We Published Peer-Reviewed AI Research (And It Powers Every Ad Decision We Make)

Most advertising platforms are built on guesswork and correlation. AppCapy is built on peer-reviewed science.

Today, I’m sharing something I’ve been working on for over two years: published, peer-reviewed research demonstrating how to reduce advertising waste by 28% using neural contextual bandits and causal inference.

This isn’t theoretical academic work collecting dust. It’s the algorithm making 12,000+ decisions per second behind AppCapy, validated on over 8.2 billion real-world impressions.

Here’s the story of how we went from research paper to production system—and why it matters for anyone spending money on digital advertising.

📄 Published Research Highlights

• 28% reduction in cost-per-incremental-conversion
• 74% incremental lift validated vs. control group
• Sub-50ms latency processing 12,000+ queries per second
• 8.2 billion impressions tested in production
• Peer-reviewed and published on ResearchGate

Read the Full Research Paper →

The $100 Billion Problem

Digital advertising is a $600+ billion global industry, yet fundamental inefficiencies persist everywhere. Despite three decades of “advancements,” mistargeted ads remain the norm:

• Vegetarians seeing meat product ads
• Licensed drivers getting driving school promotions
• Non-parents bombarded with baby formula

These aren’t just amusing anecdotes—they represent over $100 billion in wasted ad spend annually.

But the real problem runs deeper than poor targeting. Most advertising platforms optimize for the wrong thing entirely.

Correlation vs. Causation: The Critical Difference

Here’s what most ad platforms do: they show ads to people likely to click or convert, based on historical data.

Sounds reasonable, right? It’s not.

This approach conflates two completely different groups:

1. “Sure things” – People who would have converted anyway, without seeing your ad
2. “Persuadables” – People who only convert because they saw your ad

Standard advertising platforms can’t tell the difference. They waste your budget showing ads to sure things (who don’t need convincing) while underserving persuadables (who actually need your message).

That’s the difference between correlation and causation. And it’s costing advertisers billions.

💡 Real-World Example

Imagine you sell premium coffee makers. Traditional targeting shows ads to people who frequently search for “coffee makers”—they’re highly likely to click.

But here’s the problem: many of those people are already planning to buy. They’ll purchase whether they see your ad or not. Your ad spend didn’t cause the sale—it just happened to be present.

Meanwhile, the person who loves coffee but hasn’t considered upgrading their machine—the one who could be persuaded by the right message—never sees your ad because they don’t fit the “high-click” profile.

You’re paying for correlation, not causation.

The Research: A Better Approach

I spent two years teaching myself machine learning specifically to solve this problem. The result is a 52-page peer-reviewed research paper that introduces a fundamentally different approach to advertising optimization.

Three Core Innovations

1. Multi-Armed Bandits for Real-Time Learning

Traditional advertising uses static audience segments set up once and rarely updated. Our approach uses contextual bandits—algorithms that learn and adapt with every single impression.

Think of it like this: Imagine you’re trying to pick the best restaurant, but you’re new to town. You could:

• Pure exploitation: Go to the first restaurant that looks good, then keep going there forever (even if better options exist)
• Pure exploration: Try random restaurants indefinitely (wasting money on bad meals to gather data)
• Multi-armed bandits: Start by exploring several options, then increasingly favor winners while occasionally checking if circumstances have changed

Our algorithms do the third option—automatically balancing learning about new ads and audiences (exploration) with showing ads we know perform well (exploitation).

The result: Your ad targeting gets smarter with every impression, in real-time, without manual intervention.

2. Neural Networks for Complex Pattern Recognition

User behavior isn’t simple. Whether someone engages with an ad depends on:

• What device they’re using
• What time of day it is
• What content they’re viewing
• Their browsing history
• Hundreds of other subtle signals

Traditional rule-based systems can’t capture these complex interactions. Neural networks can.

Our research demonstrates how to combine deep neural networks with bandit algorithms, using neural contextual bandits that predict ad performance based on rich user context while maintaining real-time learning capabilities.

3. Causal Inference for True Impact Measurement

This is the game-changer: we optimize for incremental conversions, not total conversions.

Our system maintains a control group and uses uplift modeling to distinguish:

• Users who convert because of the ad (high value—show them ads!)
• Users who convert despite the ad (zero incremental value—save your money)
• Users who will never convert (negative value—definitely skip them)

This approach, called causal inference, ensures we’re optimizing for actual influence on behavior, not just correlation.

🎯 The Causal Difference in Numbers

In our research validation:

• Standard optimization: 0.054% conversion rate, but only 0.019% incremental (35% of conversions were actually caused by ads)
• Causal optimization: 0.051% conversion rate, but 0.025% incremental (49% of conversions were caused by ads)

Result: 31.6% more incremental conversions and 27.8% lower cost per actual customer acquired.

Slightly fewer total conversions, but dramatically more caused conversions. That’s money well spent vs. wasted.

From Paper to Production: Building AppCapy

Publishing research is one thing. Deploying it at scale is entirely different.

Real-time advertising requires making decisions in under 100 milliseconds. When a user loads a webpage, our system has to:

1. Capture 512 dimensions of user context
2. Retrieve eligible advertisements (potentially thousands)
3. Run neural network inference
4. Apply causal uplift estimates
5. Select the optimal ad
6. Return the result

All in less time than it takes you to blink.

Engineering Challenges We Solved

Action Space Reduction

With thousands of potential ads, evaluating every option is too slow. We developed hierarchical contextual zooming—clustering similar ads and evaluating clusters first, then only running full inference on the most promising candidates.

Result: 18× speedup with less than 4% accuracy loss.

Distributed Architecture

Built on Next.js and Supabase, our system distributes computation across:

• Edge layer: Lightweight feature extraction (sub-10ms)
• Prediction layer: Cached predictions and real-time inference (sub-40ms)
• Training layer: Continuous model updates on historical data (offline)

Model Optimization

• Quantization: Int8 precision reduces inference time by 3× with <0.5% accuracy loss
• Knowledge distillation: Smaller “student” models mimic larger “teacher” models at 8× speed
• Caching: Pre-compute predictions for common contexts (68% cache hit rate)

Production Performance

After months of optimization, AppCapy now achieves:

• p50 latency: 22ms
• p95 latency: 47ms
• p99 latency: 78ms
• Throughput: 12,000+ requests per second per instance
• Daily impressions: 200 million+
• Accuracy: Within 2% of theoretical maximum

This isn’t a research prototype. It’s production infrastructure serving real ads, to real users, generating real results.

Real-World Validation: 8.2 Billion Impressions

Theory is one thing. Reality is another. Here’s how the research performed when tested against actual production traffic.

The Experiment

Over 90 days, across 52 publisher properties:

• 8.2 billion total impressions
• 340 million unique users
• 1,240 active advertiser campaigns
• 18,500 unique ad creatives

We compared AppCapy’s neural contextual bandits with causal inference against four baseline approaches, including the previous production ML system.

The Results

Method	CTR	CVR	RPM	Relative Performance
Random Selection	0.82%	0.019%	$2.14	-63.2%
Linear Bandits	1.48%	0.035%	$3.87	-33.4%
Neural (No Causal)	1.89%	0.046%	$5.12	-11.9%
Previous Production ML	2.03%	0.048%	$5.45	-6.2%
AppCapy (Full System)	2.24%	0.054%	$5.81	Baseline

Key findings:

• 10.3% improvement in click-through rate vs. previous system
• 12.5% improvement in conversion rate
• 6.6% improvement in revenue per thousand impressions
• All improvements statistically significant (p < 0.001)

The Causal Validation

But the most important metric isn’t total conversions—it’s incremental conversions.

By maintaining a 5% control group shown public service announcements instead of commercial ads, we measured true causal impact:

• Treatment group (shown ads): 0.054% conversion rate
• Control group (no ads): 0.031% conversion rate
• Incremental lift: +0.023 percentage points (+74% relative lift)

This means 43% of conversions were genuinely caused by advertising (0.023 / 0.054 = 43%). The other 57% would have happened anyway.

And here’s the critical part: when we optimized explicitly for incremental conversions rather than total conversions, we achieved:

• 31.6% more incremental conversions
• 27.8% lower cost per incremental conversion

Yes, total conversions dropped slightly (5.6%), but caused conversions increased dramatically. We spent less money to achieve more actual business impact.

⚠️ What This Means for Your Ad Spend

If you’re using traditional ad platforms optimizing for clicks or total conversions, you’re likely wasting 50-60% of your budget on users who would have converted anyway.

That’s not a criticism of your strategy—it’s a limitation of the algorithms those platforms use.

The solution is causal optimization. AppCapy is the first advertising platform built on this approach from the ground up.

Why This Matters for Publishers and Advertisers

For Publishers

Higher revenue without degrading user experience.

Our research demonstrates that better ad relevance improves both publisher revenue and user satisfaction. By showing ads that