A fund buying across southern Europe cannot compare a Madrid listing to a Lisbon one until both are the same shape. That means not just collecting three national markets, but resolving the same property where it appears twice and normalising three different ideas of what a square metre includes.
The source is named because it is public; the client is not. Every figure on this page is a measured production number the client agreed to publish. Have a source of your own? Send it over, whatever its size — you get an answer within 24 hours.
property records to date
a year
measured coverage
properties in the graph
Not more rows. Every one of these projects started with somebody who already had data and could not use it for the decision in front of them.
The problem
The client invests across three markets and had three data sources with three schemas, three definitions of floor area and no way to tell whether a property in one file was the same asset as a property in another. Every comparison in their investment committee pack was being made with a footnote apologising for the data.
Fixed on day one
Every one of these is measured continuously and reported on the same dashboard the client watches. A commitment nobody measures is a sentence in a proposal.
None of these is solved by better headers or a bigger proxy pool. Each needs a different piece of engineering, and working out which one you are actually facing is most of the job.
Built area, usable area and plot area are reported differently in each market and sometimes inconsistently within one. A model that treats them as interchangeable produces valuations that are wrong by a third.
What we doArea fields are parsed into an explicit typed model with the source's own label retained, and derived comparables are computed only between like definitions.
Larger assets are listed by more than one agent, sometimes at different prices, and counting them twice distorts both supply figures and average price.
What we doEntity resolution on geometry, area, media fingerprints and description similarity, tuned against a hand-labelled set and re-measured weekly.
The two have different fields, different search behaviour and different value drivers, and most collectors handle one and bolt the other on badly.
What we doOne property model with typed extensions per asset class, so the shared fields are genuinely shared and the differences are explicit rather than crammed into a notes column.
Some listings carry a precise point, some a neighbourhood polygon, some only a municipality — and a fund's model is extremely sensitive to which.
What we doLocation precision is a first-class field. A record carries what is known and how precisely, so the model can weight it rather than treating a municipality centroid as an address.
The architecture is not interesting; every extraction system has a queue, a fetcher and a parser. These are the decisions that made this one work where the last one did not.
Built, usable and plot are separate typed fields with the source label preserved. Comparables are only ever computed between matching definitions, which is unglamorous and is the difference between a usable valuation and a wrong one.
Duplicate detection leans on what is hard to change — location, area, photographs — rather than on text, which agents rewrite freely.
Residential and commercial share a core and extend it explicitly. The alternative, two parallel pipelines, is how the two halves of a dataset drift apart over two years.
Every location record states how precisely it is known. Treating a municipality centroid as an address is the most common way property models quietly go wrong.
Coverage is measured separately for each country and each asset class. One weak country inside a healthy aggregate is exactly what per-country monitoring is for.
Production figures, not a benchmark run. Coverage is recounted daily against an independent sample of the live source rather than asserted, which is why the numbers are not round.
Daily profile
Twelve million records a day is 360 million a month, 4.4 billion a year and 12 billion since 2023, from 3.2 million requests — 37 a second across three countries. Sixty-four per cent of listings carry an exact point and the rest do not; publishing that number, rather than silently geocoding the remainder to a centroid, is what lets the fund's model weight the difference.
Headline
These four sit in the support agreement. When one of them drifts outside its band, we are alerted within fifteen minutes and fixing it is routine work under the monthly arrangement, not a change request.
The columns are the client's own numbers from before the rebuild and the measured ones from production today. The left column is the part most vendors would rather not put on a page.
| Metric | Before | Today |
|---|---|---|
Cross-country comparison |
With a footnote | Direct, one area definition |
Duplicate assets |
Counted twice | Resolved, sources retained |
Commercial coverage |
Separate, partial | In the same model |
Location handling |
Centroid where unknown | Precision as a field |
The investment committee pack lost its data footnote. That was the actual deliverable, whatever the statement of work said.
One case file per source, each opening the same way: what the client needed, what the source put in the way, how the pipeline is built and what the numbers are today.
The market, and the 1.9M people selling it
Read the case →
7.4M listings, 210M calendar-nights a day
Read the case →
46M itineraries priced a day, decomposed
Read the case →
94k routes, 210M fare quotes a day
Read the case →
340M listings watched, sold prices captured
Read the case →
Start here
Send the source, the fields you need and roughly how often. You get a straight answer within 24 hours: whether it can be done, what makes it hard, what coverage is achievable and roughly what it costs to build and to run. Whatever the size.
Direct
NDA before technical detail, as always. If it isn't our kind of work, we say so in the first reply.