The work itself · Article 2 of 3
Opens lie, clicks lie less, visits tell the truth
We built our own outreach tracking rather than buy a marketing suite. The honest reading of engagement data is the most useful thing the build taught us.
1 · A two-person firm does not need a marketing suite
Our marketing is a short list of named organizations and a few hand-written notes a week. No blasts.
At that volume the interesting question is not reach. It is whether one specific person a partner wrote to, by hand, engaged at all. The industry answer is to adopt a marketing automation suite, and we declined for the same reason we advise clients to be slow about platform adoption. It was a warehouse where we needed a shelf. It would have dragged our correspondence through a third party’s sending infrastructure. And its dashboards are, to put it professionally, credulous.
So we built the shelf. A small tracking system on Cloudflare Workers, a few hundred lines, our own data, nothing sitting in the delivery path of the mail itself.
This article covers what it does, because the build is a fair sample of how we engineer. More useful to you is the second half, on what engagement signals actually mean, because that is where the industry’s dashboards quietly stop telling the truth.
2 · The build, briefly, for the technically curious
The mechanics fit in a paragraph each, which was rather the point.
A note we send can carry two instruments. A tagged link, which is the site address rewritten per recipient, so that a visit passes through a redirect setting a first-party cookie before it lands on the page. And an open pixel, the classic invisible image whose fetch suggests the message was rendered somewhere. When the person later browses the site, a lightweight beacon reports which pages they viewed, tied back through that first-party cookie. That is the entire sensor package. No third-party scripts, no fingerprinting, nothing exotic.
Behind it, a Worker owns a small serverless SQL database as the system of record, holding prospects, messages, and events. The public website stays a plain website and hands events across an internal service binding, so if the tracking backend is ever down the site serves visitors normally and the event is simply dropped. Our marketing must never be able to break the front door.
Around it sits the boring hygiene most outreach skips. The domain publishes a strict email authentication policy, DMARC at p=reject, so mail claiming to be from us but failing authentication gets refused by receiving platforms rather than delivered to somebody. If you take one operational item from this article, audit your own firm’s SPF, DKIM, and DMARC records this week. It is an afternoon of work that protects your name from spoofing and materially affects whether your legitimate mail lands.
A status note, because this article is about honest measurement. The system is live. The site hands events across an internal binding to the Worker that keeps them, and the thresholds in section four are the ones it actually runs rather than the ones we would like to have chosen. What we will not do is quote you our own engagement figures. A firm citing its own funnel to sell you measurement advice has stopped measuring and started marketing.
3 · What the signals actually mean
Here is the part the vendors will not put on the dashboard. Each signal has a documented failure mode, and honest measurement means pricing it in.
Opens overcount and undercount at the same time. Apple’s Mail Privacy Protection, on by default for most Apple Mail users who were ever asked, routes remote images through Apple’s relays and, in Apple’s own words, “downloads remote content in the background by default, regardless of whether you engage with the email.” An open from such a user is an operating system being polite rather than a human reading. Meanwhile, clients configured to block remote images produce readers who never register at all. One metric, inflating and deflating simultaneously, which is why our system logs opens for context and never alerts on them, and why any vendor showing you open rates to two decimal places is decorating noise.
Clicks are polluted by security infrastructure doing its job. Corporate mail protection rewrites and inspects links. Microsoft’s Safe Links documentation describes URLs without established reputation being “detonated asynchronously in the background,” which is to say fetched by a sandbox to see whether they bite. To a naive tracker, a detonation is a click. The tells are timing and origin: a click arriving seconds after delivery, from a datacenter address, at machine cadence, is a scanner being careful on a human’s behalf. Clicks stay useful. They simply require the judgment the dashboard does not apply for you.
Visits are the strongest signal precisely because they are the hardest to fake by accident. A visit rides the first-party cookie set when a real browser followed the tagged link, and page-view beacons from that browser reflect a person actually on the site. Scanners rarely browse onward. Proxies do not read your services page. The signal’s honest limitation runs the other direction, since cookies are per-browser, so the prospect who clicked on a phone and later researched you on a desktop shows up as a stranger. Visits undercount, and we prefer that. A metric that errs by missing engagement is safe to act on. A metric that errs by inventing engagement, as opens do, is a machine for fooling yourself.
4 · The policy layer: alerts that respect attention
Sensors are the easy half. The system’s real design lives in what it refuses to tell us.
Clicks and visits notify instantly, because for named-account outreach the knowledge that a managing partner is on your services page right now is precisely the moment a human should hear about it. But notifications are deduplicated, one per signal type per prospect per six hours, and a visit within thirty minutes of a click counts as the same session and stays quiet. Opens never notify at all. Everything lands in a weekly digest, which also lists the prospects with no engagement whatsoever, because at our volume the silence is a finding too.
Every threshold in that paragraph is an opinion, and the opinions are encoded in the system rather than in anybody’s memory, with the decision recorded on each event row so the deduplication is exact and auditable afterward. This is the doctrine our governance work applies to client AI systems, in miniature. The policy lives in code. The code’s decisions are inspectable later. And the scarce resource being protected is human attention, because an alert channel that cries wolf stops being read, and then the one signal that mattered arrives unread with the rest.
5 · What this buys you
Nothing about our little tracker is proprietary genius. It is a shelf, well made. Its value to you is the posture it demonstrates, and the posture transfers directly to the systems you already pay for.
Your email platform reports opens, and now you know what those contain. Your CRM scores leads on engagement, so ask which signals feed the score and whether anybody priced in the scanners and the proxies. Your analytics count things, so ask which of those counts can only err downward, and act on those.
When we run the technology side of a firm this interrogation is standing practice, applied to marketing data, to vendor reporting, and to the AI-flavored scoring products now arriving with confident decimals attached. It routinely changes decisions that were about to be made on inflated numbers.
Measurement you can trust is quieter than measurement you cannot. It says less, and what it says, you can act on.
The final article in this series applies the same principle to a much bigger surface: how a firm’s accumulated knowledge, its mail and files and tasks and calendars, can be made searchable by AI without copying any of it into a second system that then has to be defended.