How to Become a Technical Marketer: A 9-Month Case Study

Every guide to becoming a technical marketer measures the skill by what you can build. Nine months of building says the thing that decided the outcome was whether I could see what changed and put it back.

Related: 794 European finance jobs, decoded.

Victor Roodhorst
24 July 2026
9 min read

In April, Emily Kramer published the team half of MKT1’s State of Marketing Report: marketing headcount, titles and open roles across 100 fast-growing B2B companies. The number everyone quoted was that 84% of the job descriptions mention AI. The detail I keep returning to is buried in her methodology note. She assembled the dataset with Claude Code and a set of MCP servers, then published the working.

That is a more useful answer to “what does technical mean for a marketer now” than any of the skills lists, because it is a marketing operator doing the research job in code rather than briefing it out. Three months earlier, LinkedIn had published its Skills on the Rise ranking for 2026, measuring which skills people added to profiles and which ones correlated with actually getting hired. AI literacy came second. Go-to-market strategy came eighth. There is not a single programming language in the top ten.

Those two facts sit oddly together only if you think technical means “can write code”. I gave myself nine months to find out what it does mean, from a seat where the answer has consequences: a marketing job inside European asset management, arriving from a decade of aggressive B2C performance acquisition where I could push a change live on a Friday and read the result on Monday. Here is the ledger, including the parts that went badly.

The trend is real, and it is not a coding trend

Take the market signal seriously first, because it is genuinely strong. Kramer’s 100 companies had 507 open marketing roles between them. Eighty-seven of the hundred were hiring. Beyond the 84% AI mention, 18% of descriptions now reference agents or agentic workflows, a phrase that barely existed in a marketing JD two years ago. LinkedIn listed over 3,000 open go-to-market engineer roles in January 2026, a title that was a niche curiosity in 2024.

Now look at what those descriptions actually name. Figma appears in 16%, Salesforce in 11%, HubSpot in 8%, Claude in 7%, Clay in 6%, Cursor in 5%, Replit in 4%. The named coding-adjacent tools are the smallest slice of the chart, and no one is asking for Python by name. The market is not asking marketers to become engineers. It is asking them to become the person who can point a machine at a job and tell whether it did it properly.

The content around this has caught up fast. Lisa Peyton is giving away a bundle of Claude skills for B2B content marketers built explicitly around thinking better before you write, rather than writing faster. The most-shared posts in my feed this month are people publishing the skills themselves as artifacts, including one built to predict which posts will travel. This is a good genre and I have learned from it. It also shares a blind spot, which is the subject of the rest of this piece.

The ledger: nine months, six builds

I did not take a course. I picked jobs I already did by hand and tried to hand them to a machine, one at a time. Six of them survive in some form.

What I built	The job it does	Where it landed
A research agent on a paper-trading account	Pre-market brief every weekday at 06:00, built from news and filings	Running. Paper money only, which is the point. The value was learning to read what an agent got wrong before the market opened.
A European financial services news digest	Watches feeds, writes a weekly digest to a spreadsheet	Running after one full rebuild. The first version was clever and unmaintainable.
An SEO pipeline for this site	Crawls the site, stores a baseline, reports what changed since last week	Running weekly. The most useful thing I have built, and the least impressive to describe.
A price-intelligence prototype for a family business	Tracks co-product prices across a small supplier network	Shipped small and still small. Deliberately.
A weekly portfolio dashboard	Pulls positions, writes a memo, flags what moved	Running. Ten minutes a week I no longer spend.
762 AI-written pages on this domain	Was supposed to win long-tail search	Deleted. The full autopsy is below.

Five of six still run. On the strength of that ledger I could write the usual post about how any marketer can now build anything. I do not think that is the lesson, because the two things that went wrong were far more instructive than the five that went right, and they went wrong in exactly the same way.

Two failures with the same shape

Failure one: scale without a check

The 762 AI-written pages produced 31 clicks and 482 impressions across twelve months. Sixty of the URLs ever surfaced in search at all. Fourteen received a single click or more. Over 99% carried no image. I have written that story up in full elsewhere, so I will only take the structural point here: nothing about the machine failed. The prose was clean. The publishing worked. What was missing was any step between generation and publication where a human looked at the output and asked whether a person would want to read it.

Failure two: speed without a diff

In June I pointed a coding agent at this site to strip leftover CSS out of twelve blog posts. It ran a pattern match, found the rules, and moved them to the global stylesheet. Correct job, about four minutes. It had also matched a stylesheet living inside a JavaScript string in a PDF-export widget, so it lifted fourteen unscoped rules for body, h1 and table into the global CSS of every page. For roughly twenty minutes, my site's typography was whatever a PDF template thought it should be. I have written up what that setup looks like on a Windows machine, including the parts the Mac-written guides skip.

Twenty minutes, on a personal site, is a funny story. The same class of mistake against a live campaign, a CRM segment or a client-facing page is not funny at all.

Both failures are the same failure. In each case the machine did precisely what I asked, at a scale or speed where I had no way of seeing the result before it was live. Not a capability gap. A verification gap.

Learn the diff, not the language

The single most valuable technical habit I picked up in nine months was not a language. It was version control, and specifically the habit of reading a diff: a line-by-line view of exactly what changed between the thing that worked and the thing in front of you now.

That sounds unglamorous next to building an agent, and it is. It is also why the twenty-minute CSS incident was twenty minutes rather than a weekend. The change was visible and reversible. Every subsequent build I have kept has that property, and every one I abandoned did not.

So the test I now apply before letting any tool touch anything that matters is three questions, and I would happily hand it to any marketer starting this month. Call it the undo test.

1. Can I see exactly what it changed? Not a summary of what it says it did. The actual difference.
2. Can I put it back in under a minute? If reversing the change requires reconstructing something from memory, it is not reversible.
3. Can I show both of those to a third person without preparing anything first?

The third question is the one that separates a personal experiment from something that can live at a real firm, and it is the one no tutorial asks.

What a regulated seat changes

Almost everything written about the technical marketer is written from inside a software company, by someone who can connect a tool to production data on a Tuesday afternoon. That is a real and useful vantage point. It is not mine, and it is not the one most people reading this from a European bank, insurer or asset manager are working from.

In a regulated firm the binding constraint is not capability, it is perimeter. You cannot point a general-purpose agent at client data. Any tool that touches a business process eventually attracts a conversation with operational resilience, and any tool that makes or shapes a decision attracts a second one about how it is classified under the AI Act. Those conversations are not obstacles invented to slow marketing down. They are the same question I was asking myself after the CSS incident, asked by people whose job it is.

Which is the genuinely good news in this piece. The undo test is not a compliance tax bolted onto the fun part. It is the same discipline, written down. A marketer who arrives at the risk conversation already able to say what the tool touches, what changed, and how it gets reversed is not asking for permission. They are handing over evidence. I came into this industry from a world where the answer to almost every question was another test, and the measure-everything reflex transferred cleanly. What did not transfer was the time horizon. B2C optimises for the quarter. This world compounds over years, and so does a badly documented automation.

What I would do in the first thirty days

If I were starting again in August with what I know now, the order would be this, and none of it requires a course.

Week one: version control, before any language

Put your working files in a repository. Campaign briefs, spreadsheets, prompts, whatever you already have. You are not learning git to become a developer, you are building the muscle of seeing what changed. Everything after this is safer because of it.

Week two: automate the boring job, not the impressive one

Pick something you already do by hand every week and can personally check in under five minutes. A report pull. A data clean-up. A monitoring digest. The impressive builds are the ones you cannot verify, which makes them the ones that fail quietly.

Week three: build the check before you scale the build

Whatever you automated, add the step that tells you when it is wrong, and make it something a person actually reads. This is the step that would have caught 762 empty pages in week one instead of month twelve.

Week four: write down what it touches

One page per tool. What data it can reach, what it can change, who approved it, how you reverse it. At a regulated firm this is the document that turns a side project into something you can defend. Everywhere else, it is the document that means the person who inherits your work can maintain it.

Notice what is absent. There is no language to choose, no bootcamp, no stack to commit to. The tools in the chart above will have rotated by next year. The habit of seeing what changed will not.

Frequently asked

How do you become a technical marketer?

Start by putting your work under version control and learning to read a diff, before you learn any programming language. Then automate one boring task you already do by hand every week, build the check before you scale the build, and write down what each tool is allowed to touch. The 2026 skills lists back this up: LinkedIn's ten fastest-growing marketing skills contain no programming language at all, with AI literacy at number two and go-to-market strategy at number eight.

Do marketers need to learn to code in 2026?

Not in the sense of writing code from memory. In MKT1's April 2026 dataset of 100 B2B companies, 84 percent of marketing job descriptions mention AI and 18 percent mention agents, while named coding tools appear far less often: Cursor in 5 percent and Replit in 4 percent. The demand is for people who can direct and verify machine output, not people who can recall syntax.

What is a GTM engineer?

A go-to-market engineer is a commercial operator who builds and maintains the systems that run go-to-market: data enrichment, routing, automation, agent workflows and the reporting on top. The title grew quickly through 2025 and 2026, with LinkedIn listing over 3,000 open GTM engineer roles in January 2026. The work is closer to operations engineering than to campaign management.

What is the undo test?

Three questions to ask before letting any tool or agent touch a live system. First, can I see exactly what it changed. Second, can I put it back in under a minute. Third, can I show both of those to a third person without preparing anything first. A tool that fails the undo test is not ready for a production system, and in a regulated firm the third question is the one your risk function is actually asking.

How is becoming technical different inside a regulated firm?

The constraint is not capability, it is perimeter. In a regulated financial services firm you cannot point a general-purpose agent at client data, and any tool that touches a business process attracts questions from operational resilience and AI governance. That makes the auditability of your setup, meaning the record of what changed and who approved it, more valuable than the sophistication of what you built.

What should a marketer automate first?

The boring repeated job, not the impressive one. Pick something you already do by hand every week where you can check the output yourself in under five minutes: a recurring report pull, a data clean-up, a monitoring digest. Automating a task you cannot personally verify is how teams end up shipping errors at machine speed.

Victor Roodhorst works in marketing in European asset management and publishes the AI Share of Voice Index for European Asset Management. He came to the industry from B2C performance acquisition, which turns out to be a useful place to have come from. Reachable here