GeoCities Deep Archeology
Recovering the lost geography of the early web. Written up at tentcities.us.
What it is
GeoCities Deep Archeology works through what survived of GeoCities, from the surface pages down to the archives crawlers never saw. It recovers what the first rescue missed, unpacks the files people parked in their accounts, and measures which sites survive whole enough to stand back up. The findings, case studies and methodology are published at tentcities.us.
Why it matters
The early web was mostly personal. Personal sites, one person one URL, no algorithm deciding who saw them. When GeoCities went offline, most of it went with it. What survived is scattered, half-broken, and in many cases still encrypted or malformed. This is the slow, unglamorous work of putting it back into a shape you can actually browse.
It also earns its way onto the InertSpace map because the shape it's rebuilding, personal sites arranged into neighborhoods, is the same shape TentCities is building. Different era, same idea.
Current state
Where it is right now: the work is written up at tentcities.us: a findings page, a methodology, and a dozen case studies. A pull from the Wayback record has brought back 4.84 million GeoCities pages and 7.08 million distinct images. 71,892 parked archives have been unpacked, surfacing files nobody had seen since they were uploaded. Of 483,485 sites scored for completeness, 39,172 come back whole: every internal link resolves and every image is present.
The finding underneath it all: the early web was built out of self-expression, not commerce. The single most-copied file in the corpus is a webring badge, not an ad.
Where the deeper story lives: tentcities.us. This dive page keeps the shape of the arc; that site is where the actual finds and the ongoing work get written up.
What is next: more recovery and more write-ups. A browsable archive is not up yet.
A note on the domain: tentcities.us is currently home to these write-ups. The domain is also being kept warm for a possible future TentCities onboarding layer that would connect AnalogCirc.us and other community servers. That is far off; for now the two live together comfortably.
Connections
Independent of the rest of the constellation. Shares a domain with the future TentCities onboarding layer. Owned by the Cooperative.
Build log
A long-arc archaeology project. The touch points below are the ones worth pointing at; the deeper story lives on tentcities.us.
-
Early years
The pile itself
Years of downloading, reviewing, and hacking through the surviving GeoCities archive. Multiple systems tried and thrown out. Tooling that eventually became the framework that runs the site today was written and rewritten many times before any of it stabilized.
-
Corralled
641 GB pulled down, 316 GB YAHOOIDS, 3.9M HTML files inventoried
The Archive Team torrent is down and unpacked. Everything downstream is transforming that pile into something a person can actually browse.
More
- 641 GB is the raw torrent size: what Archive Team preserved when GeoCities went dark.
- 316 GB of that is the YAHOOIDS directory (roughly 11 million files), the biggest single lump, and the one with the most surprises inside.
- 3.9M HTML files is what actually gets processed. Everything else (images, media, junk) is downstream of the HTML being coherent first.
-
Rewritten
1M+ pages rewritten, 2M+ links re-mapped
The rewriter chews through the pile at roughly 18,000 files per second. Broken links re-mapped, mangled encodings cleaned up, ~30 million Yahoo tracking elements pulled out along the way.
More
- 1,006,416 pages rewritten in the primary pass: a full-set edit that makes them render in a modern browser without eating each other.
- 1,997,135 links re-mapped. Old GeoCities URLs are rewritten to their new positions in the archive so internal navigation actually goes somewhere.
- About 30 million Yahoo tracking elements pulled out along the way. The pages get a lot lighter, a lot faster, and the ghost of Yahoo's ad stack stops staring back.
- Character encodings (a graveyard of half-declared, half-broken, half-cp1252 pages) normalized to UTF-8.
-
Deduped
2.3M images → 1.7M unique, converted to JPEG XL
Perceptual hashing collapses roughly a quarter of the image mass as duplicates. The 1.7M unique images that remain get converted to JPEG XL (32 GB total). Animated GIFs are preserved as-is.
More
- Perceptual hashing rather than byte-hashing: a slightly-recompressed dancing baby is the same dancing baby.
- 2,290,606 images scanned. 1,721,195 unique perceptual hashes retained. About 25% dedup rate.
- 1,709,718 converted to JPEG XL for a modern format that preserves quality at a fraction of the storage. Total footprint ends up around 32 GB.
- 77,836 animated GIFs get preserved as-is. Animation is culture; you don't reformat a dancing hamster.
-
Graphed + restored
Link graph built, Flash and MIDI put back on ~120k pages
Nearly 7 million internal links and 13 million external links across almost a million unique domains, all mapped. About 15,000 pages get Flash restored via a modern runtime; about 105,000 have their MIDI playback put back.
More
- Internal graph: 6,989,313 edges. This is the map that tells you what a neighborhood was actually shaped like: who linked to whom, where webrings ran.
- External graph: 13,218,962 edges pointing out at 996,440 unique domains. GeoCities in context: where it thought the rest of the internet was.
- Flash: ~15,000 pages had Flash content. Ruffle (a Rust-based Flash runtime) gets injected so a modern browser can still play what was there.
- MIDI: ~105,000 pages had a background MIDI track. Restored with a modern player so the ambient soundtrack of the early web comes back.
-
Publishing surface up
tentcities.us stood up as the front door
A dedicated hosting layer at tentcities.us takes over as the public face of the archive. Phase 1 is the current static surface; the richer browsing experience comes with the next two phases.
-
Aug 22, 2026
The archive starts saying what it found
tentcities.us gains a written layer: case studies drawn out of the corpus, plus a findings page, a methodology write-up, and a plain statement of what the project is for.
More
- The case studies are the point. A pile of three million pages is a statistic; one webring reconstructed, or one page traced back to why it looked the way it did, is something a person can actually read.
- Methodology is published alongside the findings, because an archive that will not show its working is asking to be taken on faith. Anyone should be able to see how a number was arrived at before believing it.
- This is the deliberate split: the dive page you are reading keeps the shape of the arc, and the archive site carries the actual finds. Writing up what was found is the work now, not extracting more of it.
-
Sep 2026 You are here
Recovery at scale, and sites scored for completeness
A pull from the Wayback record brings back 4.84 million GeoCities pages and 7.08 million distinct images. 71,892 parked archives are unpacked. Of 483,485 sites scored for completeness, 39,172 come back whole: every internal link resolves and every image is present.
-
Next
More recovery, more write-ups
A browsable archive comes after the data is connected, not before.