Gazetteer of the World (1856) ⓘ

👋 Welcome! This is a project of the World Historical Gazetteer (WHG) — an interactive map and reader of A Gazetteer of the World, a seven-volume reference book published by A. Fullarton & Co. in 1856, describing tens of thousands of places across the globe as they were known in the mid-19th century. The title page credited only “a Member of the Royal Geographical Society”; the editor is now identified as George Godfrey Cunningham, a partner in the firm.

A gazetteer is a geographical dictionary: an alphabetical list of places, each with a short description — where it is, what kind of place it is, its population, nearby features, and so on. We have turned the original printed pages into data you can search, map, and read.

What you can do here

• Explore the map — every marker is a place from the gazetteer, shown where we think it is today.
• Search — type a place name; old spellings and typos are tolerated.
• Read — open a volume and read the original entries, with the historical plates (engravings and maps) shown in place.
• Check the source — every entry links to the scanned page it came from.

How it was made

Nobody typed this in by hand — every step is automatic, and uses freely-available AI models run on our own computers (no per-use fees, nothing sent to outside services):

1. Reading the pages. We start from public-domain photographic scans of the book and turn the printed words back into text with OCR (“optical character recognition”), using the layout-aware Surya model.
2. Understanding each entry. A large language model (“LLM” — the kind of AI behind chatbots) reads each entry and pulls out structured facts: the place name, the type of place, its country, coordinates, population, and so on. We use the open Llama 3.3, with gpt-oss double-checking its work and Qwen3 repairing the cases it flags.
3. Reading the tables and pictures. The statistical tables and the engraved plates (town views and maps) are read by a vision model, Qwen2.5-VL.
4. Linking to a global index. We then try to match each place to the World Historical Gazetteer (WHG) — a free academic project that gathers historical places from many sources into one searchable index. A match gives the 1856 place a modern location, and sometimes a boundary outline. (This matching step is called reconciliation.)

Each place is also given a feature type (city, river, mountain, ruin…) chosen from the Getty Art & Architecture Thesaurus (AAT), a shared published vocabulary in which every term has a stable web address. Using such shared identifiers instead of free-text labels is what lets this data be linked to other datasets — the idea behind Linked Open Data — rather than sitting in an isolated silo.

All of this is very much a work in progress. The typing and fact-extraction are only as good as the instructions given to the language model, and can be improved simply by refining that prompt and re-running the extraction.

Over 116,000 places, most linked to a modern location, plus 1,774 statistical tables and 133 historical plates — all in a site that runs entirely in your browser, with no server doing the work as you click.

⚠️ How much to trust this

Please treat this as a research demonstrator, not an authority. Because every step is automatic:

• Many matches are wrong. An 1856 name may be linked to a same-named place somewhere else, or to nothing at all. Matching historical names automatically is genuinely hard, and never wholly perfect — but it is an active focus of the WHG, and it steadily improves as more place data is folded into its indices.
• Locations can be off, and some places are shown only by their printed coordinates, with no match found.
• Tables and illustrations are detected unevenly — some are missed or mis-read.

So take everything here as a suggestion to check against the original scanned page (linked from every record), not as fact. Spotted an error? The ⚑ Report link in any place popup or reader entry takes a few seconds and genuinely helps improve the data.

Source & credits

The original book is public domain (page scans via HathiTrust). The project grew out of Humphrey Southall’s offer of his transcripts to the WHG and the discussion that followed, though it uses its own fresh OCR rather than those transcripts. The heavy computation ran on the HTC and H2P clusters at the University of Pittsburgh Center for Research Computing and Data (RRID:SCR_022735), supported by NIH award S10OD028483 and NSF award OAC-2117681. The code and full method are on GitHub.