Your ChatGPT Property Prompts Are All Wrong: The Street-Level Data AI Cannot See

There's a story doing the rounds right now that I can't stop thinking about. A tech CEO in Sydney had a dog named Rosie diagnosed with terminal cancer. Vets said there was nothing left to try. Standard treatments had failed. The prognosis was weeks, maybe a couple of months.

So he spent $3,000 on whole-genome sequencing at UNSW. Got back 320 gigabytes of genetic data — the equivalent of roughly 700,000 pages of information. Every nucleotide, every mutation, every anomaly in Rosie's cancer cells, mapped and digitised. Then he fed the entire dataset into ChatGPT, AlphaFold, and Grok.

He designed a custom mRNA vaccine from the output. Rosie's tumour shrank by 75%.

That story is remarkable. But the lesson most people take from it is dead wrong. They hear 'AI saved the dog' when the real lesson is '320 gigabytes of precise, granular, patient-specific data saved the dog.' The AI was the processing engine. Without that mountain of data to work with, it would have generated the same generic advice any textbook could give you.

Now let me ask you something direct. Something I ask every client who tells me they've been 'using AI to research properties.'

When you ask ChatGPT for property investment advice, what are you feeding it?

'I want to buy an investment property in Victoria. Budget $700,000. Where should I look?'

That sentence contains roughly 80 bytes of data. Compared to 320 gigabytes, it's not even a rounding error. It's 0.000000025% of what the dog got. And the answer you get back? It's worth exactly what you put in — which is almost nothing, dressed up in confident-sounding paragraphs.

What AI actually knows about Australian property (and what it doesn't)

Let me give AI its due. I use it myself. My team uses it. It's brilliant at aggregating publicly available information at speed. Ask ChatGPT about median house prices in Cranbourne, population growth trends in Melbourne's south-east corridor, or the history of RBA rate decisions, and you'll get a serviceable answer. It can summarise reports from CoreLogic, Domain, and the ABS faster than any human analyst. For macro-level research — the 30,000-foot view — it's a genuine productivity accelerator.

But here's what it cannot do — and this is where the real danger lives for anyone making a six- or seven-figure purchasing decision based on AI output.

AI cannot see zoning boundaries at the street level. Take a suburb like St Albans in Melbourne's west. On one side of a single street — literally across the road — you might have properties zoned Residential Growth Zone (RGZ). Under RGZ, a single 600sqm block can potentially be developed into four or five dwellings. The land value reflects that development upside. On the other side of the same street, properties are zoned General Residential Zone (GRZ1). Under GRZ1, the same 600sqm block permits a maximum of two dwellings.

The zoning changes the development potential by 250%. It changes the land value by hundreds of thousands of dollars. And it changes on a house-by-house basis within the same street, sometimes within the same block.

Ask AI 'should I buy in St Albans?' and it gives you a suburb-level answer. It'll tell you about population growth, transport links, and median price trends. All useful at 30,000 feet. But you don't buy a suburb. You buy a specific block on a specific street. And at that resolution — the resolution where money is made or lost — AI is completely blind.

AI cannot access council-specific overlay data. Planning overlays — flood zones, heritage restrictions, bushfire zones, environmental significance overlays, design and development overlays — are maintained at the local council level. Some are publicly searchable through the Victorian planning scheme online. Many are not readily accessible without a formal property search or a direct enquiry to the council's planning department.

And here's the part that really matters: some overlay categories have sub-types that exist only in internal council records or in local institutional knowledge:

• Public flood zone designations (searchable online through council mapping tools or VicPlan)
• Council-internal supplementary flood data (requires a formal Section 32 vendor's statement search, a planning certificate, or a direct call to the council's engineering department — and even then, some councils are slow to disclose)
• Local knowledge of areas that flood in practice but aren't officially mapped — because the flood modelling hasn't been updated in 15 years, or because a drainage upgrade changed the flood behaviour, or because the area only floods in 1-in-50-year events that haven't been formally documented

AI knows about category one. It might reference category two if someone has written about it online. It has zero access to category three. And category three is often the one that actually determines whether your property floods after a heavy storm — which can destroy tens of thousands of dollars in value overnight and make the property uninsurable at reasonable premiums.

The easement trap and the face-width problem

Two data points that can make or break a development play — and neither exists in any dataset AI can access.

Easements. An easement is a legal right for a third party (usually a utility company, water authority, or council) to access part of your land for infrastructure purposes. A drainage easement running through the middle of your backyard — which might be only 2-3 metres wide on paper — can completely kill a subdivision or granny flat build. The easement doesn't just sterilise the strip of land it covers; it effectively renders the entire rear portion of the block undevelopable, because you can't build any permanent structure that encroaches on or crosses the easement.

I've seen a 650sqm block in Narre Warren that looked perfect for a dual-occupancy development on paper — flat, good shape, right zoning. But a 3-metre drainage easement ran diagonally across the back third of the block. It killed the subdivision entirely. The owner paid full development-potential price and got a single-dwelling block.

Easement data lives in the property's Certificate of Title, registered with Land Use Victoria. You have to order it individually for each property you're investigating — there is no bulk dataset, no API, no way for AI to access this information programmatically. Each title search costs $30-$50 and takes 1-3 business days. There is simply no shortcut.

I've seen clients lose $50,000-$100,000 in potential development value because of an easement they didn't check before purchasing. And I've seen other clients gain that much by buying the property next door — same street, same block size, no easement, full development potential.

Face width (frontage). The width of a property's street frontage determines how many dwellings can be built under the applicable ResCode provisions (Clause 55 and 56 of the Victorian Planning Provisions). In most Victorian councils, you need a minimum face width of approximately 15.4 metres for a dual-occupancy development on a standard residential block. Some councils apply even stricter local variations.

A 600sqm block that's 20 metres wide can comfortably accommodate two dwellings side by side. A 600sqm block that's 12 metres wide cannot be subdivided under any circumstances — the face width simply doesn't permit it. Same suburb. Same total land area. Completely different investment outcome. One block might be worth $800,000 as a development site; the other is worth $600,000 as a single dwelling. A $200,000 difference driven by a measurement that AI has no way to know.

Face width data exists in survey plans and council property records. Some councils make it searchable online; many don't. And even where it's technically available, interpreting it correctly requires understanding how the specific council applies ResCode to irregular-shaped blocks, corner lots, and lots with curved frontages. That interpretation lives in the heads of town planners and experienced buyer's agents, not in any AI training dataset.

Bushfire zones, high-voltage lines, and the data AI will never have

Let me catalogue a few more blind spots that I encounter regularly in my work — things that can swing a property's value by five or six figures and that no AI system can currently detect.

High-voltage transmission lines. Properties within 100 metres of high-voltage lines face measurable impacts on both value and desirability. Research on potential health effects (particularly childhood leukaemia risk) has been debated for decades without definitive resolution, but the perception alone affects what buyers and renters will pay. Almost all bank valuers will flag proximity to powerlines as a negative factor in their assessment. Some lenders apply a blanket 5-10% discount to properties within visual range of major transmission infrastructure.

But here's the thing: the precise routing of transmission lines — which side of the street they run on, how far specific properties sit from the nearest tower, whether the lines pass behind the property or in front — isn't captured in any standardised, searchable dataset. You need to physically inspect the site, review high-resolution aerial imagery, or check the energy distributor's asset maps (which are not publicly indexed by AI).

We bought a property in Glen Waverley for a client at $1.3 million. It was within 79 metres of a major high-voltage transmission corridor. That proximity suppressed the purchase price by an estimated $150,000-$200,000 compared to equivalent properties further from the lines. Our client knew exactly what they were buying and why — they intended to convert it to a rooming house where the per-room rental income ($2,000 per week total) would deliver an 8% yield, effectively using cash flow to offset the capital growth limitation. That's a deliberate strategy. But it only works if you know the powerlines are there before you buy. AI wouldn't have flagged it.

Bushfire overlays come in grades that matter enormously. Victoria's planning scheme distinguishes between the Bushfire Management Overlay (BMO) and the Bushfire Prone Area (BPA) designation. BMO is the severe category — it triggers mandatory building standards (BAL ratings), can prevent certain types of development, significantly impacts insurance costs (often adding $2,000-$5,000 per year to premiums), and can make the property harder to sell to risk-averse buyers.

BPA is milder — in many cases it doesn't materially affect insurance pricing, doesn't restrict development, and has minimal practical impact on property value. But both designations use the word 'bushfire,' and to an untrained eye (or an AI processing natural language), they look similar.

AI might tell you a suburb has 'some bushfire risk.' It won't tell you whether a specific property sits in BMO or BPA, whether it's at the edge of the overlay (where a small boundary adjustment could change the designation), or whether the overlay is under review by the council (which happens more often than people realise).

Slope and drainage. A block that looks flat on Google Maps — and that AI would describe as a flat suburban lot — might have a 3-metre fall from front to back. That fall changes everything about construction cost, retaining wall requirements, and stormwater drainage engineering. A sloping block can add $50,000-$100,000 to a build cost compared to a flat block of the same size in the same street. On a granny flat build, the difference between a flat site and a sloping site can be the difference between a viable project and an uneconomic one.

You can't measure slope from a database. You measure it by standing on the site. None of this is in AI's training data. It's in the legs and eyes of the person who walks the site every week.

AI will replace bad buyer's agents (and that's a good thing)

I want to say something that might sound counterintuitive coming from someone who runs a buyer's agency: AI will make the property industry better by destroying the bottom tier of buyer's agents. And I'm glad about it.

There are buyer's agents operating in this market right now whose entire value proposition is searching realestate.com.au, filtering by suburb and price range, maybe running a basic comparable sales analysis, and sending you a shortlist with some commentary. That's it. That's the $10,000-$15,000 service. Some of them don't even inspect the properties they recommend — they rely on listing photos and agent descriptions.

AI does that for free. And it does it faster, more consistently, and without charging you a percentage of the purchase price. These agents deserve to be replaced. They're not providing value that justifies their fee.

But the agents who survive — and I'd put our team at Optima firmly in this category — operate on street-level intelligence that no training dataset contains and no AI model can replicate. We know which side of Boronia's main road has 800sqm blocks and which side has 600sqm. We know which planning officer at Casey Council is receptive to granny flat applications with minor setback variations and which one will send you back to the drawing board three times for the same thing. We know that a specific street in Hampton Park experiences localised flooding after heavy rain despite not appearing in any official flood overlay — because we've seen the water marks on the fences and talked to the neighbours who've lived there for 20 years.

That knowledge comes from physically inspecting thousands of properties over years. It comes from relationships with council planners who tell you informally which applications are likely to succeed. It comes from building inspectors who've seen the same construction defects in 50 houses by the same builder. It comes from local real estate agents who call us before a property is listed because they know we'll move fast and settle without complications.

It also comes from getting things wrong and paying the price. I've bought properties with hidden easements. I've missed a slope that added $60,000 to a build quote. I've recommended a street that turned out to have a noise issue from a factory three blocks away that only operated on certain nights. Each mistake cost money and created knowledge that no dataset contains.

AI is a tool. A very good tool. We use it ourselves — for data aggregation, market trend analysis, initial suburb screening, and report summarisation. It saves us hours every week. But the jump from 'AI-assisted research' to 'AI-directed purchasing decision' is where people are going to get badly hurt. A $700,000 purchasing decision based on a 30,000-foot AI analysis is like performing surgery based on a WebMD search. The information isn't wrong, exactly. It's just missing everything that actually matters.

How to actually use AI in your property search (without getting burned)

I'm not here to tell you to ignore AI. That would be foolish. I'm here to tell you to use it properly — to understand what it's good at and, more importantly, what it can't do. Here's the framework I recommend:

Use AI for macro screening. It's excellent at suburb-level filtering: population growth rates, employment data, transport infrastructure plans, historical price trends, rental yield averages, vacancy rate data, school rankings, crime statistics. This is publicly available information that AI can aggregate and compare far faster than you can manually. Let it narrow your focus from 300 suburbs to 10 or 15 candidates. That's genuinely valuable work and it saves you weeks of research.

Do not use AI for micro selection. The jump from 'this suburb looks promising' to 'this specific property at this specific address is a good buy at this specific price' requires data that AI doesn't have and can't access: zoning boundaries at the block level, overlay status for the individual lot, easement positions from the Certificate of Title, face width measurements from survey plans, site slope and drainage conditions, power line proximity, council attitudes to specific development types, and dozens of other site-specific variables. This is where human expertise, local knowledge, and physical inspection are non-negotiable.

Pair AI outputs with boots-on-the-ground verification. If AI says Cranbourne looks promising based on the macro data, great — that's a useful starting point. Now drive every street in the target area. Check the planning scheme overlay maps for each block you're seriously considering. Order the Certificate of Title and read the easement plan. Inspect the site in person — walk the block, look at the slope, check for powerlines, note the condition of neighbouring properties. Talk to the council's planning department about what they'll approve. Call the selling agent and ask questions that aren't answered in the listing description.

Feed AI better inputs if you're going to use it. Instead of asking 'where should I invest in Melbourne?', try something like: 'Analyse the 12-month price growth, rental vacancy rate, and planned state government infrastructure spend for Cranbourne, Hampton Park, and Narre Warren. Compare the land-to-price ratios for established houses on 600sqm+ blocks in each suburb. Identify which suburb has the lowest ratio of new dwelling approvals to population growth.' The specificity of your question determines the quality of the answer. Feed it 80 bytes and you get an 80-byte answer. Feed it a structured, data-rich query and you get something approaching useful.

Never let AI be the final decision-maker. Use it as one input among many. The best property decisions I've seen come from a combination of data analysis (where AI excels), local market knowledge (where experienced agents excel), physical inspection (where your own eyes and feet excel), and professional advice from conveyancers, building inspectors, and mortgage brokers (where licensed specialists excel). Remove any one of those inputs and you're flying partially blind.

The dog's owner understood this instinctively. He didn't ask AI 'how do I cure my dog?' — a generic question that would have generated a generic and useless answer. He spent $3,000 getting 320GB of patient-specific data, then used AI to process what no human brain could handle alone. The combination of precise data plus AI capability produced a result that neither could have achieved independently.

Your property decision is a $700,000 bet on your financial future. It deserves the same rigour. Feed the machine properly, or don't use it at all.

320 gigabytes of data saved a dog's life. Your biggest financial decision is worth at least the same investment in getting the inputs right.