Where to look, in order
- The primary publisher. National statistics offices, central banks, international organisations (World Bank, IMF, OECD, WHO, UN) and government open-data portals. Primary sources document their methods and publish revisions.
- Curated secondary sources. Organisations such as Our World in Data or Gapminder harmonise many primary sources into consistent long-run series, and they document what they changed.
- Research repositories. Zenodo, Dryad, Harvard Dataverse and ICPSR host data behind published research, often with a DOI for citation.
- Community platforms. Kaggle and Hugging Face host thousands of datasets of varying quality. Excellent for practice and machine learning, but check provenance carefully.
- Dataset search engines. Google Dataset Search indexes dataset descriptions across the web. Use it to discover, then go to the original publisher.
Our open datasets directory lists 40+ of the most useful sources with formats and licence notes.
The evaluation checklist
Provenance: who made it and how?
Can you name the organisation that collected the data and find its methodology? Survey data needs a sampling method and sample size. Administrative data needs the rules for what gets recorded. Scraped data needs to say what was scraped and when. If you can't trace a dataset to a responsible publisher, treat it as unverified.
Definitions: does it measure what you think?
"Unemployment", "poverty" and "active user" all have competing definitions. Read the indicator notes. Two countries reporting the "same" metric may count it differently, and definitions change over time, creating breaks in series.
Coverage: what's missing?
Which years, regions and groups are included? Are some countries missing in some years? Are small businesses excluded? Gaps aren't errors, but conclusions shouldn't extend beyond what's covered.
Timeliness and revisions
Official statistics are often revised as better information arrives. GDP figures, for example, can change noticeably after first release. Note the release or download date, and expect figures to change.
Granularity
Annual national averages can't answer questions about monthly city-level patterns. Check the time and geographic resolution before you invest effort.
Quality signals
Profile the file before analysis: missing values, duplicates, impossible values (negative ages, percentages over 100), sudden jumps. Our Data Profiler does this in seconds.
Licences in plain English
| Licence | What it generally allows | Main obligation |
|---|---|---|
| Public domain / CC0 | Any use, including commercial | None (attribution still good practice) |
| CC BY 4.0 | Any use, including commercial | Credit the source |
| Open Government Licences (UK, Canada, etc.) | Broad reuse | Attribution statement |
| ODbL (e.g. OpenStreetMap) | Use and adapt | Attribution, and share derived databases under ODbL |
| CC BY-NC | Non-commercial use only | Credit, no commercial use |
| Custom terms | Varies | Read them, especially for commercial products |
This is general information, not legal advice. When building a commercial product on someone else's data, read the full terms.
How to cite a dataset
A good citation lets someone else find exactly what you used: Publisher (year). Dataset title, version or release. URL or DOI. Accessed date. For example: World Bank (2026). World Development Indicators. https://data.worldbank.org. Accessed 17 November 2026. Put the source line directly under any chart you publish. It's also one of the judging criteria in our Data Viz Challenge.