The brief
Two open datasets describe the world’s points of interest and neither is sufficient alone. OpenStreetMap is community-verified and often hand-tagged, with strong geographic fidelity and real gaps in commercial coverage. Overture Maps, backed by Meta, Microsoft, Amazon and TomTom, has commercial-scale breadth and lacks the granular tagging and human verification that makes OSM trustworthy.
Anyone who needs both ends up doing the same work: pull two extracts, discover they disagree about where things are and what to call them, write matching logic, and produce a merged file that nobody else can reproduce or check. The result is a private dataset with private assumptions, rebuilt from scratch by the next person who needs one.
AnythingPOI does that fusion once, in the open, and publishes both the output and the pipeline. 22.7 million records across six countries, every one deduplicated, classified against a shared taxonomy, and carrying a confidence score you can interrogate.
The insight
The useful decision was to make the pipeline conservative and then say so loudly, rather than optimising for an impressive-looking match rate.
Conflation is where this kind of project usually oversells. Loosen the thresholds and the merged count climbs, which reads as success and quietly fuses two different businesses that happen to share a name and a street. Tighten them and the match rate looks unimpressive. AnythingPOI takes the second path: a hard 50 m gate, required category agreement, and 0.85 name similarity, all three of which must hold. The result is that cross-source matches run between 1.4 and 3.2 percent by country.
That number is low and it is honest. A match here means two independent datasets agree, which is why dual-source corroboration is worth something in the confidence score at all. Inflating it would have destroyed the signal it exists to provide.
The per-country pattern is itself informative:
| Country | Total POIs | Conflated | OSM share |
|---|---|---|---|
| Germany | 6,763,796 | 3.2% | 35.8% |
| Canada | 5,565,256 | 1.4% | 8.1% |
| United Kingdom | 4,622,174 | 2.7% | 20.5% |
| Türkiye | 2,201,304 | 1.4% | 13.9% |
| Netherlands | 1,782,538 | 2.8% | 15.8% |
| Australia | 1,735,980 | 3.0% | 18.4% |
Germany and the UK carry the richest community mapping and the highest conflation rates, which is what you would expect and is worth being able to demonstrate rather than assume.
The second decision was to score every record individually instead of publishing a clean subset. A dataset that silently drops uncertain records looks better and tells the user less. Scoring keeps everything and hands the filtering decision to whoever knows their own tolerance.
What it does not do
Each release is a snapshot. OSM and Overture are live, and the fused output describes the moment the pipeline ran, which is why releases are versioned and citable rather than continuously updated.
Conflation recall is the known weakness, and it follows directly from the conservative thresholds. Real duplicates with transliterated names, address-only correspondences, or a large polygon whose centroid sits far from its entrance will not be caught. Recall is the price of precision here, paid deliberately.
Every output record is a point. A park or a shopping centre becomes its centroid, which is wrong in a way that matters for anything measuring area or containment. Address completeness varies enormously by country and category: around 42 percent street coverage for Parks and Nature against roughly 97 percent for Healthcare in Canada. And names are stored exactly as they appear in source data, with no transliteration or translation, so cross-script querying is left to the user.
