Why an AI estimate and a mortgage valuation can disagree about your home

You look up your home online, ask an estate agent about selling, then apply for a mortgage. Three different values can emerge without the building changing at all. The US Consumer Financial Protection Bureau explains that valuations may use different comparable sales, dates and purposes. Understanding those differences is more useful than assuming that the highest figure is right or the computer must be wrong.[4]

The model values the home described by its data

An automated valuation model, usually shortened to AVM, estimates value from recorded information. RICS describes a spectrum ranging from statistical models to machine learning and neural networks, often combined with human review. Automation therefore describes a method of producing evidence; it does not tell you how much property-specific checking surrounds the result.[1]

Zillow’s Zestimate is one prominent US example. Its neural network uses property records, listing-service feeds and home details. Floor area, bedrooms, location, previous sales and market conditions supply signals. A current sale listing adds information that an unlisted property lacks. The result is an estimate of today’s market value from those inputs, rather than an inspection of the building.[2]

Even a reliable source arrives on a timetable. HM Land Registry’s Price Paid Data records completed property sales submitted for registration in England and Wales. Registration follows the transaction, so the newest months remain incomplete. A newly published database entry may describe an earlier sale; refreshing a website cannot make the underlying evidence instantaneous.[3]

Imagine two neighbouring homes with the same recorded size and bedroom count. One has recently gained a well-finished extension; the other needs extensive repairs. If those differences are absent from the inputs, the model cannot assess them as a visitor could. RICS identifies both asset condition and the availability of sufficiently comparable, frequently traded properties as central issues.[1]

A lender is making an additional decision

Hometrack’s UK Digital Valuer illustrates that larger process. Its service combines an AVM with checks on property attributes, title, ownership and lender policy exclusions. The lender can configure rules to select a valuation method. Its property-risk assessment also concerns whether the home is suitable mortgage security. A reasonable market-value estimate does not automatically settle that separate lending question.[5]

There are also rules about how certain mortgage models are used. A US regulation effective from 1 October 2025 requires covered institutions to maintain controls addressing confidence in estimates, data manipulation, conflicts, sample testing and applicable nondiscrimination law. It covers specified mortgage and securitisation decisions involving a consumer’s principal dwelling. It is not a quality certificate for every public house-price widget.[8]

Why putting the property up for sale can change its estimate

Zillow explicitly says that listing a home can change its Zestimate because the model receives new information, including the asking price and recent market activity. That helps explain a seemingly strange jump: the building has not changed, but its evidence has. The company says the estimate considers other factors too and does not simply copy the asking price.[2]

Some systems also try to extract condition information from language. Hometrack describes a natural-language-processing model that reads property listings for characteristics suggesting condition. This can add useful evidence, but a description remains a description. Inferring a renovation from an advert and examining the renovation are different ways of learning about the same home.[5]

Zoopla’s UK estimates make the evidence gap visible through confidence ratings. Recent data and numerous similar nearby sales support higher confidence; sparse or older comparable information lowers it. Its documentation also acknowledges that extensions and renovations may be missing. A confidence label concerns the available support for the estimate, rather than the seller’s conviction that the home deserves a particular price.[6]

Read the error measure before applying the percentage

Redfin publishes separate US accuracy figures for listed and unlisted homes. On 12 October 2026, its current page reported the following nationwide median errors. These are provider-reported measures, not our independent assessment.[7]

Median error describes the middle of the error distribution. Half the assessed estimates fall within that percentage of the sale price; half fall outside it. It is not a maximum error or a 95% confidence interval. For a hypothetical home that eventually sells for $400,000, 7.42% of that sale price is $29,680. That calculation illustrates the scale, not a guaranteed range around a particular online estimate.[7]

The listed and unlisted figures also answer different forecasting questions because the available information differs. A system estimating a home after its sale listing appears has evidence unavailable to the same system before listing. Treating the smaller listed-home error as the accuracy of an unlisted-home estimate would compare unlike situations.[2]

Turn a disagreement into a question about evidence

Start with the property record and comparison set: the right address and property type, floor area, tenure, condition, dates and genuinely similar sales. Ask which missing or different fact explains the gap. Zoopla lets owners edit displayed property details, but its help page says users and staff cannot manually feed new data into its AVM or amend the estimate range. Editing a profile and changing a valuation are not automatically the same operation.[6]

If a lender obtains an automated estimate early and a professional appraisal later, the CFPB explains that the two can differ because of their evidence and purpose. Requesting the reasoning behind that difference is more productive than treating one number as an instruction to the other.[4]

The useful outcome is a clearer account of the home: which facts were known, which were inferred, what remained unseen and what decision the number supports. AI can assemble a large body of market evidence quickly. Its estimate becomes easier to use when the reader can see what that evidence actually describes.

Report an error or suggest a correction