Book Archive Comprehensiveness

anna's archive has 63 million books and the internet archive 25 million books is this pretty comprhenesive ?

I’ll sanity-check the headline numbers and then separate “comprehensive for published books” from “comprehensive for usable full text,” because those are very different.

Pretty comprehensive, but not “the world’s books.”

The useful framing:

1. It is probably one of the most comprehensive downloadable / full-text-ish book corpora.
Anna’s Archive claims about 63.6M books and 95.7M papers, with torrents totaling about 1.1 PB. citeturn942154search8 Internet Archive/Open Library is huge too, but more legally constrained and uneven for copyrighted books.

2. But it is not comprehensive relative to all bibliographic records.
WorldCat reports about 609.8M bibliographic records as of Oct. 2025, though that includes journals, media, archives, editions, duplicates, etc., not just distinct books. citeturn942154search2 Google estimated in 2010 that there were about 129.9M distinct book titles globally; Google Books itself says it has over 40M books in 500+ languages. citeturn942154search3turn942154search7

3. The 63M + 25M numbers are not additive.
There is massive overlap: Anna’s Archive indexes/mirrors from sources like Library Genesis, Sci-Hub, Z-Library-style collections, Internet Archive/Open Library, etc. So “63M + 25M = 88M” is wrong. It is more like overlapping reservoirs.

4. It will be biased.
Strong coverage: English, Russian, European languages, academic books, textbooks, PDFs/EPUBs that circulated online, public-domain scans, popular pirated works.
Weak coverage: obscure local-language books, small presses, newly released commercial books, children’s books, art books, scans with images/plates, archival manuscripts, corporate/government gray literature, books never digitized, and anything locked inside publisher platforms.

5. For LLM training, the bottleneck is not just count.
You care about: deduplication, OCR quality, edition/version duplication, corrupted EPUBs/PDFs, metadata quality, legal exposure, language balance, and whether images/tables/footnotes survive extraction.

So: yes, it is “comprehensive” in the practical pirate-library sense. No, it is not comprehensive in the universal bibliography sense. For training/research, it would be a giant book corpus, probably enough to cover an enormous amount of mainstream written culture, but it would still miss long-tail, recent, non-Western, visual, archival, and legally gated material.