We Asked 7 Versions of GPT to Pick Australia's Top Growth Suburbs. Then We Checked.

The One-Line Summary

How We Tested This

Every version of GPT has a knowledge cutoff date. GPT-3.5 stopped learning in late 2021. GPT-4o stopped in late 2023. That creates a natural experiment. We can ask each version to forecast property growth, then check the actual prices for the years that followed.

We queried 7 versions of GPT across all 8 Australian states and territories. Each version was asked to name 20 suburbs most likely to grow, along with expected growth percentages and reasoning. We then measured actual suburb growth from the month after each version's exact publication date through to February 2026 using monthly hedonic price data.

The critical part: we compare each version's picks against all other suburbs in the same state or territory, over the same period. This controls for the broader market. If Sydney grew 15% overall, we are not asking whether GPT predicted 15%. We are asking whether the specific suburbs it chose grew more or less than 15%.

The Prompt

Each version received this prompt for every state/territory and multiple forecast years. All queries used temperature 0.2 for consistency.

What GPT Said Back

GPT-3.5-turbo, asked for NSW growth suburbs up to 2024:

Notice the generic reasoning. "Infrastructure development", "proximity to Sydney", "strong demand". These are the kind of phrases that appear thousands of times in real estate articles. GPT is reciting patterns from its training data, not analysing live market signals.

GPT Versions Tested

Price Growth: GPT Picks vs Market

These charts show indexed price growth from each version's publication date. Both lines start at 100. The solid line tracks the average hedonic price of GPT-picked suburbs. The dashed line shows the market baseline (all suburbs in the same states). If the solid line ends above the dashed line, GPT beat the market.

Did GPT Pick Winners?

This is the only question that matters for investors. If you followed GPT's suburb recommendations, did your property grow faster than average?

The answer, for most versions, is no. Six of seven GPT versions underperformed the market. Only GPT-4.1 (1,048 picks, +0.4% per year) beat the market. The versions with the most data (GPT-3.5-turbo with 2,154 picks and GPT-4o with 1,487 picks) both underperformed. You would have been better off picking suburbs at random.

The Older the Version, the Worse the Picks

GPT-3.5-turbo (released 30 Nov 2022) recommended suburbs that grew 0.4% per year slower than the market from December 2022 onwards. Only 44% of its picks beat the state median. GPT-4 (released 14 Mar 2023) underperformed by 0.5% per year, with 43% beating the median.

In Western Australia, the gap was extreme. GPT-3.5's picks grew 44.7% from December 2022 while the WA average was 61.0%. That is 5.1% per year of underperformance. On a $1M property, you would be $162,290 worse off following GPT. Only 12.7% of its WA picks beat the state median.

Look at what happened to investors who followed GPT-3.5 in WA. It recommended Peppermint Grove ($3.1M median) and Applecross ($1.6M median). These are Perth's most expensive, most-written-about suburbs. They grew 15.6% and 16.6% while the WA average was 61.0%. On a $1M investment, following that advice cost you $454k in Peppermint Grove and $444k in Applecross compared to buying randomly in WA.

The same pattern repeats nationally. In NSW, Berrima dropped 21% while the state grew 17.5%. In Victoria, Caulfield North dropped 22.9% while the state grew 7.2%. GPT consistently picked expensive, well-known suburbs that underperformed their state. Only 21% of GPT-3.5's WA picks beat the state median.

Newer Versions: Smarter, or Just Less Wrong?

GPT-4.1 (released April 2025) is the only large-sample version to beat the market, outperforming by 0.4% per year with 54% of picks beating the median. But 0.4% per year over 9 months is not a reliable investing edge.

GPT-4o and GPT-4o-mini both underperformed slightly (-0.4% per year and -0.2% per year respectively). GPT-4-turbo was worst among newer versions at -0.9% per year. The pattern is consistent: no version of GPT reliably picks suburbs that outperform the broader market.

GPT-4.1's picks in NSW beat the market by 1.8 percentage points, with 65% beating the median. In Queensland, it outperformed by 1.0 percentage points. But in WA, it underperformed by 1.2 percentage points. The wins and losses roughly cancel out, consistent with random variation rather than genuine forecasting skill.

The Numbers at a Glance

Visual comparison of how each version's suburb picks performed against the market. A positive number means GPT outperformed. A negative number means you would have been better off picking randomly.

What Happened to Investors Who Followed GPT?

GPT named its top picks. Here is what happened if you bought where it told you to, compared to buying randomly in the same state.

GPT'S WORST ADVICE

EVEN A BROKEN CLOCK...

State-by-State Performance (Selected)

Performance varies wildly by state. Some versions do well in one state and terribly in another. This is consistent with random variation rather than genuine forecasting skill.

Absolute Accuracy: How Far Off Were the Numbers?

Beyond picking the right suburbs, how accurate were the growth percentage predictions themselves? Every version under-predicted actual growth except gpt-4.1-mini. The Australian market grew far faster than any version of GPT expected.

Data Validation: Microburbs vs CoreLogic and Domain

We cross-checked our Microburbs hedonic price data against published figures from Cotality (formerly CoreLogic) and Domain. The state rankings match exactly. Growth magnitudes are directionally aligned, with differences explained by methodology.

What This Means for Property Investors

The ChatGPT property hype is real and growing. Social media is full of people asking GPT where to buy. This data suggests that is a bad strategy.

GPT is trained on internet text. It knows which suburbs get talked about. It knows the cliches: "close to the CBD", "infrastructure investment", "lifestyle appeal". But popular suburbs are already priced in. The growth happens in the places nobody is talking about yet.

Property forecasting requires local, timely data. Supply pipelines, council approvals, demographic shifts, wage growth by postcode, days on market trends. These are not things GPT can access. It has old text, not live signals.

Methodology

Download the Raw Data

All data from this research is available for download. Every GPT prediction, every scored comparison, every baseline figure. Use it for your own analysis.

Conclusion

Property forecasting requires live, hyper-local data. Supply pipelines, demographic flows, approval volumes, wage growth by postcode. These signals exist, and they work. But they live in databases, not in the training text of a language model.

GPT is predicting the next word. Not the next hotspot.