An Email Open Is No Longer an Event. It’s a Guess.
ENGINEERING NOTES / EMAIL & MEASUREMENT
Sharad Chandel · 02 October 2026 · 7 min read
Imagine a campaign report says 42% of recipients opened an email. What does that number mean? Usually, it means the email platform received a request for an image embedded in the message. We call that an open because it has long been a useful proxy for one. But a proxy is not the event itself.
That difference matters more as email clients protect privacy, security systems inspect messages, and AI features summarise email without asking people to read every line. The uncomfortable question is not whether we can keep counting opens. It is whether the count still measures the thing we say it measures.
01 / What does a pixel actually observe?
The mechanism is simple. An email includes an image hosted on a server. Often it is a tiny, invisible image with a URL unique to a recipient or campaign. When an email client requests that image, the sender’s system records the request. Email platforms commonly report that request as an open.
There is an assumption in that last sentence. The request is observable. The person reading the message is not. A pixel can tell us that its URL was requested; it cannot tell us whether a person saw the message, read it carefully, or understood it.
Even before AI, that assumption had exceptions. A person could read a plain-text message or block remote images, leaving no pixel request. Another system could retrieve remote content for privacy, caching, or security reasons. Both cases separate the recorded event from the human action the dashboard implies.
This is not a reason to say opens are worthless. A consistent proxy can still be useful for comparing campaigns or spotting broad changes. But the result depends on the behaviour of the software between the recipient and the sender. The number is not a direct count of attentive readers.
02 / Email clients have already changed the signal
Apple’s Mail Privacy Protection is a clear example. Apple says that, when Protect Mail Activity is enabled, remote content is privately downloaded in the background when a person receives a message, rather than when they view it. That makes an image request a poor indicator of when—or whether—the person opened that message. Apple Support: Protect email privacy in Mail
Gmail takes a different approach. Google says Gmail serves message images through its secure proxy servers, and its help documentation explains that images may not display automatically for messages it considers suspicious. Proxying can protect people and reduce what the sender learns about their device or location. It also means the image request does not necessarily come straight from the recipient’s device. Google Workspace: Image URL proxy · Gmail Help: Turn images on or off
These systems were not built to help marketers measure opens. They were built to protect users and make email safer. The change in measurement is a consequence of those choices.
The uncomfortable part is that the dashboard often preserves the old label. A proxy fetch may be counted as an open; a person who blocks images may not be counted at all. Depending on the client, the number can move in either direction without a corresponding change in human attention.
03 / AI adds a different kind of ambiguity
Email is also becoming something software can process on a person’s behalf. Gmail’s Gemini features can summarise an email thread. Outlook’s Copilot can scan a thread and produce a summary. These are documented product features, and they show that a person can get value from email content without reading every message in full. Google Help: Gemini in Gmail · Microsoft Support: Summarise an email thread with Copilot
That does not establish that these features fetch tracking pixels, or that every AI assistant reads email the same way. We should not claim that. The relevant point is narrower: email content can be processed to produce an answer or summary, and that processing is not the same event as a person opening a message in the traditional sense.
If an AI system or another automated tool processes the message without requesting its remote images, the pixel may not fire. If a privacy or security layer requests the image, the pixel may fire without the person reading the message. Those are plausible paths, not a reliable way to identify an AI reader from a single request.
We end up with several distinct events that are easy to collapse into one word:
A message was delivered.
A remote image was requested.
A client or assistant processed some of the message.
A person saw or acted on the message.
These events may be related, but the pixel alone cannot tell us which sequence occurred. Treating them as interchangeable gives a clean-looking metric by hiding the uncertainty that matters.
04 / Can we make the open more reliable?
It is tempting to respond with better classification. Look at the request time, IP address, user agent, proxy characteristics, repeat fetches, or other signals. Some patterns may help separate likely human activity from likely automated activity.
But a pattern is evidence, not identity. Privacy proxies deliberately hide or transform information. Security products change. A user agent can be shared or misleading. Systems can fetch once and cache, so a later human view may produce no new request. Rules that fit one mail client or scanner may fail for another.
There is a hard limit when the only thing we can change is the email HTML. We can include a pixel URL, perhaps vary it by recipient, and observe requests that reach our server. We cannot make an email client or AI assistant disclose why it fetched that URL. Nor can we assume an assistant will call a separate endpoint just to report that it processed the message.
A classification system might still improve the estimate. It could label an event as likely automated, likely human, or unknown, and show what evidence supports that label. But that would be a better account of uncertainty—not proof that a human read the email.
This distinction matters if a customer asks for “reliable open tracking.” Reliable at what? Reliably recording image requests is achievable. Reliably identifying the actor may be possible only for some cases. Reliably confirming human attention from HTML alone is a much stronger claim, and the pixel does not provide that evidence.
05 / Measure the outcome you care about
If the purpose of an email is to get someone to visit a page, submit a form, reply, or make a purchase, measure that action where it happens. Those signals are closer to the outcome than an image request.
They still do not prove that a person read the email. Link scanners can follow URLs; a click does not tell us how carefully a message was read; a conversion may happen after someone sees the same offer elsewhere. The lesson is not that another metric is perfect. It is to choose a metric that answers the question we actually have.
For campaign diagnostics, opens may remain one imperfect signal among others. For a decision about whether people found the message useful, replies or downstream actions may matter more. The right measure depends on the job of the email, and every measure needs a clear definition.
I would want reporting to distinguish what was observed from what was inferred: “image requested” is an event; “likely human open” is a classification; “recipient understood the message” is not something the pixel can establish. That vocabulary may be less convenient, but it is more useful when people are making decisions from the data.
Conclusion / Keep the uncertainty visible
An email open was never quite the event we named it after. It was a remote image request used as a proxy for a person opening a message. Privacy features made the gap visible. AI-assisted email makes it more important to ask what counts as processing, viewing, and acting.
There may be ways to classify some requests more accurately. That is different from confirming that a human read an email. With only the email HTML to work from, we should be careful about promising more than the channel can show.
The useful question is not simply “Did they open it?” It is: What did we observe, what are we inferring, and what decision is this measurement meant to support?