Back to blog
AI

How AI property matching works (and why keyword search isn't enough)

The technical and practical guide to AI-powered property recommendations for real estate platforms.

January 20, 2026
8 min read

"I want something quiet, with a garden, near a good international school, and not too far from my office in New Cairo." Parsing this request with a keyword search returns anything with "garden" and "New Cairo" — potentially thousands of irrelevant results. AI property matching returns the 3–5 units that actually match.

The limitation of SQL keyword search

Traditional property search uses SQL WHERE clauses: SELECT * FROM units WHERE area='New Cairo' AND bedrooms=3 AND price < 4000000. This works for explicit filters but fails for implicit requirements like "quiet", "good school district", or "investment potential".

Vector embeddings for property matching

Zold uses Google's text-embedding-005 model to encode every property listing into a 768-dimension vector. This vector captures the semantic meaning of the unit's features — not just keywords. A unit described as "south-facing corner apartment with panoramic golf course view, 3 minutes from Gate Academy" will have its vector positioned near units with similar characteristics, even if the exact words don't match a query.

Hybrid retrieval architecture

The most accurate results come from a two-stage approach: first, SQL hard filters eliminate impossible matches (unit is Sold, price is wildly outside budget, area is completely wrong); then pgvector similarity search ranks the remaining candidates by semantic similarity to the buyer's requirements.

Re-ranking with buyer context

Raw similarity scores don't account for buyer-specific priorities. Zold's property matching agent applies a learned re-ranking: if the buyer mentioned "for my kids' schooling", school proximity is weighted more heavily; if they said "investment", rental yield data and capital appreciation history adjust the final ranking.

Acceptance rate as a success metric

The property matching agent's accuracy is measured not by "did it return results" but by "did the agent accept the recommendations". Acceptance rate of 91% means agents see the AI's suggestions as genuinely useful — not noise to scroll past.

Ready to transform your real estate operations?

Try Zold free for 14 days. No credit card required.