Air France-KLM: Driving €1M Monthly Revenue with Smart Recommendations
Published:
The Challenge
Picture this: You land on Air France’s website to book a flight from Paris to New York. Dozens of options appear—different airlines, times, prices, connections. Which flight should appear first? Should we recommend the Basic fare or try to upsell you to Flex?
This is the problem I solve.
With millions of monthly visitors and average bookings worth €300-500, even tiny improvements have massive impact. A 0.5% lift in conversion doesn’t sound like much—until you realize that’s €1 million in monthly revenue.
My job? Design the recommendation systems that personalize these choices, then prove they work through rigorous A/B testing before we deploy them to production.
What I Do
I lead the experimentation and evaluation for two critical recommender systems:
1. Flight List Ranker
When you search for flights, we rank them based on what matters to YOU—not just price. We consider your search behavior, booking history, preferences, and hundreds of other signals to show you the flights you’re most likely to book.
2. Branded Fare Recommender
Should we show you the €150 Basic Economy fare or highlight the €250 Flex fare with free cancellation? We predict which customers are willing to pay for premium features and personalize the recommendation accordingly.
My Process: From Idea to Production
Before anything goes live, I put it through the wringer:
Step 1: Offline Testing I test new models on historical data to see if they would have performed better than what we currently have. No point running a real experiment if the idea is doomed to fail.
Step 2: Experiment Design If it looks promising, I design the A/B test. How many visitors do we need? How long should we run it? What could go wrong? I calculate all of this upfront so we’re not wasting time or traffic.
Step 3: Live Monitoring Once the experiment is running, I watch it like a hawk. Are conversion rates moving? Is anything breaking? Are certain customer segments reacting differently? I catch problems early and make decisions fast.
Step 4: Ship or Kill After the experiment, I analyze the results, present findings to stakeholders, and make the call: deploy to everyone or kill it. No gut feelings—just data.
The Impact
Conversion Rate: +0.6% improvement
- That’s a 20% relative lift over baseline
- Thousands more bookings every month
Average Revenue Per Visitor: +0.5% improvement
- Better conversion + smarter upselling
- More customers choosing premium fares
Business Value: €1M+ incremental monthly revenue
- €12M+ annualized impact
- From a single successful experiment
- Sustainable gains that compound over time
Why This Matters
Scale: Air France-KLM is one of Europe’s largest airline groups. We’re talking millions of users, massive revenue, and zero room for error.
Rigor: This isn’t “we built a model and hope it works.” Every recommendation system goes through scientific A/B testing. We measure real impact on real customers.
Ownership: I don’t just analyze data. I own the entire lifecycle—from designing the test to monitoring it live to making the final shipping decision.
Complexity: Aviation is tough. Customers search multiple times before booking, demand fluctuates wildly with seasons and events, and we’re competing with dozens of airlines and OTAs.
What I Learned
Business metrics beat technical metrics every time. Nobody cares if your ranking algorithm has a great “score.” They care if it makes money.
Fast iteration wins. The faster you can test ideas, the faster you learn. I’ve built processes that let us run experiments in weeks, not months.
Simple usually beats complex. The best solutions are often embarrassingly simple. Don’t over-engineer.
Trust but verify. Models that look great in testing can fail in production. Always validate with real experiments.
The Stack
- Data: Google BigQuery for petabyte-scale customer behavior analysis
- Experimentation: Custom A/B testing platform for millions of concurrent users
- Languages: Python, SQL
- ML: Scikit-learn, XGBoost, ranking algorithms, ensemble methods
- Collaboration: Working across data science, engineering, product, and commercial teams
Let’s Talk
I’m open to freelance projects and full-time opportunities where I can apply this experience:
- Building recommendation systems that drive revenue
- Designing and scaling A/B testing programs
- Applying causal inference to marketing problems
- Leading data science teams tackling complex challenges
Interested? Get in touch.