sender or email service provider inserts a unique, usually invisible 1 × 1 pixel into the message. A request for that image generates an event in the provider’s records. The event establishes that the image was requested, not why.

Several different processes can produce that request. An email client may load the image while displaying a message. An image proxy may fetch it on a recipient’s behalf. A privacy service may retrieve it automatically, a security system may inspect the message, or a cache may fetch the image again. These are different activities sharing a measurement channel.

The blind spot runs in the opposite direction, too. Someone can read every word with remote images disabled and produce no pixel request. Mailgun notes that some spam-folder views also block remote images. A missing open therefore cannot establish that the recipient ignored the email, just as a recorded open cannot establish attention.

Implementation adds another possible source of distortion. Mailgun says it has seen template design prevent tracking pixels from loading correctly, depressing reported open rates. Fixing that problem can expose separate measurement quirks. Before treating a sudden change as an audience verdict, senders need to establish whether the tracking implementation changed.

That distinction gives open tracking a narrower role in email marketing: it records observable image activity rather than a verified reading decision. It also keeps engagement measurement separate from other inbox projects, such as BIMI logos and DMARC enforcement.

The background

Apple Mail Privacy Protection gets much of the blame for unreliable opens because its automatic image retrieval caused conspicuous inflation, Mailgun says. But Google’s image proxy predates that debate: Gmail has served external email images through Google’s own infrastructure since 2013, rather than having recipients request them directly from senders’ servers.

As described in Mailgun’s account of Google’s guidance, the proxy protects recipients from image-based attempts to learn about their computer or location, set cookies or deliver known malicious software. Google also acknowledges that senders may still be able to detect when a message containing images has been opened. Privacy protection does not necessarily erase the event.

Instead, the sender may see Google’s infrastructure rather than the recipient’s IP address, device, browser or location. Caching and repeated requests can also change recorded counts and timing. The consequence is not simply fewer details in a report: it is less evidence for attributing a request to a particular action by a person.

A request from Google is therefore not automatically evidence of an unattended, machine-generated open. The proxy may be retrieving an image because someone displayed the message; other Google systems may retrieve images automatically. Mailgun warns that these events can look similar from the sender’s side despite having different causes.

Security scanning makes a clean split harder still. Mailgun cites guidance from the Messaging Malware Mobile Anti-Abuse Working Group, or M3AAWG, on nonhuman interaction, or NHI. Security systems deliberately conceal scanning patterns so malicious senders cannot easily evade them, and some simulate behavior intended to resemble human activity.

That protective ambiguity creates a trade-off for analytics. More aggressive filtering can remove automation, but it can also discard legitimate activity. A machine may act on behalf of a person, before a person makes a decision, or simply because a message arrived. The same recipient can generate several image requests—or none.

What people are saying

Deliverability expert Laura Atkins calls these events “pixel loads,” according to Mailgun. The terminology matters because it names the evidence rather than its presumed meaning. A pixel load is observable; whether someone saw, read or valued the message is an interpretation layered on top of it.

Mailgun’s own position is not to abandon automation detection. The company says it identifies detected Apple, Gmail and generic automated activity in event data, giving senders more context. Its caution is about the conclusion: even a correctly classified request does not become proof of human attention.

The M3AAWG guidance, as summarized by Mailgun, takes a similarly practical line. Filtering changes reported numbers, but senders cannot know whether they have identified every nonhuman interaction. The effort may not be worthwhile. That leaves teams with a decision to make about what a classification actually helps them do.

Mailgun puts recipient permission ahead of forensic certainty. It asks whether people knowingly subscribed, recognize the sender, still find value in the messages and can easily leave. Identifying every automated request cannot repair a relationship in which the recipient never wanted the email in the first place.

What happens next

For senders investigating strange numbers, Mailgun’s hourly reporting example offers a concrete starting point. Deliveries were counted in the hour they occurred, while unique opens were counted in the later hour when they happened. The two populations were not matched using the Message-ID, the message identifier.

Its illustration starts with 100 messages delivered at 9 a.m., many opened at 10 a.m., and only 20 new messages delivered during the later hour. Dividing those later opens by that hour’s new deliveries can yield more than 100%. Counting each message as uniquely opened only once does not fix an unmatched denominator.

Daily or campaign-level reporting may make that apparent impossibility disappear because more deliveries and subsequent activity fall inside the same window. The practical check is to examine the population and time period behind each number before blaming bots. A different reporting window can change the percentage without changing recipient behavior.

Teams turning events into operational notifications should make the same distinction between an event and its interpretation; email alerts routed into Discord are a related workflow. For open-rate reporting, Mailgun’s recommended question is what decision the classification supports. Correcting an obvious distortion is useful; claiming a perfectly recovered human open rate is certainty the pixel cannot supply.