Everyone uses AI.
Almost no one has a system.

Why now

The marketer is becoming an engineer

McKinsey’s 2025 State of AI survey found that 88% of organisations use AI in at least one function, but only 7% say it is fully scaled. Almost everyone has access to the same models. Very few have built anything around them that compounds.

Each era of marketing added a new layer of leverage. Traditional marketing ran on story and psychology. Digital added funnels and analytics. Growth hacking added product loops and the AARRR framework. The fourth era adds continuous learning: systems that ingest market signal, run experiments and update their own playbooks, with a human setting direction.

The problem

Scattered prompts vs systemic memory

Most teams are in the void. Someone types “write 10 posts” into a chat window, gets ten passable posts, and starts from scratch again next week. Nothing the model learned about the business survives the session.

The system looks different. A growth repo holds what the business knows, and research, content and optimisation agents all read from it before they act. Marketing memory compounds instead of evaporating.

Adoption is near-universal and durable advantage is not. Only around a third of organisations have begun scaling their AI programmes. The gap between using AI and running on it is the whole opportunity.

Step 1 · Initialise

Build the Growth Repo

The repo is a version-controlled folder of plain-text context that every agent reads before it acts. This is the difference between AI generating random copy and AI studying your business: feed it the actual receipts.

Five folders carry the whole system:

• customer-truth/ holds sales calls, support tickets, churn notes and live product feedback.
• content-engine/ holds founder voice guides, winning hooks and performance notes.
• outbound-engine/ holds ICPs, trigger events and banned SDR language.
• creative-testing/ holds ad angles, landing-page tests and offer results.
• agent-specs/ holds job definitions, triggers and expected outputs for each AI worker.

Architecture

The marketing engineer’s tool stack

Four roles, not forty apps. Each layer has one job, and all of them read from and write to the same repo.

• The edge watcher. Connected to social ecosystems; watches competitors, creators and shifts in cultural language in real time.
• The builders. Coding agents that write landing pages, build internal tools and structure repo data.
• The schedulers. Run scheduled operations with persistent memory and human-in-the-loop approval.
• The secure vault. Local models that process sensitive, regulated or private data on your own hardware, without calling cloud APIs.

System 1

The Customer Truth Engine

Pipe every raw customer signal into one processor that writes a living markdown brief. Sales calls, Stripe data, support tickets and CRM notes go in. A file called what-the-market-is-telling-us.md comes out, and every line in it links back to its source.

The goal is not a vague summary; it is insight with receipts that makes the business harder to lie to. A generic output says “HVAC customers want better collaboration.” An engineered output says which calls cited emergency dispatch versus the calls that actually converted, and quotes the missed follow-ups.

System 2

The Founder Content Engine

Record the founder talking: customer calls, podcasts, voice notes. An extraction agent pulls claims, stories and contrarian takes, then rewrites each into five platform-native formats using the voice guide and winning-hook library in content-engine/.

Per interview the agent extracts four things. Claims: what you believe that others don’t. Stories: specific customer moments. Proof: numbers you can cite. Hooks: scored against past winners in the repo.

System 3

The Outbound Signal Engine

Good outbound doesn’t start with a spreadsheet of names; it starts with timing. Agents watch for trigger events (funding, hiring, bad reviews), match them to your ICP, draft a hyper-specific message and queue it for human approval.

Timing matters more than most teams admit. Harvard Business Review studied how fast 2,241 US firms responded to a web lead. Firms that made contact within an hour were about seven times likelier to qualify the lead than those that waited even an hour more. The average response time was 42 hours, and 23% never responded at all.

System 4

Creative testing as a loop

Agents generate structured variations from the core offer (for example 20 hooks by 10 angles), ship them, read the metrics and feed the winners back into creative-testing/. Each cycle starts smarter than the last. It is not a treadmill; it is a learning system.

The reason to invest here is in the data. Nielsen attributes 47% of ad-driven sales lift to creative, ahead of reach at 22%, brand at 15%, targeting at 9%, recency at 5% and context at 2%. Creative is the biggest lever and the one most teams test least rigorously.

System 5

Get cited by AI search

Buyers increasingly ask an answer engine instead of scanning ten blue links. The job shifts from ranking pages to being the source the model quotes: structured, factual, citable content.

The surfaces are already large. Google AI Overviews reports over 2 billion monthly users, ChatGPT over 800 million weekly, the Gemini app 450 million monthly and Google AI Mode over 100 million monthly in the US and India. Gartner predicts traditional search-engine volume falls 25% by 2026. The GEO study presented at KDD 2024 found that adding statistics, quotations and citations lifted visibility in generative answers by up to 40%.

System 6

The Growth Cockpit

The sum is greater than the parts. The cockpit rolls all five engines into one weekly executive view: market shifts, experiment results and the next tests to run. The memo answers three questions. What changed this week, which angle won on the metric that matters, and what to test next.

Most ideas lose, even at the best experimenters. Kohavi and colleagues put the share of A/B tests that improve the target metric at roughly 33% at Microsoft overall, around 15% at Bing, and 8% to 10% at Booking.com, Google Ads, Netflix and Airbnb Search. So the cockpit optimises for learning velocity: more valid tests per week, not more confident guesses.

Execution

The 30-day playbook

Don’t build all six at once. One month, four weeks, one working machine at the end.

1. Week 1, the audit. Study the website, ICP, founder content and funnel. Output: a market map of customer pain.
2. Week 2, the Growth Repo. Initialise the five core folders. Output: the first what-the-market-is-telling-us.md.
3. Week 3, the first machine. Pick one system and build it completely. Output: a live agent executing one job spec.
4. Week 4, the results. Document conversion and booked meetings. Output: a case study with business value.

The moat

Agents are a commodity. Judgment is the moat

McKinsey estimates agentic AI will power more than 60% of the additional value AI is expected to create in marketing and sales. The agents themselves will be available to everyone. What will not be evenly distributed is the judgment about what to point them at.

“The agents are going to be a commodity at some point. Your judgment about what to point them to is the machine your company cannot run without.”

Quoted in The Growth OS Blueprint

The best marketing engineers combine marketer, data analyst and engineer. They know what should exist in the first place.

Where to start

Build one engine

Start with the Growth Repo and one engine, ideally Customer Truth, since every other system reads from it. Ship it in 30 days, measure it, then add the next.

Every number in this guide is cited in its figure caption. Check the sources before you quote them.