TL;DR

Running incrementality experiments in 2025 is like navigating international airspace in a hot air balloon—technically possible, but painfully slow, weather-dependent, and entirely out of touch with how the world actually moves. In this post, I break down how INCRMNTAL evolved past experiments into a real-time attribution engine—one that actually learns. I’ll walk through our methodology, best practices from working with hundreds of advertisers, the signs that signal optimization is already happening, and a peek at where this is all going (hint: platforms using causal data to improve themselves).

Preface

As INCRMNTAL, We were lucky. Not in the startup-lottery kind of way - but in the “right tech, right time, right problem” kind of way. Advertisers had been trying to fix incremental marketing measurement for years. Everyone knew there was waste. Everyone had a hunch that performance reports were mostly storytelling. But until recently, the only way to figure out what worked was to run experiments: geo lifts, blackout tests, A/Bs that killed your biggest channels.
I’ve said it before: “Using incrementality experiments in 2025 is like mailing floppy disks to stream Netflix - anachronistic, inefficient, and missing the point of progress.” 
So when we started INCRMNTAL, we made a very intentional decision. We weren’t going to be another fake incrementality service, masked in a “platform” offering incrementality “tests”. We were going to build something smarter - something that would allow advertisers to actually use incrementality in their day-to-day decision-making. Not just to measure the past, but to optimize the present.

From Incrementality Testing to Always-On Incrementality

Marketers have always known the truth: not all conversions are equal. Some are driven by ads. Some happen in spite of them. The hard part has always been figuring out which is which - without breaking your campaign or relying on creepy tracking.

INCRMNTAL was built to answer that. Our core is a reinforcement learning (RL) system that continuously estimates and updates the true impact of your marketing efforts. Not hypothetically. Not after a six-week blackout. In real-time. This kind of technology represents a leap forward in AI in data analysis for marketing.

How it works:

Step 1: Interrupted Time Series (ITS)
We start by modeling what would have happened without a given marketing event. That’s the counterfactual. We compare it to reality to determine lift - without needing control groups or user-level data.

Step 2: Incrementality Vectors & RL
Each marketing entity (channel, campaign, geo, etc.) gets a daily contribution vector. If predictions and reality don’t match, the model self-corrects. It learns. (Yes, real learning, not just a buzzword.)

Step 3: Attribution Functions
We then map changes to outcomes, giving marketers both total and marginal attribution curves. That’s how we help you shift budget before performance tanks - not after.

This approach gives you a live, always-on view of true incremental contribution - across everything you’re doing. No tracking. No waiting. No tests.

Best Practices in Incrementality (for the 2025-era marketer)

So what do the smartest advertisers do with this new power?
Focus on Lift, Not Hype: Don’t fall for big conversion numbers. Ask: which conversions were caused by the ad? If the answer is “most of them were going to happen anyway,” congratulations - you’ve just identified cannibalization, which directly affects your incremental revenue.

Zoom In and Out: We don’t just look at channels. We look at campaigns, regions, days of week, budget shifts. Attribution isn’t one-size-fits-all. It's layered, messy, and dynamic - and we embrace that. Using a powerful marketing data analyst tool helps marketers visualize and act on these nuanced layers efficiently.

Make It Operational: Top clients have our models embedded in dashboards. They’re adjusting budgets daily based on marginal ROI. It’s not an annual analytics report - it’s a living system.

Account for the World Around You: Trends, holidays, outages - legacy models often miss them. Ours doesn’t. That’s why marketers trust what it says even when the data gets weird.

Early Signs of Advertising Signal Optimization

This is where it gets fun.
Advertisers are no longer just analyzing incrementality - they’re optimizing based on it. Some of the things we’re seeing:

Dynamic Attribution Windows: A one-size window doesn’t fit all. Clients are shifting their attribution windows based on when value actually shows up—3 days for social, 10+ for influencer, etc.
We have been working closely with the largest ad platforms in the world, learning how we should teach our customers to use the platforms to account for incremental value. 

Smarter Targets: ROAS and CPA goals are being recalibrated based on incremental value, not raw numbers. That $20 CPA might look great - until you find out it only drove 5% lift.

Internal Integration: INCRMNTAL’s outputs are showing up in custom dashboards, planning tools, even Slack alerts. It’s not a “report” anymore - it’s part of the operating system.

What’s Next? We’re working on an AI agent inside our dashboard. It won’t just show you the data - it’ll tell you what to do. Think of it as a junior performance analyst that never sleeps, doesn’t take credit, and actually understands statistics. :-) 

The Future: Closing the Loop

We’re not stopping at insights.
The future we’re building is one where INCRMNTAL can send aggregated, privacy-safe daily postbacks to platforms - giving them causal signals on what’s actually working. Not correlation. Not attribution games. Real lift.

We call this Reinforcement Attribution: a closed feedback loop where platforms get the ground truth from advertisers, and adjust delivery accordingly. You get better outcomes. They get better signals. Nobody loses (except maybe the people still relying on last-click attribution in 2025).
In a privacy-first, ID-less, AI-everywhere world—this is the only kind of attribution that makes sense. Clean. Causal. Continuous.

Summary: Key Takeaways

• Experiments are out. Reinforcement learning and causal modeling are in.
• ITS modeling helps us estimate what would have happened without an ad.
• The system learns over time - updating its beliefs based on real-world feedback.
• Incrementality isn’t a metric. It’s a method. And it should be part of daily ops.
• Clients are already optimizing based on INCRMNTAL output: dynamic attribution windows, smarter ROAS targets, and integrated decision loops.
• An AI agent is coming soon to guide next steps directly in the dashboard.
• The future: closed-loop postbacks to platforms. Real optimization. Reinforced.