Build, don't buy: how we architected AIMG's marketing intelligence stack

Most marketing teams buy attribution. We built it.

I joined Ai Media Group as a web developer in 2008. By 2013, I was running Advertising Technologies as Director — and by then the marketing measurement market was already crowded with tools that promised to unify cross-channel data. Most of them shipped versions of the same problem. They assumed a model of measurement that fit every client equally, which meant they fit no one well once spend got serious. We had clients managing eight-figure annual media budgets across a dozen ad networks, and the attribution they were getting back from third-party tools wouldn’t survive a CFO’s scrutiny.

So we built our own. Not because it was ideologically pure — because it was the only path that produced numbers our account teams could actually defend.

A decade later, that decision compounded into a three-layer architecture that I now lead as CTIO. Each layer exists for a specific reason, and the whole stack only works because the layers are coherent with each other.

The data layer: AiTRK

AiTRK is a patented (US Patent 10,091,321) tracking pixel that captures every meaningful user interaction — clicks, impressions, view-through conversions, cross-domain journeys, cross-device recognition. It’s lightweight (under 5KB gzipped), privacy-first, first-party, and Google Ads–certified.

The reason AiTRK matters isn’t the feature list. It’s that we control the data layer end-to-end. If you license your tracking pixel from someone else, you don’t control your own measurement. You’re consuming a model of reality that someone else built — and when their model changes, your numbers change with it. Owning the pixel is the moat. Everything downstream — attribution, AI, optimization — depends on whether the data layer is yours.

The attribution engine: Atrilyx

Atrilyx is the multi-touch attribution platform that sits on top of AiTRK data. It unifies online and offline media into a single source of truth for marketing ROI — direct, indirect, and assisted conversions, all attributed via the same model with consistent lookback windows.

The substantive thing Atrilyx solves is reconciliation. Every ad network reports its own performance using its own attribution logic. Networks have a structural incentive to over-credit their own touchpoints. Without an independent measurement layer, marketing leaders are making budget decisions on a ledger where every column was scored by a different referee. Atrilyx is the referee that doesn’t work for any of the networks.

Under this stack, our clients have seen client engagement lift 95% against legacy attribution and our data governance posture improve 40% through GDPR/CCPA-aligned controls — not because the platform is magical, but because the operating discipline that ties the layers together produces decisions you can stand behind.

The agent: AI on top of clean attributed data

The Atrilyx Agent extends the stack with AI. It’s an AI marketing assistant that operates on attributed data — surfacing anomalies, recommending optimizations, and answering the kinds of questions our account teams used to spend Tuesdays building decks for.

The reason this works is that the agent runs on a clean, layered foundation. Most “AI for marketing” products today layer an LLM on top of a fragmented data stack and produce confidently wrong answers faster than humans can verify them. We didn’t have that problem because the agent reads from data that’s already been collected by our pixel, attributed by our engine, and governed by our compliance controls. The AI doesn’t have to reconcile reality — reality is already reconciled by the time it gets there.

What I’d tell anyone considering the same path

Three things, briefly:

1. The pixel is the moat. If you don’t control your data collection, you don’t control anything downstream. License the data layer and you’ve licensed your own truth.

2. AI agents only work on top of clean attributed data. Building an AI surface before you’ve solved the attribution problem just compounds noise. The order matters: data → attribution → AI. Skipping a step doesn’t save time; it imports inconsistency that you’ll spend years untangling.

3. Operating discipline beats platform sophistication. The hardest part isn’t building the technology. It’s the team and the process that runs on top of it — the people who defend the numbers in QBRs, who handle attribution discrepancies without blaming the data, who turn down the easy answer when the right answer requires more work. Platforms can be copied. The operating discipline can’t.

More than a decade after the decision to build, I’m still convinced it was right. Not because the alternative was wrong for everyone — but because for what AIMG was trying to do at the scale we wanted to do it, the only stack that worked was the one we owned.

— Joel