AVM Accuracy in 2026: What Zillow, Redfin, CoreLogic, and Quantarium Each Get Right — and Wrong

Every AVM vendor tells you their model is accurate. Every one of them grades their own homework. If you're an investor trying to decide which estimate to trust — Zillow's, Redfin's, CoreLogic's, Quantarium's — the honest answer is that "which one is most accurate" is the wrong question. Here's the right way to think about each model in 2026, what each is structurally good and bad at, and the blind spots they all share.

How accurate are AVMs in 2026?

For a typical on-market home in a data-rich metro, the major AVMs land within low single digits of the eventual sale price at the median — Zillow's own published figures put the Zestimate's median error around 2% for listed homes. For off-market homes, the same published figures jump to roughly 7%, and that's a median: half of the misses are bigger. No vendor has published independently verified numbers showing they've solved the off-market problem.

Two things follow from that:

1. On-market, most major AVMs are genuinely useful. The models are looking at a live listing — photos, list price, agent-entered facts. They're triangulating against strong signal.
2. Off-market — where investors, wholesalers, and acquisition teams actually operate — every model is guessing from stale records. This isn't a Zillow problem or a Redfin problem. It's a data-starvation problem, and it applies to all of them.

Why can't you compare AVM accuracy claims directly?

Because each vendor measures accuracy on its own test set, with its own metric, filtered by its own confidence rules. One publishes median absolute error on homes that sold; another cites the percentage of predictions within 10% of sale price (PPE10); most only score themselves where they chose to produce an estimate at all. Same word — "accuracy" — different exams.

The specific traps in vendor accuracy claims:

• Survivorship in coverage. A model that declines to estimate hard properties scores better on the ones it accepts. Coverage rate and accuracy trade off against each other, and marketing pages only quote the second number.
• On-market skew. Homes that sell were, by definition, mostly listed — so "accuracy vs. sale price" studies structurally over-sample the easy case.
• Metric shopping. Median error, mean error, PPE10, PPE20 — a vendor can be best-in-class on one and mid-pack on another.
• Timing. Scored against sale price at contract? At close? Using the estimate from when? A model that quietly re-fits after a listing goes live looks better than it behaves off-market.

There is no independent, apples-to-apples public benchmark across all the major AVMs. Anyone who tells you definitively that model X is "the most accurate" is either quoting a vendor or selling something.

What does Zillow's Zestimate get right — and wrong?

The Zestimate's strength is scale and freshness on the consumer side: coverage of on the order of 100 million U.S. homes, fast ingestion of listing events, and a published accuracy methodology — which is more transparency than most of the industry offers. Its structural weaknesses: homeowner-editable property facts, condition-blindness, and the documented on-market/off-market accuracy gap.

For investors specifically: the Zestimate is a reasonable first anchor on standard on-market product, and it's the number your seller will quote you, so you need it either way. Treat it with the most suspicion on off-market targets, heavy-renovation properties, and anything with unusual lot or use characteristics.

What does Redfin's Estimate get right — and wrong?

Redfin's Estimate is built by a brokerage with direct MLS participation, which gives it high-quality agent-entered data in the markets Redfin serves. The flip side is narrower coverage than Zillow — where Redfin has thin or no MLS presence, you get no estimate or a weaker one.

For investors: where both exist, the Zestimate–Redfin gap on a specific address is itself a signal (we wrote a whole piece on reading that spread). Where Redfin doesn't cover, you've just learned something about data density in that market.

What do CoreLogic and Quantarium get right — and wrong?

CoreLogic (the company rebranded as Cotality in 2025; the AVM is still labeled CoreLogic where consumers see it) and Quantarium are the AVMs behind much of the lender and portal world — both appear in Realtor.com's RealEstimate panel alongside Collateral Analytics. Their strength is public-records depth and mortgage-industry tuning; their weakness, from an investor's chair, is opacity. Neither publishes consumer-facing accuracy data the way Zillow does, so you can't independently check how they perform in your market or segment.

• CoreLogic sits on one of the largest property-data assets in the country — parcel records, mortgage data, tax data. Its models are built for lender risk workflows, which tends to mean conservative, defensible numbers on standard collateral.
• Quantarium markets an AI-heavy approach (including computer-vision signals) and is one of the fastest-refreshing models in the RealEstimate panel. The marketing claims are strong; the public evidence is thin. Score it in your own market before you weight it.

For investors: these two are worth pulling precisely because they aren't the portal models — they come from different data and different objectives. Agreement between a consumer model and a lender-grade model is worth more than agreement between two consumer models.

What about the AVMs you never see?

A handful of institutional AVMs rarely show up on consumer sites but move real decisions: Collateral Analytics (now part of ICE, via Black Knight), First American's valuation models, and ICE's own suite price collateral for lenders, servicers, and REITs every day. Consumer-facing, Homes.com (a CoStar company) shows a panel of third-party provider estimates rather than a single house number, Xome (a Mr. Cooper Group company) publishes its own value, and Movoto carries an estimate of its own.

The practical point isn't the org chart. It's that the estimates exist in silos. Zillow will never show you Collateral Analytics. Realtor.com won't show you the Zestimate. Each silo has one opinion, formed from different data. The full board — consumer portals plus lender-grade models side by side — is what an acquisition desk actually wants, and no portal has an incentive to show it.

What do all AVMs get wrong?

Every AVM, regardless of vendor, is blind to the same four things: current condition, interior quality, location externalities, and fast market turns. No model sees the failing roof, the gut renovation, the rail line behind the fence, or last month's shift in buyer behavior that hasn't closed into comp data yet.

• Condition. Public records don't know about deferred maintenance or the $80K renovation. This is the single biggest miss on investor-grade (i.e., distressed) inventory — and it cuts both ways.
• Geo-risk. Flood zones, transmission lines, highway and rail proximity, environmental sites — these move real transaction prices and largely aren't priced by the models. See our pieces on flood zone costs and transmission lines.
• Off-market staleness. Covered above; worth repeating because it's where investors live.
• Turning markets. AVMs learn from closed sales, which lag contracts by one to two months. In a fast-moving market, every model is looking in the rear-view mirror — and they don't all lag by the same amount, which is one reason spreads widen at turning points.

How should investors measure AVM accuracy in their own market?

Stop asking which model wins nationally and run the test that matters: score the models against recent closed sales in your own buy box. Pull estimates on 20–30 properties you know sold, compare each source to the actual price, and look at both individual error and — more useful — how the spread behaved when the models were wrong.

What you'll typically find is that no single source wins consistently, but the pattern is informative: tight spreads correlate with reliable consensus values, and wide spreads flag exactly the properties where every model struggled. That's the operational insight — use the spread as a confidence meter, the consensus as your anchor, and your own eyes for condition.

Frequently asked questions

How accurate is a Zestimate in 2026?

By Zillow's own published data: a median error of roughly 2% for on-market homes and roughly 7% for off-market homes, with figures that shift over time and by market. Half of all errors are larger than the median — and off-market is where the estimate is weakest.

What is the most accurate home value estimator?

No independent public benchmark crowns one. Each vendor publishes self-measured accuracy on its own coverage and metric, so the claims aren't comparable. Accuracy varies by market, property type, and listing status; the defensible move is to compare several models on the specific property and treat their agreement or disagreement as the signal.

Is CoreLogic or Quantarium more accurate?

There's no public apples-to-apples answer — neither publishes consumer-facing accuracy data the way Zillow does. Both appear in Realtor.com's RealEstimate panel; the useful investor move is watching when they agree or split on a specific address, and backtesting both against known sale prices in your own market.

Which AVM do lenders use?

Lenders and servicers typically use institutional AVMs — CoreLogic, Collateral Analytics (ICE), First American, and similar — often in cascades where a second model checks the first. That's worth noticing: the mortgage industry itself doesn't trust a single AVM. Multi-model checking is the institutional standard, not an investor exotic.

Why do investors use multiple AVMs instead of one?

Because the disagreement between models is information you can't get from any single model. Tight agreement means the property is algorithm-readable and the consensus is trustworthy; wide disagreement flags data errors, unusual properties, or turning markets — exactly the properties that need human diligence before an offer.

AVM values are estimates, not appraisals, and should not be relied on as the sole basis for any transaction.