With AI assistants
Your assistant, with the European used-car market in it
Connect this service and an assistant stops guessing what a car is worth abroad. It reads the same measured figures this site publishes — and repeats the limits with them.
> Where in Europe is a 2018-2022 BMW 3 Series cheapest?
⏺ auto-market · corridor_scan · BMW 3 Series 2018-2022
| Market | Median asking | Median odometer | Adverts |
|---|---|---|---|
| Lithuania | EUR 19,750 | 152,000 km | 290–300 |
| Latvia | EUR 20,000 | 186,000 km | 110–120 |
| Poland | EUR 20,750 | 123,000 km | 906 |
| … 15 more markets | |||
| France | EUR 28,000 | 88,000 km | 965 |
| Netherlands | EUR 28,000 | 108,000 km | 1,382 |
| Norway | EUR 33,000 | 102,000 km | 260–270 |
Read 21 markets. EUR 13,250 between Lithuania and Norway — and that is not a margin. The cheap end also reads 50,000 km further on the odometer, and registration tax, transport and VAT are all still to come. Asking prices, never transaction prices.
Measured across 4,267,191 live adverts in 23 national markets, release 2026-09-11. These are the figures this site publishes free — see Lithuania.
Why an assistant gets this wrong on its own
Nobody publishes cross-border used-car aggregates. So an assistant with web search does the only thing it can: it reaches a few dozen adverts it happens to be able to open, and reports what they say.
That is a sample. It has no median, no sample size, no odometer comparison and no idea which adverts are already sold. It will still answer in a confident sentence, because the shape of the answer does not depend on the strength of the evidence.
The same question is answered from every live advert in 23 markets, with the sample size attached to each figure and the caveats attached to the answer.
The gap between a handful of adverts and a market is the whole product. An assistant cannot close it by reasoning harder, and neither can a longer search.
Four things worth asking it
Find a corridor worth driving
Compare this specification across every market that publishes it. Give me the three widest gaps against the cheapest market, and tell me for each whether the mileage is comparable.
The mileage caveat is the point. A spread between a market quoting 265,000 km and one quoting 159,000 km is not an opportunity, and the answer says so in a field the assistant can branch on rather than in prose it might skip.
Sanity-check a car in front of you
A dealer is asking EUR 18,000 for this car in Poland. Where does that sit against the Polish market — above or below the median, and inside or outside the middle half?
Inside or outside the middle half is what separates a keen price from a normal one. A median alone cannot answer it, which is why every market carries its interquartile range wherever the sample is large enough to publish one.
Pick the vintage, not just the market
Show me the depreciation curve and find the steepest fall between neighbouring model-year bands. I want the band just below that drop — what does it ask, and how many cars is that figure built on?
The steepest fall is measured only between abutting bands, so a gap in the published years is never reported as a one-step collapse. A dramatic fall over forty cars is a different claim from the same fall over four hundred.
Build a buying list for the week
For each specification I trade, find the cheapest market and the spread, skip any where only one market publishes, and rank by spread as a share of the cheapest price. Check my coverage budget first.
Share rather than absolute euros, because a EUR 3,000 gap on a EUR 9,000 car is a trade and the same gap on a EUR 60,000 car is noise. The budget check matters too: the assistant can find out the sweep will not fit before it starts.
What it will not do
This service publishes aggregates and never rows. An assistant connected to it cannot be talked into an advert, a seller, an individual price or a link to a marketplace, because none of those exists anywhere behind the connection — not as a permission it lacks, but as data the published artefact does not contain.
- Figures are asking prices, never transaction prices. A gap between two markets is not a margin.
- Landed cost is not modelled. Registration tax is published per market, with its source and the date it was verified; transport and VAT applicability are not.
- Prices are snapped and small samples are withheld, so a figure is a published cell rather than a raw measurement. How these figures are made says exactly how.
Connect it in about a minute
It is an MCP server with an OAuth sign-in, so most clients take a URL and handle the rest. The setup page has the connector URL, the commands for Claude Code and Codex, the menu path for Claude and ChatGPT, and what to do if your client only speaks stdio.
Set it up