How listings reach your feed
The full distribution algorithm, in plain language, with the exact weights that are live in the product. Nothing here is claimed to be a prediction of which property is "best" - scores rank preference fit and novelty, not investment quality.
The pipeline, stage by stage
Authorized supply only
Only properties that agencies/developers have published into the shared catalog enter the pool (latest records only, with a price). Your feed never invents supply.
Your criteria as hard filters
If your saved preferences name cities, property types, or a price ceiling, non-matching listings are excluded, not down-ranked: excluded.
Scoring
Every candidate gets a composite score from six components (weights below). Higher score = served earlier.
Diversity pass
Before results are shown, an equalizer checks each 6-card window; if it is all one type or one price band, a different candidate is moved up so the feed is not a monoculture.
Your profile learns
Your saves, shares, detail views, and skips update your taste profile (favorite +1.0, view +0.6, skip −0.4, dismiss −0.2, share +1.2). Cities and types you respond to are added; your price band follows your actual browsing average. You can see that profile on the Discover page.
The exact ranking weights
| Component | Weight | What it means |
|---|
Components sum to 1.0. Exploration randomness is capped (max ±2% of final score) and is re-seeded each page render, so order is stable within a session but fresh between sessions. At launch, the "quality" component is a fixed reserve pending a data-completeness measure.
Weights lab, try the algorithm on sample supply
The lab ranks a labeled sample of 12 fictional listings with the product's real scoring code
(app/js/feed-algorithm.js - the same file the feed uses). Choose a preset to see how preference changes re-rank supply,
and read each card's why shown reasons. No predictions are shown for these listings: market signal requires the paid engine (it contributes only when a forecast exists).
Rules the ranking follows
No paid boost inside discovery
Promotions exist as a separate agency product and are visually labeled; they do not silently add weight inside the matching score. [owner decision 2026-10-17]
Predictions are paid, and labeled
Forecast-based ranking only activates for listings with a real current forecast; forecasts come from the membership product and always show uncertainty bands, see the public Model Accuracy Report.
No "best investment" claims
The score ranks how well a listing fits your stated and learned preferences. It is not a valuation rank, a return forecast, or a recommendation to buy.
Sample lab data is illustrative. This surface itself is a sample-data preview; live listings arrive from the authorized supply pipeline.