Vibe coding an EV charging map out of Australian open data

The short version

Australia has no national EV charger dataset. What it has is one community database, four state government publications, one territory's policy page, and a lot of inconsistency between them. Stitched together carefully, that gets you about four fifths of the country's public charging.

I wanted to know whether you could build a genuinely useful national charging map without paying for commercial data. The answer is mostly yes. The interesting part isn't the total, though: it's everything that goes wrong on the way there.

This post covers three things: what the combined data actually says about sites and plugs, who operates them (with a table of every operator I could identify and a verified link to each), and every problem I hit. I've been deliberate about including the mistakes that were mine, because they were more instructive than the publishers'.

Explore the map →

Sites and plugs: what the data says

After reconciling 3,806 source records into 3,157 distinct sites (merging 649 duplicates along the way), here is the shape of Australian public charging as open data sees it.

Sites by state and territory

Tasmania's bar is split: 81 sites with real coordinates, plus 55 known only to the town (purple). NSW dominates partly through genuine density and partly because NSW simply publishes more than anyone else.

Sites by charging speed

The grey bar is the problem: 1,447 sites (46%) have no usable power rating in any source. Only 54% of sites state their kilowatts.

How complete is each field?

Coverage is not the same as completeness. A site can be in the data and still tell you almost nothing useful:

Every operator I could identify

This is the table I most wanted when I started and could not find anywhere.

38 operators run five or more sites, accounting for 2,928 of 3,157 sites (93%).

Where the data comes from

Six sources, all openly licensed, none requiring an API key. Note how lopsided the record counts are, and that three of the six are limited to government-funded sites rather than everything in their state.

Everything that went wrong

This is the useful part. Not one of these was visible from a dataset's description page: every one turned up only by running the data.

1. The open-data portal is not where the open data is

I originally recorded the ACT as publishing nothing, because its open-data portal returns only federated US Department of Energy records. The ACT in fact publishes its full funded-charger list as HTML tables on a policy page. Re-checking the other jurisdictions with that lesson in hand found Tasmania doing the same thing.

2. A field that means the opposite of what it looks like

Victoria publishes estimated_project_completion. The obvious reading (populated means "not yet built") is wrong. 96 of 152 rows carry a value and most are past dates describing finished projects. Taking it at face value mislabelled 96 live sites as unbuilt. Parsing the date and comparing it to today gives the real answer: 11. Australian dates are day-first, so Date.parse() can't be used either.

3. Truncated names, in the publisher's own export

The TfNSW dataset contains "Fast Cities A" (17 rows) and "Viva Energy A" (7 rows): names cut off mid-word. Inconsistently, too: one row has the full "Viva Energy Australia", which splits one company across two operator values. A third appears as "PLUS ES Manag". I left them as published and flagged them, because silently "correcting" a publisher's data is how you end up with a dataset nobody can reconcile against the source.

4. HTTP 200 is not the same as good data

Two separate incidents, both silent:

• Queensland's CSV returns HTTP 403 with an HTML error page to any non-browser client. A lenient parser reads that as a zero-row CSV. The same file also has newlines inside quoted fields, so it holds 17 records where wc -l claims 34.
• When the primary OpenStreetMap endpoint timed out, the fallback mirror returned a complete-looking 1,493 elements against the primary's 1,590. Nothing truncated, no warning: the mirror's database was simply 43 days stale. A silent 6% coverage loss, small enough to slip under any record-count threshold.

5. Publishers state totals: use them

The single most useful trick I found. Both the ACT and Tasmanian pages state their own figures: "74 public EV chargers with 131 charging bays", and a Total row of $567,000, so the parse can reconcile against them and refuse the page if it doesn't add up. That converts a fragile HTML scrape into one that fails loudly when the page is restructured, instead of quietly returning fewer rows.

// Tasmania: three tables, three stated grant totals, all reconciling
Chargesmart 3 2025    parsed $567000  stated $567000  OK
Fast charging         parsed $710500  stated $710500  OK
Destination charging  parsed $62500   stated $62500   OK

6. Some coordinates are not locations

Tasmania publishes a town name and nothing finer. Geocoding "Miena" gives you the centre of Miena, which for a charger is kilometres out. So precision became a graded property rather than a boolean: surveyed (~10 m), geocoded_address (~100 m, the ACT), and geocoded_locality (~5 km, Tasmania). Town-level records are excluded from the map by default, never merged with anything, never returned as "nearest", and confidence-penalised, while staying fully searchable, because "is there a funded charger in Miena?" is a real question nothing else answers.

7. Stale and rebranded links inside the data

While verifying operator websites I found OpenStreetMap still carrying goevie.com.au for Evie Networks, whose site is now evie.com.au. EVNet's domain redirects to chargengo.au, a rebrand. Electric Highway Tasmania's domain now lands on an NRMA page. Any URL you inherit from crowd-sourced data needs checking before you put it in front of a user.

8. "Operator" is four different concepts

214 distinct values, 143 of them appearing once. "Non-networked" (218 sites) is a TfNSW placeholder, not a company. "ChargeHub" (12 sites) and "Charge Hub" (7) may or may not be the same business: I could not resolve it, so I left them separate and said so. Same for "EVE Australia" and "EVSE". Alias folding fixes the obvious cases ("BP" and "BP Australia"; "Tesla" and "Tesla Motors") but you cannot fold what you cannot verify.

9. The bugs that were mine

Worth separating out, because they were the most instructive failures and none of them were the publishers' doing.

What I'd tell you before you start

1. Settle the licence question first. It's the one decision that is expensive to reverse, and it is a product decision disguised as a footnote.
2. Budget real effort for reconciliation. NSW alone has more records than all-of-Australia OpenStreetMap, and neither is a superset. Concatenate them and you ship doubled pins; round coordinates to deduplicate and you merge distinct chargers in one car park. This is where the perceived quality of the whole thing lives.
3. Never invent a value. An unparseable rating stays null. A missing fee is unknown, not free. A town name is not a location. Guessing corrupts silently and is invisible in testing.
4. Look for an invariant the publisher already asserts: a stated total, count or date you can reconcile against. It's the cheapest possible way to make a scrape trustworthy.
5. Validate before you cache, or an error page becomes your dataset.
6. Send corrections upstream. Fixes belong in OpenStreetMap, where they benefit everyone and flow back on your next ingest. A private correction store forks from your sources and rots.
7. Say what you don't know. No pricing, no live availability, roughly one site in five missing entirely, and four jurisdictions where every pin came from a volunteer. A map that admits this is more useful than one implying uniform authority.