user@gitdiot:~/blog/how-i-operate/email-opens-delivery-receipts$
● online ~/index ~/about
$ cat ./content/operating/email-opens-delivery-receipts.md --render
~/blog/how-i-operate/email-opens-delivery-receipts
how-i-operate email-tracking deliverability lead-scoring Oct 8, 2026 · 10 min read

Most of our email opens were delivery receipts.

I went back through every open our tracker logged, one request at a time. The dashboard had been counting mail servers, security scanners and our own phones.

In March I wrote about the tracking image we use: the company logo in the signature, served from a per-email URL, so the tracker is something the email needs anyway. Near the end of that post I said opens are a weak signal and should be treated as prioritization input.

This spring I measured how weak. The answer changed how I think about the most quoted number in email.

The setup

We sent about 1,100 outbound emails through AgentCRM, nearly all of them in April and May, to marketing and technology leaders at agencies and enterprises. The leads dashboard looked healthy: lots of opens, a few dozen clicks, a pipeline worth talking about.

Then a simple question ("which of one rep's accounts are ready to email again?") turned into a long session with Claude Code going through the raw tracking table. Each row is one HTTP request for our logo. We grouped them by network, client, and how long after sending they arrived, and looked at what each group actually was.

1,694 image requests came back on those emails. More than 80% of them, on inspection, had nothing to do with a person reading.

Who actually loads the image

An open is a request for an image. The tracker can't see a person. It sees a network address, a user agent and a timestamp, and anything that renders the email's images makes the same request a reader would. Sorting by who made the requests gave five groups.

Apple's privacy relay

Apple Mail Privacy Protection downloads every remote image in every message when it arrives, through Apple's relay network, whether or not the message is ever read. Apple documents this, and our data matched: 410 requests came through the relay, almost all at delivery.

So a relay hit tells you the message reached a mailbox read in Apple Mail with privacy protection on. It tells you nothing about whether anyone read it.

Corporate security gateways

Large companies inspect mail before it reaches the inbox. Microsoft Defender, Proofpoint, Mimecast and the like render the message, load the images and follow the links to check them.

The tell is timing. Mail infrastructure is uniform: a gateway handles every message to a company the same way. We found 81 cases where one network and one client fetched most of a company's emails from that day within the same 30 seconds of delivery. Coverage was high and the spread was tight. None of them spread out beyond five minutes. A batch of humans doesn't behave like that.

Us

When a rep scrolls the Sent folder on a phone, or a mail app syncs an account for the first time, every message on screen loads its images. That includes our logo, with a tracking ID that still belongs to the recipient.

More than half of all opens that arrived a day or more after sending were our own mailboxes. They're easy to spot once you look for them: one sender's emails to many different companies, all fetched in one sitting from the same client. One rep's mail app re-loaded 128 sent emails across 15 companies in a single hour. Another day, my own mail app did 65 across 12. Our office and server addresses were already filtered out, but the list was incomplete, and phones on cellular networks never match an address list anyway.

Security tools that come back later

The strangest group showed up on one afternoon in May. Security tooling at three unrelated companies, each on its own Microsoft tenant, re-opened our emails within about three hours of each other. The messages were five or six days old. Each re-open loaded the image and tested the link in the same second, an exact replay of what the gateway had done at delivery.

It happened once and never again. It hit only the emails from the mailbox dedicated to each of those companies, and those emails were paraphrases of one template: same opening claim, same closing line, same link. Microsoft documents automated investigations that search a whole tenant for messages similar to a suspicious one. Something like that is the likeliest explanation, though I can't prove it from the outside. What matters for measurement: nobody opened anything.

It also gave us a governance ticket. We weren't sending duplicate emails, but repeating one template's structure inside one company looks like a campaign to a filter. That's now a pre-send check.

Webmail proxies

Gmail fetches images through its own proxy, so you see Google's address instead of the reader's. That's usually a real view. But Gmail caches the image after the first load, so repeat views are invisible, and some of its fetches arrived within a minute of sending, faster than anyone plausibly read.

Clicks have the same problem

I assumed clicks were cleaner, since a click takes intent.

Gateways follow links to check them, often within seconds of delivery, and the better ones use real browsers on networks that rotate around the world. Of 623 tracked clicks, our tracker had marked 40 as human. After reviewing each one, 12 looked like a person clicking.

Most of the rest had a signature: an open at 5 to 60 seconds after delivery, then a click at 18 to 140 seconds, repeated at every recipient in the company. One company's checker came from a different network for each recipient, all on the same 30 to 70 second rhythm.

This mattered more than it sounds. Our lead score gave up to 25 points for landing on a high-intent page, and the code read the page from every link event, including the ones already marked as scanner traffic. A gateway checking our migration-guide link gave every recipient at that company credit for visiting it. That one bug produced most of what the dashboard called "nurture."

What an open is actually good for

An open is good evidence of delivery. It says the message reached a real mail system, didn't bounce, and made it through the security layer. A relay fetch says it landed in an Apple Mail inbox. A gateway fetch says it reached the company and was inspected.

Used that way, opens are a deliverability instrument:

What an open can't do, alone, is tell you someone read the email.

How we score reading now

We rebuilt the scoring around a few rules. These are decisions, with numbers attached, and we're about to test each one against mailboxes we control.

Weight every hit. Each request gets a number from 0 to 1 for how likely it is that a person caused it. A relay fetch at delivery is 0. A mail app on a phone the next afternoon is close to 1. An AI assistant summarising the inbox for someone counts too, at 0.2, because a digest is a lesser read. Weights replace the old human-or-bot verdicts, so nothing gets thrown away; it just counts for less.

Use the company as the baseline, to loosen. If we send ten emails to a company and the gateway fetches nine of them identically at delivery, that pattern is the machine and weighs zero. A fetch that breaks the pattern, hours later on one message, is a person or a scan triggered by a person opening it. We weight it 0.8, the same as a Gmail open. Intent doesn't depend on whether the company runs Microsoft or Google, and when in doubt I'd rather assume interest than blame some unknown security layer.

Count reads. A read is a credible hit, and hits on the same email within an hour are one read. A second read, on another day or another message, means someone came back. It's worth 25 points, more than a single click.

Name tiers after evidence. Our old tiers were score bands, and they'd been cut on numbers full of scanner noise. The only reply in the whole program scored "warm," because a reply added 20 points. Now the tier says what we saw, and points only order people inside it:

Tier What we saw
Hot A reply or a booked meeting
Warm A click plus another signal: a second read, a real site visit, or a form
Engaged A click, a second read, or a site visit with real time on the page
Read One credible read
Inconclusive Something a person might have caused
Not seen Nothing we can attribute to a person

Never read "not seen" as "not interested." Of everyone we emailed, about a quarter read in Apple Mail, where reads are invisible by design, and roughly another quarter sit behind gateways that hide the reader. Another third produced no requests at all. "Not seen" describes the sensor, so the tier carries a note saying which of those it is, and nobody gets dropped from follow-up because we couldn't watch them read.

We also switched off the age discount for a few months. Everything in this data is five months old, and halving every score made the tiers meaningless. Age is now a tie-breaker.

What changed

Run against the same data, the pipeline shrank and got more honest. Contacts who looked engaged because a gateway walked their links, or because a rep scrolled the Sent folder, dropped to "not seen." People who read twice, and had been sitting in cold because no single hit looked dramatic, moved up to Engaged. 39 contacts are Engaged now, whichever mail client they read in.

The one person who replied is at the top.

It's a less flattering dashboard. It's the first one I'd hand to a salesperson as a call list.

If you run outbound

  1. Filter your own team, and treat one sender's old emails opening across many companies in one sitting as your own Sent folder. An IP list won't catch phones.
  2. Separate Apple relay hits from opens. Count them as delivery.
  3. Look at timing within a company. If everyone there "opened" in the same 30 seconds, nobody did.
  4. Treat a click in the first minute after delivery as a link check until something else says otherwise.
  5. Check that your page-visit credit reads only human clicks. Ours didn't.
  6. Ask whether you could have seen a read before you call someone cold.

Next we're sending through the production pipeline to mailboxes we own, across Outlook, Gmail and Apple Mail, with and without Defender, logging every action to the second. After that, every weight in this post will rest on a measurement, and I'll write up the ones that survive.

subscribe.sh

Get the field notes

Weekly dispatches from an aging tech worker's refactoring. No spam, no thought leadership.

no spam · only high-signal logic