Most Accurate Home Value Estimator? Zestimate, Redfin and AVMs Compared

Two estimates for the same house, $60,000 apart: that is the normal experience of anyone who has typed an address into Zillow and Redfin on the same evening. This article explains what these tools are, what their publishers say about Zestimate and Redfin Estimate accuracy, why they disagree, and how to test any estimate in about ten minutes. Every route to a home value, from the assessor's card to a licensed appraisal, is laid out in the main guide to how much is my house worth, while this piece stays with the online estimators.

What an AVM is and why every online estimator is one

Every free home value tool, from the Zestimate to your bank's calculator, is an automated valuation model (AVM): a statistical model that predicts a sale price from data it can read without visiting the house. Federal banking regulators describe AVMs as computerized models used to estimate the value of a dwelling and regulate the ones lenders rely on²; Congress wrote the same concept into 12 U.S.C. § 3354, which requires AVMs used in mortgage lending to follow quality-control standards³.

The inputs are the same everywhere: county assessor and recorder data (size, lot, year built, last sale), MLS listing feeds where the company has access, and whatever the owner has typed in. The model finds recent sales it considers similar, weights them and produces a point value, sometimes with a range. What it never has is a walk-through: the kitchen you gutted in 2023 or the roof that leaks does not exist for the model until it shows up in a listing or a permit.

That is the design, not a flaw: an AVM puts a number on every home in the country without visiting any of them. The question is how far to trust one of those numbers, and the publishers have answered it themselves.

Most accurate home value estimator: published error rates

Redfin states on its Redfin Estimate page (checked September 2026) that the estimate has a median error rate of 1.88 percent for homes that are for sale and 7.35 percent for off-market homes, that it draws on full MLS access, and that it is recalculated daily for listed homes and weekly for off-market ones¹. Zillow publishes its own nationwide median error rates for the Zestimate on its Zestimate page, split the same way; Zestimate accuracy figures change with every model update, so read the current ones there rather than a number quoted secondhand.

Two things matter more than the decimals. First, the on-market number does not measure the model's skill. Once a home is listed, the model sees the asking price, the photos and the agent's description, and the sale price it is later scored against usually lands within a few percent of that asking price; the estimator is largely predicting that the list price is about right. The off-market figure is the honest one for an owner who has not listed yet.

Second, a median error is not a maximum. A median of 7.35 percent means half of all off-market homes were estimated more than 7.35 percent from their sale price, in either direction, and the statistic says nothing about the worse half. At the national median list price of $424,500 that Realtor.com reported for August 2026⁴, 7.35 percent is about $31,000, and half of all homes are further off than that.

Why estimates diverge: data coverage, condition, renovations, unique homes

Two models aimed at the same house can land $60,000 apart because they are not looking at the same evidence.

Data coverage. Texas is the largest of ten states where the sale price never enters the public record; Alaska, Idaho, Wyoming, Kansas, New Mexico, Louisiana and Mississippi are the others, and Missouri and Montana disclose in some counties but not in the rest⁵. A model built on recorder data has little to learn in those places; a model with MLS access at least has the agent-reported price. There, the gap between estimators is structural rather than accidental.

Condition and renovations. County records carry the year built and the square footage from the last permit, not the condition. A 1962 house with a new kitchen, roof and windows and its untouched twin next door look identical to the model until one of them is listed. Owners who type their renovations into one portal move that portal's number and nobody else's.

Unique homes and thin markets. Models need similar recent sales. A hillside lot with a view, a house twice the size of its neighbors, a farmhouse on ten acres, a street where three homes sell a year: each produces a wide, unstable estimate, and the tools rarely say so.

Market tempo and time adjustment. Realtor.com's August 2026 report had the typical US home sitting 60 days before it sold, with list prices below the previous year for the tenth month running⁴. A model trained on last spring's sales drifts high in a cooling market until new sales pull it down, at a speed that depends on each company's time adjustment. The independent check is the FHFA House Price Index, a weighted repeat-sales index published down to county, ZIP-code and tract level⁶.

None of this is dishonest; it is why two numbers exist and why neither is a price.

The free-estimate business model: your contact data is the product

The estimate is free for a reason: portals earn their money from agents, who pay for advertising and for the leads a homeowner generates by asking what the house is worth. The number on the screen is the hook; the form that asks for a phone number "to refine your estimate" or "to connect you with a local expert" is the product, which is why people search for a home value estimator without giving their information.

There is nothing wrong with talking to an agent when you want to sell, and something wrong with treating a number produced to trigger that call as evidence. CheckValue sits on the other side of that line: the report is paid for by the person who orders it, the address is not sold as a lead, and no phone number is required.

Two portal estimates $70,000 apart, one Los Angeles sale at $958,000

Suppose a 1,600 sq ft, three-bedroom house in Los Angeles County, where the median single-family price was $946,950 and the median time on market 30 days in August 2026⁷. The figures below are a hypothetical example of how estimates behave, not a real address.

Portal A learned nothing new about the house between month six and listing day; the kitchen, the roof and the lot were the same. It learned the asking price and moved $23,000 toward it. That is the whole on-market effect, and it is why "my Zestimate jumped after I listed" is such a common story: the model saw your agent's number, not your granite.

Both off-market estimates were inside the published median error and still $70,000 apart. The agent's three closed sales, which cost an afternoon, were closer than either model: they were the sales an appraiser would later use.

How to sanity-check any estimate in ten minutes

You do not need a model to test a model; you need the evidence it is supposed to summarize. The best home value estimator for your situation is the one that survives this check.

1. Pull three closed sales. Same neighborhood, similar size and age, closed within the last 12 months: the standard Fannie Mae sets for appraisals, which requires a minimum of three closed comparables and an explanation for older sales⁸. Closed prices are in the county recorder's index in disclosure states and on the portals; the comparable sales guide shows where to find them and how to adjust.
2. Compute price per square foot for each sale and for the estimate. If the estimate implies a figure outside the range of the three sales, either the model is off or your house is genuinely unusual. Keep the comparison to sales within about 20 percent of your size: a small house carries a higher figure per square foot than a large one on the same street, for reasons the price per square foot guide takes apart.
3. Check the tempo. How long did those three sales sit, and were they cut first? A street at 25 days on market and one at 90 need different caution; the national median was 60 days in August 2026⁴.
4. Time-adjust the oldest sale. Look up the FHFA index for your metro or ZIP code and move a sale from ten months ago to today's level before comparing it⁶.
5. Look for a range. A point value with no range hides its uncertainty; if there is a range, its width shows how much the model trusts itself, and a 15 percent range on a $950,000 house is about $140,000 of admitted doubt.

If the estimate sits inside the adjusted range of the three sales, it is a fair starting point; if not, trust the sales.

The 2025 AVM rule: what regulators now require of lender AVMs

Since October 1, 2025, AVMs used by mortgage originators and secondary market issuers to value a consumer's principal dwelling must follow an interagency quality-control rule issued by the OCC, the Federal Reserve, the FDIC, the NCUA, the CFPB and the FHFA². Institutions must adopt policies and controls designed to ensure a high level of confidence in the estimates, protect against manipulation of data, avoid conflicts of interest, require random sample testing and reviews, and comply with applicable nondiscrimination laws². The first four factors come from the statute, 12 U.S.C. § 3354; the fifth was added by the agencies³.

Two consequences for homeowners. First, the rule does not cover the free portal estimate. It governs models used in mortgage decisions; a consumer-facing Zestimate or Redfin Estimate is outside its scope². Second, if your lender uses an AVM or an appraisal to decide on your loan, Regulation B gives you the right to a copy of that appraisal or other written valuation, promptly on completion or three business days before closing, whichever comes first, and the lender may not charge for the copy⁹. That document, or the closed sale price, is the honest benchmark for any estimate. This is general information, not legal or tax advice.

A range, cited sources and the same answer twice from CheckValue

CheckValue is an AI valuation report, so it is also a model, and it claims no accuracy percentage; the sections above explain why such a number would mislead. What it does differently is show its work, which for a reader who has just compared two portal numbers comes down to three things:

• A range, not a bare figure. The point value arrives with the range around it and its price per square foot placed beside the district average, so the model's own doubt is visible before you compare it with anyone else's number.
• The evidence on the page. The closed sales behind the value, the official index for the neighborhood and the pace at which houses there are selling are printed rather than implied, and each adjustment (a garage, a pool, solar, the condition, the year built) carries its coefficient and the source it came from, followed by a numbered reference list. Steps 1 to 4 of the sanity check above can be run against the report itself.
• The same answer twice. Ask for the same address with the same facts a week later and the value does not move, because every result goes into a registry, is calculated twice by independent runs and has to pass a plausibility check before you see it. A portal number that shifts every Tuesday cannot offer that.

The page on how CheckValue keeps results consistent describes the method and its limits, and how an AI valuation works walks through comparables, adjustments and consistency checks. Before you decide, the free preview of your report lays out that same structure for your own address.

It is not a licensed appraisal and does not pretend to be one: nobody inspects the house. Once a lender, a court or the IRS is involved, the appraisal is theirs to order; the report gets you ready for that conversation and does not replace the document.

When an estimate is not enough: lender, court, IRS

A short list of situations where no estimate will do. A mortgage or refinance: the lender orders the valuation and you are entitled to a copy⁹. An estate, a divorce settlement or a dispute between co-owners: courts want a dated, signed opinion of value from a licensed appraiser. A gift, a step-up in basis or a donation of property: the IRS wants a qualified appraisal, not a screenshot. In each case the appraiser finds at least three closed comparable sales⁸, adjusts them for the differences and signs the result under a license number.

For everything short of that (deciding whether to sell, checking a listing price, settling an argument with a neighbor) the sales themselves and a report that shows them are faster, cheaper and more honest than a single free number. What an appraisal, a CMA and an AVM each cost and where each carries legal weight is set out in appraisal vs. CMA vs. AVM.

This article is general information, not legal, tax or investment advice. Figures and rules carry the year they were published; check the cited source for the current version.

Christian Eckmair

Co-founder, CheckValue · More than 20 years in real estate in Tenerife and Austria · Reviewed by Josef Eckmair MBA

Christian Eckmair is co-founder of CheckValue and of CBDNOL GmbH (Ansfelden, Austria). He has worked in real estate for more than 20 years in Tenerife and Austria – buying, renovating, letting and selling residential property – and writes about home values, buying checks and ownership records.

Articles by Christian Eckmair →