Domain has one. REA has one. CoreLogic has one. Every major Australian property platform now offers an automated valuation model (AVM) — an AI-generated estimate of what your property is worth. These numbers influence buyer expectations, seller pricing decisions, and even bank lending assessments.

But are they any good?

I’ve spent the last month comparing AVM estimates against actual sale prices for 50 Melbourne properties that sold at auction in February and early March 2026. The results are illuminating — and should make you cautious about treating any single estimate as gospel.

What I Found

Across 50 properties spanning inner-city apartments to outer-suburban houses, here’s how the three major platforms performed:

CoreLogic AVM: Average deviation from sale price was 7.2%. That means on a $1 million property, the estimate was typically off by about $72,000 in either direction. For 68% of properties, the estimate was within 10% of the sale price. For 14%, it was off by more than 15%.

Domain Price Estimate: Average deviation of 8.5%. Slightly less accurate overall, with a noticeable tendency to overestimate prices in the inner suburbs and underestimate in the outer suburbs.

REA Property Estimate: Average deviation of 9.1%. Similar pattern to Domain but with wider variation. Several estimates were off by more than 20%.

For context, a qualified human valuer typically achieves 3-5% accuracy. So the AVMs are in the right ballpark, but they’re not a substitute for a professional valuation — especially when six-figure sums hinge on the number.

Why They Get It Wrong

Understanding why AVMs miss helps you interpret their estimates more intelligently.

Renovation quality is invisible to algorithms. An AVM can tell that a house in Hawthorn has three bedrooms and sits on 600 square metres. What it can’t reliably assess is whether the kitchen was renovated in 2024 with $80,000 of Miele appliances or hasn’t been touched since 1987. Interior quality is the single biggest source of AVM error.

Unique properties break the model. AVMs work by finding comparable recent sales and adjusting for differences. When a property is genuinely unusual — a converted warehouse, a house on an irregular lot, a heritage-listed terrace with restrictions — there aren’t enough comparables to anchor the estimate.

Timing gaps matter. AVMs rely on recent sales data, but in a moving market, even a two-month lag can introduce significant error. If the last three comparable sales in a street happened in November and the market has shifted since then, the AVM is calibrated to old conditions.

Auction premiums aren’t predictable. A property that attracts five bidders might sell $150,000 above a realistic pre-auction estimate. That’s not the estimate being “wrong” — it’s the auction dynamic producing an outcome that no model could reliably forecast.

How to Use AVMs Properly

Despite their limitations, AVMs are genuinely useful tools when used correctly.

Use them for rough screening, not final decisions. When you’re browsing listings and trying to gauge whether a property fits your budget, an AVM estimate saves time. Just treat it as plus or minus 10%.

Compare multiple estimates. If CoreLogic, Domain, and REA all suggest a property is worth $850,000-$900,000, that convergence is more meaningful than any single number. If they diverge wildly, that’s a sign the property has unusual characteristics that algorithms struggle with.

Check the confidence level. Most AVMs provide a confidence indicator. High confidence means lots of recent comparable data. Low confidence means the estimate is essentially a guess. Treat low-confidence estimates with extreme skepticism.

Supplement with recent sales data. Look at the actual sale prices of similar properties within 500 metres that sold in the last 90 days. This manual comparables check often reveals more than the automated estimate.

The firms building more sophisticated AI valuation models are getting better at incorporating data that traditional AVMs miss. An AI consultancy I spoke with recently described how natural language processing can now extract renovation details from listing descriptions and photos — feeding richer data into valuation models. This kind of multi-modal analysis is where AVMs are heading, but we’re not fully there yet.

What This Means for Buyers and Sellers

Buyers: Don’t walk into an auction with an AVM as your maximum bid. Get a bank valuation or hire an independent valuer for any property you’re serious about. The $300-$500 cost of a professional valuation is trivial relative to the potential error in an automated estimate.

Sellers: Don’t set your reserve based solely on what Domain tells you your house is worth. Have your agent prepare a comparative market analysis using actual recent sales, and get an independent appraisal if the stakes are high.

Investors: AVMs are better for portfolio-level screening (identifying suburbs or property types that appear undervalued) than for individual property assessment. The errors that make them unreliable for a single property tend to wash out across a larger sample.

The Future

AVMs will keep improving. Better data, better models, more training examples. Within five years, I expect average deviation to drop below 5% for standard properties. But the outliers — the renovated gems, the unique sites, the auction fever — will continue to confuse algorithms for a long time.

In the meantime, treat AI valuations as one input among many. They’re a starting point, not a verdict.

Chris Palmer is a Melbourne-based property analyst who spent twelve years as a licensed real estate agent.