Stop Running Campaigns, Start Building Growth Machines: The 5 Steps to Work Like a Growth Engineer!
TL;DR: The first working version of a software product is nearly free these days. Distribution is not, and the execution of marketing is moving to machines. The founders who win this game stop running campaigns and start building growth machines. Here are the five steps I use to build one, from the Single Point of Truth to the first running loops. I call the method Growth Machine DNA (GMDNA).
Why Distribution Is the New Bottleneck
Two numbers from this summer frame the shift. Meta's Q2 results showed its Advantage+ suite, the AI system that writes, tests and reallocates ads largely on its own, passing a $75 billion annual revenue run-rate. And on Lovable, the Stockholm vibe-coding platform, users now start one million new projects every week.
Let me be precise about what that second number means, because this is where the story usually gets overstated. What has become nearly free is the first working version of a product. UX, testing and taste still need humans, and the more human-centred the product, the more of them it needs. But that precision makes the problem worse, not better: the market is flooding with decent first versions. Getting found and getting chosen is now the scarcest part of building a company, and the machinery that decides both is increasingly automated.
So the real question for founders is not whether you can build your product. It is whether you can build the machine that reliably finds and keeps your customers.
From Growth Hacker to Growth Engineer
Marketing went analytical around 2010, when Sean Ellis coined "growth hacking": tracking, attribution, A/B tests. The next step is happening now: AI agents do the execution itself, from creative to copy to budget shifts. The person who thrives in this setup is the growth engineer: a builder who is measured on lift in one metric (sign-ups, activation, or revenue), and who orchestrates a stack of agents instead of running campaigns by hand.
The role is going institutional: Andreessen Horowitz launched a Growth Engineer Fellowship this year, and OpenAI has a dedicated GTM growth engineering team. In Episode #29 I covered agentic productivity, in #30 the shift to Service-as-a-Software. In this edition I show how growth engineering works in practice.
I hold this role myself at one of my ventures, and I help startups build the same setup. Here is the sequence, in the order that works.
Step 1: Build Your Single Point of Truth
Before any agent runs, build the repository that everything else derives from: one Single Point of Truth for everything the machines cannot invent, from your brand and personas to tone and voice, value propositions, competition and market trends. This repository is the DNA of your growth machine, and it is why I call the method Growth Machine DNA (GMDNA): the go-to-market counterpart to my BMDNA framework, which decodes the genome of a business model.
This is where purpose gets encoded into the machine. When agents later run thousands of touchpoints, this repository is the reason every one of them still sounds like you. It is also your only durable edge on this layer: everyone can rent the same models, and two competitors prompting the same system get roughly the same ad. The difference is what you feed it.
What belongs in the repository is easy to guess. How it is structured, maintained and kept in sync with reality, so that dozens of agents work from it consistently without drifting off brand, is the actual craft. And this step carries extreme leverage: like in real DNA, a defect here does not stay local. It gets replicated into every campaign, every ad and every touchpoint the machine produces.
Step 2: Wire Your Pipelines
Now build the plumbing, in two directions. First, take your existing campaigns and rebuild them as loops: agents draft variants, launch small, measure lift against your one metric, kill the losers, scale the winners. Second, build pipelines into the channels you do not serve yet. The newest of these deserves its own pipeline: buyers increasingly get their shortlists from AI answers rather than from a results page, and Ahrefs' latest study measured a 58% click-through drop for top-ranked pages where Google's AI answers appear. Visibility there is a build task with shipped artefacts, not a media buy.
Start with one channel, not five. A single loop that demonstrably lifts your metric earns you the credibility, and the learnings, to wire the next one. Most teams that try to automate everything at once end up with five half-connected pipelines and no lift at all.
The integration layer is where vibe engineering shines. Connect your ad platforms, CMS, CRM and analytics through official APIs, MCP servers, or connectors you vibe-code yourself; with today's coding agents this takes days, not quarters, and tools like n8n wire the agents together. At one of my ventures the entire ad account runs through scripts I vibe-coded myself, from building the campaigns to the daily checks. Which scripts, and with which decision logic, is a longer conversation. The point here is that the account runs better than it did by hand, with a fraction of the manual work.
Step 3: Generate the Content From the Source
Only now does content enter the picture: posts, ads, graphics, videos, landing pages, all drafted by agents from the Single Point of Truth, all curated by a human before anything ships. The order matters. Teams that start with content tools and skip Step 1 produce faster generic output, which is worse than slower generic output. Salesforce's State of Marketing study puts AI adoption among marketers at 75%, and finds most of that usage producing the same one-way messaging, only faster. Adoption is not advantage. Your source repository is.
The human role in this step is curation, and it is not a formality. The agents deliver volume and consistency; you deliver taste. In practice that means a hard review gate with a handful of questions before anything ships. Which questions, and how strict the gate, depends on your brand and your risk; that the gate exists at all is the difference between a distinctive presence and automated noise.
Step 4: Make Tracking Non-Negotiable
Before a single loop goes live, make sure the machine can see. The ad platforms need your conversions, your dashboard needs every channel separately, and AI-assistant referrals deserve their own tracking, because they behave differently from search traffic. For EU founders, GDPR and the AI Act's transparency rules belong in the definition of done here, not in a legal review afterwards.
A first version of the scorecard, your one metric broken down by channel and loop, is an evening of vibe-coding; a version your team actually trusts takes a few more iterations. Build it anyway, and give it to whoever needs the answer: no BI ticket queue, no quarterly reporting project. If the scorecard project outlives the first campaign, something is wrong with the order of operations.
I learned the importance of this step the honest way: at my own venture, the true bottleneck was not creative or budget, it was a tracking gap that kept purchases invisible to the platform. Everything the machine optimised before that fix was noise. Tracking is prerequisite number one; without it, every later decision is guesswork dressed up as data.
Step 5: Start the Loops, Stay in the Loop
Now the loops go live, and your job changes: you become the human in the loop and the orchestrator of the whole machine. You define what a winner looks like before a test starts, in writing, so that the result decides and not the mood of the day. You review on a schedule, thirty minutes each morning: read what the loops did overnight, make the kill and scale calls, write what you learned back into the Single Point of Truth. That last part is the compounding engine of the whole machine: every review makes the repository smarter, and every agent works from the smarter repository.
The shift is easiest to see in my own calendar. Campaigns used to be handwork: briefings, drafts, uploads, reports, spread across the whole week. Today the machine does that part overnight, and my mornings are thirty minutes of decisions. Same channels, same budgets, a different job.
Two things separate the professionals here. The first is restraint: the platforms' bidding algorithms need stable learning phases, so a system that changes everything daily performs worse than one with a disciplined cadence. The second is patience: some tests need longer runtimes and a sufficient data basis before they mean anything, and killing them early means deciding on noise. When generating fifty variants costs nothing, the bottleneck moves from production to judgement.
Those principles fit in a newsletter. The parameters behind them, how much data is enough, which cadence fits which platform, when a quiet week is noise and when it is signal, are learned on a live system.
The Machine Is Buildable
That is the Growth Machine DNA method: one source of truth, wired pipelines, content from the source, tracking first, then loops with you in the loop. Every piece of it is buildable today with tools you already have access to, and it scales with you: a solo founder runs a small version of this machine alone, a Series B team runs it with two or three growth engineers. The founders who build it this quarter will compound learnings while their competitors are still approving campaign calendars.
Which of the five steps is your bottleneck right now? Tell me in the comments or send me a DM. I read everything, and your questions shape the next episodes.
And if you want hands-on support building your growth machine, from your GMDNA to the first running loops: that is exactly what I do with startups and scale-ups. Reach out.
I speak tech, product and business. Let's build! 🚀
Visit my website: www.arndtschwaiger.com
Thank you for reading and sharing!
Best regards,
Arndt
