Back to Case StudiesReal-Time Bidding Optimization for a Global Ad-Tech Leader
October 14, 2025 .
The client’s challenge was to optimize bidding strategies in real time, ensuring relevant ads reached the right users, minimising wasted spend
The client is a leading global player in the mobile advertising space based in the US, helping advertisers and publishers run offers across multiple platforms. With operations in over 25 countries and 100M+ users, the platform facilitates billions of ad auctions daily across various campaign types like video ads, engagement offers, rewarded ads, click ads, and app installs.
Each campaign competes for user attention, but capturing attention alone doesn’t guarantee conversions. Some ads require substantial user interaction, while others are simpler but lower value. Users may drop off mid-engagement, be disqualified, or attempt fraudulent interactions. The client’s challenge was to optimize bidding strategies in real time, ensuring relevant ads reached the right users at the right moment to maximize conversions and minimize wasted spend.
Their existing bidding algorithms were static and delayed in responding to changing user behaviors, resulting in overspending in some campaigns while underutilizing others. The absence of robust experimentation frameworks further limited their ability to fine-tune strategies.
APPROACH
Understanding KPIs & User Behavior
We began with a structured two-pronged approach.
1. Benchmarking Performance
- Used quick-turnaround dashboards (Looker Studio) to assess current bidding efficiency, conversion rates, and campaign ROI across geographies for different advertisers & ad-types.
- Identified key levers influencing user engagement, such as ad duration, reward size, and interaction complexity.
2. Exploring User Preferences
- Analyzed patterns showing how users respond to various bid-triggered ads.
- Found that ads with shorter interaction times but meaningful rewards drove higher completion rates.
- Observed that overly complex offers led to drop-offs and reduced conversion.
Assessing the Technical Landscape
The analysis led to deeper insights about:
- Data volume & velocity expected in real-time bidding.
- Latency requirements, with sub-200ms thresholds for bid decisions.
- Data refresh rates, sourcing from user sessions, geo-location signals, device types, etc.
- Algorithms & experimentation frameworks suited to high-throughput decisioning.
SOLUTION
Rapid Prototyping with Scalable Architecture
We set out with our eyes on speed to impact and setting ourselves up for continuous iteration. By assembling our Avengers (Engineers and Data scientists) together, we rapidly turned our hypothesis into working solutions while laying the groundwork for future experiments.
Feature Engineering Pipelines
We built our own data infrastructure (including a feature store) by connecting with data sources like BigQuery, AWS S3 etc. for real-time and batch feature extraction. Big-Data pipelines used PySpark on AWS EMR clusters, while others used python or SQL scripts. These Data pipelines were orchestrated with Airflow. This framework of pipelines + Feature Store evolved further serving multiple models downstream.
Modeling Strategy
While the pipelines were under development, Data Scientists created training datasets and launched experiments using AWS SageMaker.
- Initial experiments used multi-armed bandit algorithms to dynamically balance exploration and exploitation.
- Contextual bandits were applied to personalize bids based on user segments and interaction history.
- A cold-start framework was implemented for newly launched campaigns, using auditioning algorithms to ensure even less-experienced campaigns had optimized bid opportunities.
Real-Time Inference Infrastructure
The inference pipeline began on AWS API Gateway and Lambda, but as the model complexity increased, we migrated to EMR clusters with auto-scaling, ensuring uninterrupted service at scale.
Ads were served through an API endpoint where engineering teams made an API request at the time of rendering a bid, receiving near-instantaneous optimized bid values.
We implemented caching layers and query optimizations to keep API latency consistently below 200ms.
IMPACT
Over the course of one year
70% increase
in campaign conversion rates (YoY)
20% reduction
in bid waste through smarter allocation
15% improvement
in overall ad engagement metrics
- Scaled from 1 geography (US) serving ~1M bid requests per day to 21 countries serving 10M+ requests daily
- Enabled automated A/B testing, giving campaign managers continuous feedback loops without manual overhead
KEY LEARNING
Behind the Numbers: What Really Matters: 6 Hard-Earned Lessons
1. Real-Time Latency Is a Bigger Challenge Than Anticipated
- Even with scalable cloud services, ensuring sub-200ms response times requires continuous optimization.
- It’s not enough to track failures or latency spikes, you need dashboards that correlate bid outcomes with user behavior and campaign performance.
2. Data Quality & Consistency Drive Model Accuracy
- Inconsistent signals (missing user context, stale session data) can degrade bid predictions rapidly.
- Feature drift (when user behavior changes over time) can make previously accurate models obsolete if retraining isn’t automated.
- Contextual monitoring (which campaign, which region, which user segment) helps isolate root causes faster.
3. Exploration vs. Exploitation Is a Constant Tradeoff
- Multi-armed bandits and contextual models need fine-tuning between trying new strategies and leveraging known wins.
- Over-exploration can increase costs, while over-exploitation can prevent discovering better opportunities.
- Dynamic strategies adjusted based on campaign size, geography, or user volatility yield better results than static approaches.
- New campaigns without historical data (Cold Start) perform poorly unless audition frameworks are built for exploration.
4. Models need to be Localized, not one-size-fits-all
- Bidding strategies that drive clicks in one region may underperform in another due to cultural and behavioral differences.
- Device & network conditions shape outcomes. RTB models optimized for high-speed networks may misfire where latency or bandwidth is constrained.
- Localized tuning of bid logic, frequency capping, and creative selection consistently outperform generic models.
5. Model Interpretability & Explainability is important as it
builds Trust
- Campaign managers need visibility into why certain bids are being recommended. Even if algorithms outperform manual strategies, explainability leads to adoption and alignment with business goals.
6. “Test & Learn” Frameworks Unlock Innovation
- Without A/B testing pipelines, experimentation becomes ad hoc and slow, limiting the team’s ability to iterate & improve the system.
Your ads can do more
Let’s build your RTB strategy together over a call with our experts.
