Methodology, updated 13 September 2026

Does AI property valuation actually work?

This is the sceptic's question, and it deserves a plain answer rather than a testimonial. The short version: a photo-based analysis sees what a photograph sees, the evidence that it lands close is thin and is stated here at its actual weight, and the method earns its place on consistency and speed across many properties, not as a substitute for looking at one.

Every data page on this site links here so a reader can check how the numbers were produced. The measured back-tests for valuation and rent, with their sample sizes and failure cases, are on the accuracy page. This page is about the part that cannot be back-tested.

What the model can see

• The room type, from the floor plan and the photograph.
• Its condition: worn, dated, damaged, or recently done.
• The finish level and the fixtures: the units, the suite, the flooring, the windows.
• Obvious defects that show in a photograph: a stained ceiling, a cracked tile, a missing radiator.
• The floor area, from the plan where there is one and from the EPC record where there is not.

What it cannot see

• Behind plaster.
• Under floors.
• Into a roof.
• Damp, unless it has stained a visible surface.
• Wiring and plumbing.
• Structural movement.

That list is why any photo-based estimate is a floor, not a forecast. It prices the visible work. The final bill is the visible work plus whatever the survey and the first day on site find, and the estimate cannot know that number, so it does not pretend to.

The evidence, at its actual weight

For: one customer who has checked our refurbishment estimates against builder quotes on three deals reports landing within £500 to £1,000 of the quote. That is three deals and one person. It is encouraging and it is not a measurement, and this page will not describe it as anything broader than three deals from one customer.

Against: an investor viewed a property we had costed in September 2026 and judged it needed roughly £10,000 more than our estimate, for work the photographs could not show. He was almost certainly right. A real buyer, on a real viewing, moving a number by about a third is exactly the case the list above describes, and publishing it is worth more than any testimonial.

For the part of the analysis that can be measured, the end value, the back-test against 126,264 real UK sold prices gives a median absolute error of 15.3% with no systematic bias, and it is published in full on the valuation accuracy page. Refurbishment has no equivalent public dataset to test against, which is why no accuracy percentage is published for it and why the refurbishment accuracy page says so.

Where the method earns its place

On the first pass. The same method is applied to every property, so fifty listings can be compared on the same footing and three chosen to drive to. A person screening by hand applies a slightly different standard to the fiftieth listing than to the first, and takes a working week to do it. The analysis takes about three minutes a property and does not get tired.

The judgement stays with the buyer. The viewing, the survey, the builder's quote and the decision to bid are not things a photograph can do, and the product is built to get you to those three properties with the numbers laid out, not to replace what happens when you get there. The room by room cost data on this site is the aggregate of that first pass across thousands of properties; it is useful for exactly the reason above, and limited for exactly the reason above.

How to use it, then

• Treat the refurbishment figure as the floor for the visible work, and carry a contingency above it.
• Treat the end value as evidence from comparable sales, with the comparables shown so you can disagree.
• Use the analysis to choose which properties get your time, and the survey to decide whether to buy.
• When a figure surprises you, check the evidence behind it on the report before you act on it. Every number is shown with its source.

The honest test is a property you know. Run your own property through PropDetect and the same figures arrive for that address in about three minutes. No card needed for the first analyses.