Marketing as Systems Design

How to think about growth as an engineering problem: inputs, outputs, feedback loops, and compounding returns.

The first quarter I ran affiliate growth at AliExpress US, my team shipped eleven campaigns. Nine of them worked. Revenue went up. Everyone was happy. The next quarter we had to do it again from zero, because nothing we built in Q1 made Q2 any easier.

That is the failure mode of most marketing organizations, and it took me an embarrassingly long time to name it. We were producing outputs when we should have been producing capacity. A campaign is an output. A recruitment engine that brings in 400 qualified partners a month without anyone opening a spreadsheet is capacity. The difference compounds, and the compounding is the entire game.

When I finally stopped thinking in campaigns and started thinking in systems, the affiliate program went from a few hundred active partners to over 10,000 across 40+ countries, and contributed a 30% revenue lift. Not because the campaigns got smarter. Because the system did work while we slept.

A marketing system has four parts, and most teams only manage one

Borrow the vocabulary from engineering and a lot of confusion disappears. Every growth system has inputs, a transformation, outputs, and feedback. Marketing teams obsess over the transformation, the creative, the copy, the campaign mechanics, and treat the other three as somebody else's job.

Here is how I break down any growth program before I touch a single tactic:

• Inputs: what raw material enters the system every week, and can I increase the supply without increasing headcount linearly? For an affiliate program, that is qualified publisher leads. For creator commerce, it is applications. If the input is capped, nothing downstream matters.
• Transformation: the conversion step that turns an input into an active revenue-producing unit. Onboarding, activation, first-payout. This is where most programs quietly leak 60 to 80 percent of what they recruit.
• Outputs: GMV, orders, incremental revenue. The number everyone reports and nobody can move directly.
• Feedback: what the system learns from its own outputs and feeds back into inputs or transformation. This is the only part that produces compounding, and it is the part almost nobody builds.

Ask a growth team what their weekly input volume is and most cannot answer. They can tell you last month's revenue to the dollar. That asymmetry tells you exactly where the system is undermanaged.

Feedback loops are the only thing that compounds

Linear marketing is spend in, revenue out, at a roughly fixed ratio. You buy traffic, you get conversions, you stop buying, it stops. Nothing you did last month makes this month cheaper. Paid acquisition is the purest form of this, which is why CAC in most paid channels has gone one direction for a decade.

A loop is different. In the AliExpress creator program, every creator we onboarded produced content. That content produced performance data. The data fed a recommendation layer that got better at matching products to creators. Better matching meant higher earnings per creator, which meant creators told other creators, which fed recruitment. Four connected steps, each one feeding the next, and each cycle made the next cycle cheaper.

We went from onboarding roughly 400 creators a month to over 2,000 active creators and 40% GMV growth, and the marginal cost of the last thousand was a fraction of the first hundred. That is not a campaign result. That is a loop result, and loops are the only asset in marketing that appreciates.

If your growth program produces no artifact that makes next quarter easier than this quarter, you do not have a system. You have a job.

Instrument the transformation, not just the output

The most expensive mistake I see is teams that measure only the endpoint. GMV is up 12%. Great. Why? Nobody knows, so nobody can repeat it deliberately.

When I rebuilt measurement for the AliExpress affiliate engine, I made every stage of the funnel a first-class metric with its own owner. Leads sourced. Leads contacted. Applications received. Approved. Onboarded. First link placed. First conversion. First payout. Second month retained. Nine stages, nine numbers, reviewed weekly.

The result was that when GMV moved, we knew which stage moved. And more usefully, when GMV did not move, we could see the blockage before it showed up in the revenue line, usually four to six weeks earlier. A program that sources 2,000 leads and activates 40 of them does not have a revenue problem. It has a transformation problem, and no amount of extra lead volume fixes it. You just pour more water into a leaking bucket at greater expense.

The rule I use now: every stage where a unit can die needs a number and a name attached to it. If a stage has no owner, it has no improvement rate.

Design for the second-order effects, because they arrive anyway

Systems thinking is mostly about taking the second step seriously. You change an incentive, and the change propagates in ways the spreadsheet did not model.

We once raised commission rates on a category to drive volume. Volume went up. It also pulled our best content publishers off higher-margin categories where they had been performing well, because we had just made the alternative more attractive. Net contribution was flat. The model said the change was positive because it only modeled the category in isolation. The system had a substitution effect the model did not contain.

Now, before any structural change to a growth program, I ask four questions:

• What behavior does this make more profitable for the partner, and what does it make less profitable?
• What did the resource being reallocated used to do, and who notices when it stops?
• If this works and doubles, does the system hold, or does something break at 2x volume?
• What does this look like when someone games it? Because if a lever exists, somebody will pull it harder than you intended.

That last one matters more in affiliate marketing and performance channels than anywhere else. Any payout rule is an optimization target for thousands of independent operators. Design a rule that rewards last-click and you will get last-click harvesting rather than demand creation. The partners are not being dishonest. They are responding correctly to the system you built.

What this actually changes about how you work

The practical shift is where you spend your best hours. A campaign-oriented marketer spends their best hours on execution. A systems-oriented one spends them on the input pipeline and the feedback mechanism, and lets execution be handled by whoever or whatever is cheapest that clears the quality bar.

At Next2Market, running DTC for consumer electronics brands including Samsung, TCL, Levoit and Anker, the 350% GMV growth did not come from better ads. It came from building a repeatable launch sequence that any new SKU could run through, plus a conversion measurement layer that told us within days which step of the sequence was underperforming. The 18% CVR improvement was a byproduct of finally being able to see which step was broken. We were not smarter. We were better instrumented.

I am also skeptical of the version of this idea that turns into a tooling purchase. A dashboard is not a system. Most teams buy visibility into outputs they already knew and call it transformation. The system is the loop, the owned stages, and the discipline of asking what happens on the second step. The tooling is downstream of that, and it is usually much less than people expect.

The test I apply to any growth program: if the whole team took a month off, what would keep producing? In a campaign shop the answer is nothing. In a well-built system the answer is most of it, slightly degraded. That gap is the only honest measure of how much system you have actually built.