About the Data

Each of the nine reciprocal networks publishes its own member list — typically as a PDF or on a web page — on its own website, on its own schedule. The Museum Rover dataset is built by pulling all nine of these lists, cleaning them, and merging them into a single unified table.

The raw data, as published by the networks to their members, comes in the form of PDF directories, web pages, and downloadable brochures. Nine different formats, nine different naming conventions, nine different ideas about what constitutes an address. None of it is structured. None of it is queryable. None of it talks to each other.

Turning this into a single, clean, mappable database — the kind of thing you can ask “what can I visit for free?” and get a real answer — is the kind of work that would have required a small office army in the old days. Data entry clerks, spreadsheet jockeys, someone to call every museum and verify the spelling. Producing a tool like this was historically the domain of companies that wanted to sell you a subscription, not an enthusiast who just wanted the information to be accessible.

But we live in the age of AI and language models that can process large, messy datasets. That changed the math. One person can now do what used to take a team — though “can” doesn't mean “easy.”

Each network publishes its member list differently. NARM is a dense 7-page PDF with 1,525 institutions. ROAM is a 13-page PDF. AZA is a 5-page reciprocity chart with tier information embedded in the layout. Time Travelers is 29 pages with per-institution benefit codes. MARP and SERM are web pages. None of these are tables — they're formatted for humans to read on paper, not for machines to parse.

The PDFs are the hard part. Institution names, cities, and states are extracted from running text. Some networks include extra information encoded in symbols that had to be decoded:

The real work begins when you try to merge nine lists into one. The same institution can appear in multiple networks under different names, different abbreviations, different punctuation, or different corporate identities.

“The Cleveland Museum of Art” in one list is “Cleveland Museum of Art” in another. “The New York Botanical Garden” vs “New York Botanical Garden, The.” “Assoc. of Science & Technology Centers” vs “Association of Science and Technology Centers.” Ampersands vs. “and.” “Mus.” vs “Museum.” “Ctr” vs “Center.” “St.” vs “Saint.”

Some institutions have rebranded between network list publications — one network still uses the old name, another has the new one. Some umbrella organizations (like the Wisconsin Historical Society) list a dozen individual historic sites under the parent name in one network but as separate entries in another.

The first pass used automated normalization to catch the mechanical differences — lowercasing, expanding abbreviations, collapsing whitespace. The second pass was slower: reviewing hundreds of flagged potential matches by hand. Is “Gilcrease Museum” in Tulsa the same institution as “Thomas Gilcrease Institute of American History and Art” in Tulsa? (Yes.) Is “Children's Museum of Indianapolis” the same as “The Children's Museum of Indianapolis”? (Yes, but only after checking it's the same city and not a different children's museum.) 106 true duplicates were confirmed and merged. Along the way, 19 Massachusetts entries in the ROAM data turned out to have corrupted city names — a systematic glitch from PDF text extraction that prepended “tt” to every city (e.g. “ttCambridge” instead of “Cambridge”).

After merging, every institution in the dataset was checked against known facts — verifying that the name, city, and state were correct and the institution actually exists where the source list says it does. 59 errors were found and corrected.

Some were subtle: the Wexner Center for the Arts was listed under Columbus, GA instead of Columbus, OH — wrong Columbus. The Detroit Zoo's address is technically Royal Oak, MI, not Detroit. Some were serious: the Country Doctors Museum was filed under Winterville, GA when it's actually in Bailey, NC — entirely wrong state. An Australian science center was accidentally classified as being in the USA. And some were just messy: phone number fragments jammed into institution names (a PDF extraction artifact), systematic misspellings across lists (“Musuem,” “Sommerville,” “Perrysberg”), and one institution whose name included half of the next institution's address.

Every correction is documented. The goal isn't perfection — it's transparency about where the data came from and what was done to clean it.

Every institution needed map coordinates. The Google Maps Geocoding API was used with three fallback strategies: institution name + city + state (most precise), then name + city, then city + state centroid as a last resort. In the end, 3,113 of 3,114 institutions were geocoded to a precise address — essentially 100%. The one holdout turned out to be an online-only museum. Would you have believed such a thing existed before 2020? Turns out we did pretty good.

Geocoding also revealed another round of duplicates. Institutions that resolved to the same coordinates (within ~100 meters) were grouped and reviewed. Of 123 coordinate groups, 94 turned out to be the same institution listed under different names in different networks — they got merged. The remaining 29 were legitimately separate institutions that share a building or campus: the Smithsonian American Art Museum and National Portrait Gallery (same building, two distinct collections), three separate museums inside Casa de Balboa in San Diego's Balboa Park, the Yale Center for British Art and Yale University Art Gallery (adjacent buildings, entirely different collections).

This is the Spring 2026 dataset. Networks update their member lists on their own schedules — some annually, some quarterly, some whenever they get around to it. Institutions join and leave. New institutions open. Museums merge, rebrand, or close.

Much of the pipeline can probably be automated — reprocessing fresh published documents against the existing database, flagging what changed, and only requiring human review for the ambiguous cases. That's the next step. A full refresh shouldn't take days once the tooling matures.

The dream scenario: the networks themselves — or better yet, individual museums — push their updates directly into the database. We'd welcome that. All they'd have to do is spell their institution the same way twice. (Based on the cross-referencing experience, this may be the hardest part.)

In the meantime, the pipeline exists, every step is documented, and the data gets refreshed when the source lists change. If you notice an institution that's missing, miscategorized, or pinned to the wrong place on the map, we want to know.

This dataset is as accurate as we could make it, given what the source documents contain. But there are real limits to what any dataset built from published PDFs can capture. Here's what you should know.

To the best of our ability, this list is accurate as of the dates shown above. If you find an error — a museum in the wrong place, a network membership that's out of date, a museum that closed — we want to know about it.