Why online estimates fail on luxury homes, measured

Every valuation site admits, somewhere in the small print, that automated estimates struggle with luxury homes. None of them says by how much. We measured it, against officially recorded sales, and the answer is 30 to 72%, always on the low side. Here are the numbers, and what an honest estimate looks like instead.

The whole field agrees, and nobody publishes a number

Ask in any language whether online estimates work for a luxury home and the industry answers no. In the United States, the press covered the story of Zillow’s own then-CEO selling his Seattle home roughly 40% below its Zestimate. In France, every guide to valuing a high-end house ends with the same sentence: online tools do not work for luxury, contact an agency. In Germany, the valuation portals say luxury needs a human appraiser. The leading French portal’s documented method goes further: it excludes the top 5% of prices from its own model as anomalies. The luxury tier is not where these tools are weak. It is what they cut out before they start. And in all three languages, one thing is missing: a measurement. How wrong, in which direction, on how many real sales? Nobody says. We did the measurement.

What we measured, and how

We ran a median model, the same kind of model every online estimator is built on, over the officially recorded sales of six French luxury markets (the state’s open DVF register), and asked one question: for each real sale, how far off would the model’s prediction have been? Each sale was tested against a model built from all the other sales, never against itself. The result, market by market, for 2024:

Market	Segment	Median error vs real price	Recorded sales
Saint-Jean-Cap-Ferrat	houses	23.5%	16
Cannes	apartments	20.3%	2,246
Cannes	houses	29.2%	89
Courchevel	apartments	25.3%	112
Megève	apartments	20.5%	155
Gordes	houses	27.8%	29

The median error is 20 to 30% everywhere. Not on strange homes: on half of all recorded sales the model was at least this wrong. We know of no estimation tool, in any country, that publishes an equivalent figure about itself. Note Saint-Jean-Cap-Ferrat: one of the most expensive addresses in the world recorded sixteen clean house sales in a full year. Nobody can be statistically precise on sixteen observations. Including us, and unlike the others, we say so.

On luxury homes the error is not noise, it is construction

Split the same sales by where they sit in their own market and the error stops being random. Homes in the middle of their market: an honest 17 to 25% spread, in both directions. Homes priced above the middle: the model lands systematically UNDER the real price, by 30 to 48% on the premium tier and by 52 to 72% at the top, across every market we measured. The reason is mechanical. A median model prices the middle of a market, and the things that make a home worth multiples of that middle, the view, the address, the finish, the service that comes with it, appear in no public register. An automated estimate does not merely miss the premium. It cannot see it. That is why the same model that is respectable on a standard apartment tells the owner of an exceptional house that it is worth a third of its price.

What an honest estimate looks like

Since 26/07/2026 our own estimator does what we believe every estimator should do. For a home in the middle of its market, it gives the value and the band where the middle half of real comparable sales fell, with the sample size and the year of the sales behind it. For a home above the middle, it refuses to print a central figure it can measure to be wrong. Instead it publishes a floor built from the other thing we know about these homes: what they earn. A home that brings in a given rental income a year, counted in nights actually sold and never in occupancy percentages, is worth at least that income capitalised at 4.5% gross, the top of the yield band we measured across our covered markets. The real value sits above that floor, and how far above is exactly the information no public register holds.

If you own one of these homes

Then you hold the one number every model in this article is blind to: what your home really earned, or what it would really sell for. That number, next to the recorded sales and the rental benchmark we hold for your market, is the only honest way to place a luxury home, and it is what we build. Run the estimate, and if it refuses to give you a central figure, that refusal is the accurate answer, tell us your number and we place your home against its true market. Individual owner figures stay private, always. Sources: French DVF recorded sales (Etalab/DGFiP), backtests 07/2026, LuxVacation AI. Where we cannot measure, we say we do not know.

Published 26/07/2026. Figures generated from our live benchmark data and updated on recalibration.