How Accurate Are AI Home Valuations? A Deep Dive into the Data

Every automated valuation provider publishes an accuracy figure. Almost all of them publish the one statistic that makes the product look best, and it is not the one that tells you whether to trust the estimate on your house.

How these models work — public records, recent comparable sales, adjustments for the differences the data can see — and what they structurally cannot see is covered in full on our valuation page. This is about the layer above that: how to read an accuracy claim, and what a given number does and does not entitle you to conclude.

It matters because these figures are doing real work. People decide whether to sell, what to list at, and whether an asking price is plausible partly on the strength of them. A statistic chosen for marketing rather than for information is not a small problem.

Why the median error rate flatters every model that publishes it

The industry standard disclosure is median absolute percentage error: line every estimate up against the price the property eventually sold for, and take the middle one. If a provider reports a median error of a few percent, that is a real and non-trivial claim.

It is also, by construction, silent about half the estimates. The median says nothing whatsoever about the shape of the distribution above it. A model whose worst half misses by a little and a model whose worst half misses catastrophically can publish the same median. Nothing in the figure distinguishes them, and the difference is the entire question if your house happens to sit in the tail.

This is not an accusation of bad faith. The median is genuinely the right summary if you want one number for a whole portfolio of estimates. It is simply the wrong number for the question an individual homeowner is asking, which is not “how does this model do on average” but “how much should I believe this figure, for this house”.

The disclosure that actually answers your question

Ask instead for the share of estimates falling within a stated band of the eventual sale price — what proportion landed within five percent, and within ten. That is a statement about the whole distribution rather than its midpoint, and two providers with identical medians will usually separate sharply on it.

Better still is that share broken down by property type and by market. Accuracy is not evenly spread and never has been. It is highest where the model has abundant, closely matched evidence — dense tract housing, standard condominium buildings — and degrades predictably as evidence thins, through architecturally unusual homes, large or irregular lots, rural property, and neighborhoods with few recent sales.

• What is the share of estimates within five percent of sale price, and within ten? Not the median.
• How does that share vary by property type — and is my kind of property in the well-covered group or the thin one?
• What is the sample: which markets, over what period, and does it include the homes the model declined to estimate at all?
• Is it measured against sale price, or against another model’s output? Only the first is a test of anything.
• How recent is the measurement window, and does it span a period when prices moved?

A provider that cannot answer the first two of those is not necessarily hiding a bad model. But you are entitled to treat an unanswered question as an unanswered question.

Selection effects, and the estimates that never appear in the figure

There is a subtler issue that almost no published figure addresses. Accuracy can only be measured against homes that actually sold. Properties that were listed and withdrawn, or never listed at all, have no sale price to check against — so they are absent from every accuracy statistic ever published.

That absence is not random. Homes that fail to sell are disproportionately the unusual ones, the mispriced ones, and the ones in thin markets — which is to say, precisely the homes automated models handle worst. Every accuracy figure is therefore measured on a sample that is systematically easier than the full population, and is to that extent optimistic. Nobody can fully correct for this. It is worth knowing that it is there.

The same applies to coverage. A model that declines to estimate the hard cases and reports accuracy only on the ones it attempted will look better than one that estimates everything. Ask what proportion of properties the model returns a figure for at all, and compare like with like.

The free diagnostic that beats any published figure

Run your address through several tools and compare the answers. This costs nothing and tells you something no provider discloses: how much comparable evidence exists for your specific property.

When the estimates cluster tightly, the models are drawing on abundant closely-matched sales and are probably close to right. When they scatter, your home is unusual for its area and no automated model can price it well — regardless of what any of them publish about their median error. The disagreement is not a defect in the tools. It is the most honest signal available, and it is specific to your house rather than to a national average.

Where the spread is wide and the amount at stake is real, that is the point to get an appraisal — a licensed professional inspecting the property, seeing the condition, layout and setting that no record captures. In California, remember too that a sale triggers reassessment to market value, so the number has consequences beyond the transaction itself.

General information about how accuracy claims are constructed and should be read. Not an appraisal, and not financial advice. An automated valuation should not be relied upon where a formal valuation is required.