How Unitrank works

A technical note on data sources, modelling, and known limits. Last updated May 2026.

Unitrank ranks every apartment by how much of your life it costs. Not just rent: commute, noise, grocery access, workspace quality, and every other daily activity. The apartment that costs you the fewest minutes per day ranks first.

Results are personalized to your work location, household, income, and lifestyle. Because we aggregate across all major platforms, the model compares an Airbnb in Palermo against a direct rental on Argenprop, a temporary listing on Zonaprop, and an owner-listed apartment on MercadoLibre on equal terms: same cost model, same commute calculation, same noise score.

1. Why this is harder than it looks

Apartment data is not a clean dataset. It is a hostile, continuously degrading environment.

Listings go stale. Prices change. Photos get replaced, sometimes with images from older units or staged shots that do not represent what you would actually walk into. The same unit appears across two or three platforms with different prices, different bedroom counts, and sometimes different coordinates. Categories are used loosely: “1 bedroom,” “studio,” “temporary,” “furnished,” and “dueño directo” mean different things on different sources, and sometimes different things within the same source. Commissions, building fees, utility assumptions, and guarantee costs are often hidden, softened, or only mentioned after first contact.

Even basic comparisons break down without correction. A listing that looks cheap can end up more expensive than a neighbour once building fees, utilities, amortised guarantee, and broker commission are added back in. A listing that appears similar to another on paper can represent a very different apartment, building, or living situation in reality.

That is why apartment search is not a database problem. It is a continuous process of ingestion, deduplication, normalisation, extraction of missing attributes, validation, and correction against a market that is noisy, strategic, and actively changing underneath the data.

2. Why preferences aren’t literal filters

Standard portals treat preferences as clean binary filters. Real renter intent is not that simple.

When someone states a “budget of $1,200,” they may mean maximum advertised rent, maximum true all-in cost, a comfortable target that flexes if the apartment is clearly better, or a number that moves as commute or furnishing improves. When someone says “close to work,” they may mean walking distance, a short subway ride, or simply “not exhausting by the end of the week.” “Furnished” ranges from fully move-in ready to “has a bed.” “Safe” can mean low crime, quieter streets, better upkeep, or simply a neighbourhood that feels calmer.

Two renters can give identical inputs and want very different outcomes. Two identical apartments can be the right choice for one household and the wrong choice for another. The reason is that preferences are compressed expressions of tradeoffs (time against money, convenience against price, quality against friction), and those tradeoffs are not captured by a literal label match.

Unitrank therefore treats many preferences as interpreted signals. The ranking evaluates tradeoffs against the specific household, stay length, income, and lifestyle, not just the words the user typed into a filter.

3. What we measure

Every apartment is evaluated across 21 life outcomes. Each is a daily or weekly activity your apartment makes easier or harder.

No single outcome dominates the ranking. The apartment that wins is the one where all 21 outcomes combined cost the least.

For each outcome, multiple strategies compete. An apartment with a pool may skip the gym commute for cardio. One without a desk factors in coworking costs. Cook at home, eat out, or order delivery: all three compete for every meal, and the cheapest wins.

For outcomes with multiple nearby options, trips are distributed across competing venues weighted by distance and quality. A closer, better-stocked supermarket captures more of your grocery trips than a distant corner store. Breakfast, lunch, and dinner are modeled separately with different venue preferences and household scaling.

4. What the model catches

Four decisions the ranking makes differently than a human browsing listings.

Noise within the same building

A street-facing unit and a courtyard-facing unit at the same address can differ by 15-20 dB. That is the difference between sleeping with earplugs and sleeping in silence. Both units list the same rent. No platform distinguishes between them. The ranking knows which side of the building faces the avenida.

The commute triangle

A renter working in Microcentro, with a gym in Belgrano and kids at school in Núñez has three commutes, not one. "Palermo is central" is true in general but wrong for this household. The ranking finds the apartment that minimizes the triangle of all three destinations. Choosing Palermo because everyone says Palermo costs 20-40 extra minutes per day.

The expensas blindside

Two apartments at $800/mo rent. One has $50/mo in building fees, the other $250. The $800+$250 apartment is more expensive than a $900 apartment with $80 in building fees. Renters filter by rent. The ranking sorts by all-in cost.

Furnished vs unfurnished: the real comparison

Most platforms label apartments "furnished" or "unfurnished" but never compare them on equal terms. The ranking does. It calculates the true cost of an unfurnished apartment including furniture purchase, delivery, assembly, and disposal at exit. For stays under 12 months, unfurnished apartments frequently cost more than furnished ones once these hidden costs are included. Each user sees the comparison based on their actual planned stay duration.

5. All-in cost calculation

Most platforms show base rent. We show what you actually pay each month.

6. Commute model

Walking and driving distances use real routing on OpenStreetMap road data. These are actual walking routes, not straight-line distances.

Distances are pre-computed for 428 categories and 465k+ points of interest: supermarkets, subte stations, parks, gyms, schools, coworking spaces, and more. For custom locations such as your work address or school, exact routes are calculated at query time.

7. Noise model

Every apartment gets a noise impact score from 0 (silent) to 100 (major exposure).

Outdoor noise: 5 sources

• Road traffic: CoRTN road noise model using OpenStreetMap road geometry with calibrated traffic volumes. Major roads contribute most.
• Aircraft: Point-source attenuation from Aeroparque. Primarily affects Palermo, Belgrano, and Núñez.
• Nightlife: Nightclubs and bars within 200m. Relevant for Palermo Soho, San Telmo, and entertainment districts.
• Bus stops: Colectivo stops on high-frequency routes along major avenidas.
• Combined: All 5 sources summed in the power domain via acoustic addition.

Indoor attenuation

Outdoor noise is reduced by the building envelope. Closed windows with AC attenuate 14 dB. Open windows without AC attenuate 10 dB. Building age adds 0-3 dB, double-pane windows add 5 dB, and soundproofing adds 3 dB. Open-window values from Locher et al. 2018, a study of 102 Swiss residences.

Impact scoring

Formula: (indoor_dB - 35) / 30 × 100, clamped 0-100. The 35 dB baseline sits between the WHO sleep guideline of 30 dB and the ASHRAE office standard of 35 dB. The 65 dB ceiling is the WHO "major public health concern" threshold.

8. Converting money to time

All costs are converted to minutes per day using a formula that accounts for both earning power and psychological weight of spending:

minutes_per_dollar = √(labor_rate × perception_rate)

9. Statistical validation

We validate our cost model against market prices using hedonic regression, a statistical method that isolates the price premium for each individual feature.

• Method: Semi-log regression with hexagonal grid fixed effects, controlling for location at block level (~460m resolution)
• Features: 60+ amenities tested simultaneously: pool, parking, washer, elevator, AC, workspace, dishwasher, and more
• R²: 0.51, explaining 51% of price variation. In line with published hedonic studies
• Observations: 6,046 listings with complete data

Where our model agrees with market pricing, we have confirmation. Where we disagree, the model finds value others miss.

Each arbitrage opportunity is personalized. The workspace advantage only appears when you specify remote work. The dog park advantage only appears when you indicate you have a dog. The model adapts to your life, not to average market preferences.

10. Data sources

New sources are added regularly. Each new platform increases coverage and strengthens cross-platform comparison.

11. Refresh cadence

Data freshness directly affects ranking quality. Stale listings waste your time.

• Search crawls: Continuous per source. Newest-first discovery plus rotating full sweeps. New listings typically surface within hours of publication.
• Availability verification: Calendar sync and detail re-fetch on staleness or search-miss signals. Empty availability removes listings from the ranking.
• Tier-weighted prioritisation: A-tier barrios (Palermo, Recoleta, Belgrano, Núñez, Vicente López, and peers) are refreshed ahead of B and C barrios.
• LLM enrichment: Continuous batch processing with a premium/economy tier split. New listings typically enriched within hours of discovery.
• Ranking rebuild: Incremental updates on every listing change. Full rebuild nightly.
• Trust scores: Refreshed on user feedback, listing age, and verification signals.
• Ranking computation: Personalised results in under 41 ms (p95), typically 32 ms (p50). Evaluates all 21 life outcomes across the full inventory, selects the optimal strategy for each outcome per unit, and exposes every decision with a complete cost breakdown. Computed per request, not pre-cached.

12. Known limits

The ranking is a model, not an oracle. Five areas where signal is weaker than the rest of the system and users should read results with that in mind.

• Furnishing detection is photo-based. Roughly 61% of units have detectable bed capacity from LLM photo analysis. The remaining 39% are treated as unknown, not as unfurnished. This is the best available signal when listing text is unreliable, but it misses units whose furniture is off-camera.
• Review and trust signals are uneven across sources. Airbnb has review counts, superhost badges, and guest-favourite flags. Argenprop, Zonaprop, and MercadoLibre do not. Trust scores on non-Airbnb sources fall back to listing age, agent volume, price-vs-market, and user feedback, a narrower set of inputs.
• Expensas estimation has a three-tier fallback. Actual expensas are used when the listing publishes them. Otherwise the model uses the median for other units in the same building, then the barrio average per square metre, then a percentage of rent. Each fallback is less precise than the one before. About 40% of direct rentals fall back at least one tier.
• Noise model omits indoor orientation. A courtyard-facing unit is typically 10–20 dB quieter than a street-facing unit in the same building. The model has fields for this (bedroom_faces, faces_street) but they are extracted from photos at only 14% coverage. For the other 86% the noise score reflects outdoor exposure without the orientation adjustment, so street-facing units inside quieter buildings look better than they are and interior units inside noisy buildings look worse.
• Hedonic R² is 0.51. The price model explains 51% of observed variation across 6,046 listings. The other 49% is noise, unobserved features, and listing-specific quirks (condition, dispute history, owner motivation). The ranking uses hedonic regression for validation, not for pricing, so the ceiling is not binding on the final score. But users who read "R² = 1" into price estimates should temper the expectation.

Each of these limits is actively being narrowed. The raw data is open. Errors are welcome: [email protected].

13. How to cite this work

Unitrank data is released under Creative Commons Attribution-ShareAlike 4.0. You may quote figures, embed tables, or republish datasets with attribution. Commercial use is permitted under the same licence terms.

We release the data because we would rather be cited than hidden behind a paywall. Buenos Aires rental-market figures are worth more in circulation (in journalism, policy work, academic research, and partner tools) than locked up. The standard is simple: attribute the source, keep derivatives under the same licence.

Suggested citation

Unitrank (May 2026). Buenos Aires rental market data. Retrieved from unitrank.com/en/methodology. Licensed CC BY-SA 4.0.

For journalists and researchers

• Named entity: Unitrank on Wikidata (Q138792840).
• Bulk data: monthly snapshots at github.com/unitrank/public-data.
• Custom queries: barrio, stay duration, or household composition cuts available on request.
• Review copies: pre-publication fact checks within 24 hours for deadline-bound work.
• Attribution line: "Source: Unitrank, Buenos Aires rental market data, May 2026".

Questions about methodology: [email protected].