QuintoAndar's LLM Re-Ranker boosts housing search CTR by 5.3%
New LLM-based re-ranker improves click-through and scheduled visits by ~5% each...
QuintoAndar Group, the largest housing marketplace in Latin America, has published a research paper proposing an LLM-based re-ranker to improve conversational real estate search. The platform, which digitizes the traditionally paper-heavy rental and sales process, faces the challenge of matching users to homes from a vast catalog. As conversational AI becomes more common, users expect to express complex, multi-dimensional preferences through open dialog rather than rigid filters. The researchers fine-tuned a large language model to reorder retrieved candidates based on the nuanced context extracted from user conversations. To evaluate the approach, they constructed a proprietary dataset of 960,000 query-item pairs, combining synthetic and production queries, with annotations validated by both an LLM-as-a-judge framework and human reviewers.
Offline experiments on this dataset showed consistent ranking improvements, and a production A/B test confirmed real-world impact. The LLM re-ranker delivered a statistically significant +5.3% increase in click-through rate and a +4.8% increase in scheduled property visits. These results demonstrate that integrating conversational context into the ranking pipeline can meaningfully improve user engagement and conversion in housing search. The work highlights a practical path for other marketplaces to adopt LLM-based personalization without overhauling their entire search infrastructure.
- QuintoAndar built a 960,000 query-item evaluation dataset using LLM-as-a-judge with human validation.
- Production A/B test resulted in +5.3% higher click-through rate and +4.8% more scheduled visits.
- LLM re-ranker captures multi-dimensional conversational intent, replacing rigid filter-based search.
Why It Matters
Proves LLMs can directly improve key business metrics in large-scale real estate marketplaces.