The Expedia analysis started as a raw dataset: 100,000 hotel search records with outcome signals (clicks, bookings, prices, ratings). The challenge wasn't the data—it was turning raw numbers into insights and understanding what the data was really saying.

This is where Claude Code made the difference. Instead of spending days writing boilerplate data cleaning code, I focused on asking better questions and letting Claude Code handle the mechanical work. Here's how.

Step 1: Load and explore. I imported the dataset into Claude Code and asked it to profile the data—shape, missing values, distributions. Within seconds, I had clarity on what I was working with.

Step 2: Hypothesis generation. Instead of running blind statistical tests, I used Claude Code to help structure hypotheses: "What if price isn't the conversion lever? What if position matters more?" Claude Code then wrote the queries to test each hypothesis.

Step 3: Segmented analysis. Breaking down booking rates by segment (Budget, Mid, Luxury) was straightforward in pandas, but Claude Code automated the comparison tables and cross-segment calculations, catching patterns I might have missed manually.

Step 4: Visualization. The deck you see was built with Claude Code writing Chart.js, HTML, and CSS. Instead of struggling with charting libraries, I described what I wanted to show—the trust gap, position elasticity, segment differences—and Claude Code structured it visually.

Step 5: Documentation. The case study writeup was generated with Claude Code helping structure the narrative, ensuring each finding was backed by data and clearly explained.

Data cleaning and transformation: Claude Code wrote efficient pandas pipelines without me having to debug data types or handle edge cases manually.

Exploratory analysis: Instead of writing multiple scripts to test different hypotheses, Claude Code generated them on-demand as I asked questions.

Visualization: Building interactive charts and presentations is usually time-consuming. Claude Code wrote the HTML/CSS/JS deck in a fraction of the time it would take manually.

Documentation and storytelling: Claude Code helped articulate findings clearly, ensuring the technical analysis was accessible to non-technical stakeholders.

1. Ask better questions, not more questions. With Claude Code handling code generation, I could focus on asking more targeted questions about the data.

2. Outcome-level thinking. Claude Code doesn't care about convention. When I asked "which levers actually move conversion?", it tested outcomes directly instead of proxies.

3. Iteration is fast. Changing a hypothesis or exploring a new angle took minutes instead of hours. This enabled deeper analysis in less time.

4. Documentation is first-class. Because Claude Code generates narratives alongside code, the analysis is self-documenting. Anyone can understand the journey from data to recommendation.

The real win wasn't speed—it was depth. With less time spent on mechanical coding, I spent more time thinking deeply about marketplace dynamics, questioning assumptions, and stress-testing conclusions. The analysis is stronger because Claude Code handled the grunt work.

If you're building data-driven insights, start with Claude Code. Define your outcome metric, articulate your hypothesis, and let Claude Code do the mechanical work. You'll find yourself asking better questions faster, and your findings will be stronger.

Building the Expedia Analysis with Claude Code

This case study explores a 100,000-sample analysis of Expedia's hotel search and booking marketplace, uncovering why the luxury segment underperforms by 27% compared to budget hotels—and what data reveals about the real drivers of conversion.

Using outcome-level data (clicks, bookings, prices), I applied causal measurement techniques to isolate which levers actually move the needle on conversions. The findings challenge conventional optimization assumptions and point to a clear strategic priority: quality trust, not price.

Expedia's overall booking rate sits at 2.8% of impressions, but this varies significantly by hotel segment: Budget at 3.04%, Mid at 2.89%, and Luxury at 2.20%. Luxury is 27% lower than Budget.

This gap is material—on a platform handling millions of searches, a 0.84 percentage point difference in luxury bookings represents substantial lost revenue. The question is: why?

Most marketplace analyses rely on statistical correlation or user surveys to infer causation. This project took a different approach.

Our dataset includes direct outcome signals—booking_bool (did the user book?) and click_bool (did they click the hotel?). This allows us to measure actual conversion behavior, not proxies.

The approach included: funnel analysis, ranking impact quantification, price elasticity testing, quality analysis comparing promised star ratings to actual review scores, and experience effects measuring returning-visitor lift vs. new visitors.

Finding 1: Price is NOT the Conversion Lever

Hypothesis: Underpriced hotels convert better because they represent value. Result: No. Booking rate differences by pricing tier are negligible (0.83 percentage points). The correlation is statistically irrelevant.

Stop optimizing price. The marketplace is already at equilibrium.

Finding 2: Position is the Strongest Lever

Position 1 books at 13.37%, while Position 10 books at 2.56%—a 5.22x difference. More surprising: booking elasticity is stronger than click elasticity. This means position matters more for post-click conversion than for visibility itself.

The ranking algorithm should optimize for conversion signals, not just relevance. Hotels shown higher have more credibility and trust.

Finding 3: Ranking Quality is Equal Across Segments

Click rates are flat across segments (4.6% Budget, 4.7% Mid, 4.1% Luxury). The problem is post-click abandonment, not search relevance.

Finding 4: The Quality Trust Gap Explains It All

Budget travelers expect to be pleasantly surprised (they exceed expectations). Luxury travelers expect perfection and are disappointed when reality doesn't match the promise.

Budget: Promise 2.40★, Delivery 3.39 = +41% gap. Luxury: Promise 4.30★, Delivery 4.20 = −2.3% gap. This cognitive dissonance drives abandonment in the luxury segment.

Returning visitors book at 3.66% vs. 2.74% for new visitors—a 1.34x lift. But click rates are nearly identical. Trust, not visibility, is the differentiator.

Luxury buyers have elevated expectations. When listed hotels deliver below promise, they abandon. This is a credibility problem, not an execution problem. The star rating system doesn't distinguish between truly exceptional and merely adequate.

Priority 1: Audit all 4.3★+ luxury listings for promise-delivery gaps. Delist properties where the gap exceeds 0.2 stars. Expected impact: +0.6pp lift in luxury booking rate.

Priority 2: Retrain the ranking algorithm to surface high-confidence hotels first. Expected impact: +0.3–0.5pp lift.

Priority 3: Highlight verified buyer reviews and trust signals. Expected impact: +0.2–0.3pp lift.

Priority 4: Remove effort from price-matching and dynamic pricing. The data shows no ROI. Redeploy resources to higher-leverage levers.

1. Causation matters. Correlation analysis would have suggested price optimization was worth pursuing. Outcome-level analysis revealed the difference is noise.

2. Segment differences are real. Budget travelers are optimism-biased. Luxury travelers are skepticism-biased. Marketing and UX should reflect this.

3. Position is a trust signal, not just a relevance signal. The algorithm should weight quality and reliability equally with search relevance.

4. The data already tells the story. With the right hypotheses and outcome metrics, you don't need surveys or guesses. The data is explicit.

This analysis demonstrates how outcome-level data and hypothesis-driven measurement can cut through marketplace complexity and identify true conversion drivers. The path forward isn't price optimization—it's quality alignment and trust signals.

Writing

Building the Expedia Analysis with Claude Code

Expedia Marketplace Analysis: Why Luxury Underperforms