"I set one up for this magazine in under 30 minutes using free tools that any engineer with a GitHub account can access today."
NIXON V ANTONY · SECOND ENGINEER · MAERSK A/S · APRIL 2026
Most engineers think reader analytics means big corporate budgets, data science teams, and months of infrastructure work. I set one up for this magazine in under 30 minutes using free tools that any engineer with a GitHub account can access today.
But this article is not really about analytics. It is about something more fundamental — the feedback loop that sits at the core of every AI system we discuss in this magazine, and what happens when you apply that same loop to the thing you are building yourself.
The Problem
I publish Marine Intelligence Weekly every week. I write about DeepSea HyperPilot autonomous propulsion, MEPC 84 net-zero negotiations, ClassNK's Genbu certification, OCEANS-X data platforms, and FuelEU compliance deadlines. What I do not know is this: which of those articles do you actually read?
Without that data, I am operating the way a vessel runs without engine room telemetry — making decisions based on assumption rather than measurement. I assume the regulatory deep-dives matter most. I assume engineers care more about IMO compliance than classification society rankings. I might be completely wrong. Every week I make editorial decisions without a feedback signal. That is not a sustainable engineering practice. And it is not how AI systems are built.
The AI Connection
In Andrew Ng's machine learning workflow, every system begins the same way: Define the problem → Collect data → Train the model → Evaluate → Iterate. The system does not guess what the right output is. It receives a signal — a feedback loop — that tells it whether its output matched what was wanted. Without that signal, the model cannot improve. It simply repeats the same behaviour regardless of result.
This is not abstract — it is the operational logic behind every AI system discussed in this magazine, and it applies just as precisely to predictive maintenance on a container vessel. Sensor data comes in. The model outputs a prediction. The prediction is checked against what actually happened to the machinery. The delta between prediction and reality becomes the training signal for the next cycle. My magazine had no training signal. This week, I added one.
The Architecture — Mapped to Andrew Ng's Workflow
READER VISITS ARTICLE
↓
GA4 COLLECTS: time on page, section reached, scroll depth, traffic source, return visits
↓
DASHBOARD SHOWS: which sections readers complete vs exit — broken down by section label
↓
EDITORIAL DECISION: increase depth in high-engagement sections / restructure sections readers exit early
↓
BETTER ARTICLE → higher session duration → measure again → iterate
Tools used: Google Analytics 4 (free), GitHub Pages (free), custom JavaScript events to track which sections readers scroll through. Setup time: 28 minutes. Cost: zero.
The Success Criteria — Defined Before Data Exists
This is the part most editorial projects skip. They collect data, then decide retrospectively what "good" looks like. That is not how Andrew Ng teaches it. You define your evaluation metric first — before you see any results. Here is mine, committed publicly in this issue before a single data point exists:
20%
Improvement in session duration
AI articles vs regulatory-only
Measured at Issue 26 · 8 weeks from now
30%
Issue 18 readers return
for Issue 19
First data published at Issue 20
Primary metric: average session duration per issue. A reader who spends 8 minutes with an issue read it. One who spends 22 seconds did not. Secondary metric: return visitor rate. First results published at Issue 20 — two weeks of live data. I will report the actual numbers regardless of whether they are flattering.
The Limitation I Am Watching
GA4 cannot tell me why a reader left a section. It tells me that they left. This is the same limitation that applies to the LLM-based systems we cover each week. The model tells you the output. It does not show you the reasoning that produced it. The instrument gives you the signal. The engineer interprets it.
GA4 might show me that readers exit the Regulatory section faster than the AI Feature. That could mean: the regulatory content is too dense. Or it could mean readers already know the regulatory story and skip ahead. Or it could mean the section is positioned after a long AI feature and readers are simply fatigued. The data cannot distinguish between these interpretations. I have to reason about it as an engineer would reason about an anomalous sensor reading — the number is real, but the cause requires judgment.
What GA4 Is Now Tracking in This Magazine
Each major section of Marine Intelligence Weekly is tagged with a custom label. The IntersectionObserver API fires a GA4 custom event (section_view) when you scroll 40% into a section. Session duration is tracked per issue URL. Return visitor rate is measured against the previous issue's cookies.
section_view events
avg session duration
return visitor rate
traffic source
scroll depth
Next issue: the first two weeks of data. What the numbers showed. What I changed in the editorial structure as a result. Whether the target metrics moved in the right direction. If they did not, I will report that too. This is the AI workflow applied to editorial work. Define the problem. Collect the data. Measure. Iterate. It works for autonomous propulsion systems. It will work for a weekly maritime magazine.
— Nixon V Antony
Second Engineer · Maersk A/S · April 2026
