Where our numbers come from

1. Our data sources

Every figure on the site carries its own source and date at the point it is shown. Where a page combines several, the sources are listed at the foot of that page.

Property X-Ray and Suburb Insights cover Victoria only, because the state datasets they depend on are Victorian. Market Pulse and the calculators are national.

2. What we do to the raw data

We resolve your address to one parcel

This is the step everything else depends on, and it is not a lookup. The state property map is keyed on parcel identifiers rather than addresses, and addresses are ambiguous in every direction — units against lots, renamed streets, subdivisions that split one title into six. Your address is matched to a single parcel, and from then on every other dataset is queried by that parcel’s geometry, not by its name: the overlays whose boundaries intersect it, the easements that clip it, the zones that contain it, the sales recorded around it.

Getting the parcel wrong would put another property’s constraints on your report and you would have no way of telling. That is why an address that does not resolve cleanly returns nothing rather than a near-enough guess.

We measure the block, rather than repeat what we were told

Land area, frontage, side lengths and orientation are computed from the parcel polygon itself, not taken from a listing. That means walking the boundary edge by edge, working out which edge fronts the street and which edges face north — and a boundary that runs north to south is an east or west facing edge, not a north-facing one, which is the kind of distinction that quietly reverses a sun diagram if you get it wrong.

Easements are handled the same way. Rather than noting that one exists, we intersect its geometry with the parcel to work out how many square metres it actually takes, and what share of the block that is — because “there is an easement” and “the easement removes 28 m² from where you could build” are different pieces of information.

We aggregate, and we suppress

Suburb figures are averages across many properties. Comparables are selected by narrowing until there is enough to be meaningful — matching property configuration over a recent window first, widening to type and a longer window when that is too thin. Where too few sales sit behind a median for it to mean anything, we do not publish a number: a median built on two sales is noise wearing a dollar sign. We say the figure is unavailable and why. An absent number here is a deliberate choice, not a gap.

We simplify shapes for drawing

Boundary outlines on maps are simplified so they load quickly. That is a display decision and never affects a calculation — area, zone and overlay checks all run against the full-precision boundary.

We use a language model for the writing

Summaries, suburb profiles and the buyer-fit ratings are generated by a language model working only from the data above. It is instructed to cite specific figures and not to recommend. It can still get things wrong, which is why every claim carries a number you can check independently. How those ratings work is on how we score suburbs.

3. How AI and agents fit in

The rule: numbers are computed, not generated. Every figure — land area, easement coverage, travel time, median price, match score — comes out of deterministic spatial queries and arithmetic. Artificial intelligence enters afterwards, to organise, explain and prioritise what was measured. No language model is ever asked to produce a measurement.

A pipeline, not a prompt

A Property Finder search is not one call to a model. It runs as a multi-stage graph of around seventeen steps: normalising your brief, fetching candidate properties, applying exclusion filters, enriching each survivor with geography, travel times, points of interest and school zones, then scoring several dimensions in parallel before aggregating and explaining the result. Each stage is a discrete node with its own inputs and outputs, which is what makes a score traceable back to the step that produced it.

Because the stages complete at different speeds, results stream to you as they land rather than after everything finishes — a first ranking appears, then commute evidence fills in, then local amenities and schools, then the written analysis.

Agents that query, rather than recall

The conversational parts of Abora — asking for suburbs in plain English, comparing areas — run as tool-calling agents built on a state-machine framework. The agent does not answer from memory. It decides which tool it needs, calls it, and that tool runs a real query against the property and Census databases. The numbers in the answer are the numbers the database returned.

This is the design decision that matters most for trust. A language model asked about a suburb from memory will produce plausible figures that are quietly wrong. An agent that must call a tool to get a figure cannot invent one — the worst it can do is choose the wrong tool, which shows up as an odd answer rather than a convincing false number.

Governed data access

Tools are not ad-hoc. Data capabilities are declared in a central registry — what each provides, where it comes from, who owns it, and its operational status — so the set of things an agent can reach is enumerable and reviewable rather than scattered through the code. Long-running conversations checkpoint their state so a multi-step search can resume rather than restart.

How we constrain the model

• Invent figures. It works only from data supplied to it, and is instructed to cite the number behind every claim.
• Deliver verdicts. Prompts direct it to describe rather than judge — setting out what a fact enables and constrains instead of rating a property as good or bad.
• Manufacture balance. Where a fact has only a downside, it is required to say so rather than invent an upside to sound even-handed.
• Override a measurement. If the arithmetic and the prose disagree, the arithmetic is right and the prose is a bug.

Output is checked before it is stored — that ratings sit in range, that claims carry figures, that the structure parsed cleanly. Those checks feed the confidence rating shown on suburb pages, explained in how we score suburbs.

4. How to check us

Every figure we publish comes from a source you can go to yourself, and on anything that decides a purchase you should.

• Planning and overlays — the council’s planning department, or a planning certificate.
• Boundaries and title — the title and plan of subdivision, via your conveyancer.
• School zones — the Department of Education’s zone finder, and the school itself.
• Demographics — ABS Census QuickStats.

If a number here disagrees with its source, that is a bug on our side and it is fixable. Tell us and we will tell you what our source says.