Internet Search Tips

A description of advanced tips and tricks for effective Internet research of papers/books.
archiving⁠, technology⁠, shell⁠, Google⁠, tutorial
2018-12-112021-04-18 finished certainty: certain importance: 4

Over time, I developed a certain google-fu and expertise in finding references, papers, and books online. Some of these tricks are not well-known, like checking the Internet Archive (IA) for books. I try to write down my search workflow, and give general advice about finding and hosting documents, with demonstration case studies⁠.

Google-fu search skill is something I’ve prided myself ever since elementary school, when the librarian challenged the class to find things in the almanac; not infrequently, I’d win. And I can still remember the exact moment it dawned on me in high school that much of the rest of my life would be spent dealing with searches, paywalls, and broken links. The Internet is the greatest almanac of all, and to the curious, a never-ending cornucopia, so I am sad to see many fail to find things after a cursory search—or not look at all. For most people, if it’s not the first hit in Google/Google Scholar, it doesn’t exist. Below, I reveal my best Internet search tricks and try to provide a rough flowchart of how to go about an online search, explaining the subtle tricks and intuition of search-fu.



Human flesh search engine. Last resort: if none of this works, there are a few places online you can request a copy (however, they will usually fail if you have exhausted all previous avenues):

Finally, you can always try to contact the author. This only occasionally works for the papers I have the hardest time with, since they tend to be old ones where the author is dead or unreachable—any author publishing a paper since 1990 will usually have been digitized somewhere—but it’s easy to try.


After finding a fulltext copy, you should find a reliable long-term link/place to store it and make it more findable (remember—if it’s not in Google/Google Scholar, it doesn’t exist!):

  • Never Link Unreliable Hosts:

    • LG/SH: Always operate under the assumption they could be gone tomorrow. (As my uncle found out with shortly after paying for a lifetime membership!) There are no guarantees either one will be around for long under their legal assaults or the behind-the-scenes dramas, and no guarantee that they are being properly mirrored or will be restored elsewhere. Download anything you need and keep a copy of it yourself and, ideally, host it publicly.
    • NBER: never rely on a or URL, as they are temporary. (SSRN is also undesirable due to making it increasingly difficult to download, but it is at least reliable.)
    • Scribd: never link Scribd—they are a scummy website which impede downloads, and anything on Scribd usually first appeared elsewhere anyway. (In fact, if you run into anything vaguely useful-looking which exists only on Scribd, you’ll do humanity a service if you copy it elsewhere just in case.)
    • RG: avoid linking to (compromised by new ownership & PDFs get deleted routinely, apparently often by authors) or (the URLs are one-time and break)
    • high-impact journals: be careful linking to or Cell (if a paper is not explicitly marked as Open Access, even if it’s available, it may disappear in a few months!); similarly, watch out for,,,,, &, who pull similar shenanigans.
    • ~/: be careful linking to academic personal directories on university websites (often noticeable by the Unix convention .edu/~user/ or by directories suggestive of ephemeral hosting, like .edu/cs/course112/readings/foo.pdf); they have short half-lives.
  • PDF Editing: if a scan, it may be worth editing the PDF to crop the edges, threshold to binarize it (which, for a bad grayscale or color scan, can drastically reduce filesize while increasing readability), and OCR it.

    I use gscan2pdf but there are alternatives worth checking out.

  • Check & Improve Metadata.

    Adding metadata to papers/books is a good idea because it makes the file findable in G/GS (if it’s not online, does it really exist?) and helps you if you decide to use bibliographic software like in the future. Many academic publishers & LG are terrible about metadata, and will not include even title/author/DOI/year.

    PDFs can be easily annotated with metadata using : : exiftool -All prints all metadata, and the metadata can be set individually using similar fields.

    For papers hidden inside volumes or other files, you should extract the relevant page range to create a single relevant file. (For extraction of PDF page-ranges, I use ⁠, eg: pdftk 2010-davidson-wellplayed10-videogamesvaluemeaning.pdf cat 180-196 output 2009-fortugno.pdf. Many publishers insert a spam page as the first page. You can drop that easily with pdftk INPUT.pdf cat 2-end output OUTPUT.pdf, but note that PDFtk may drop all metadata, so do that before adding any metadata. To delete pseudo-encryption or ‘passworded’ PDFs, do pdftk INPUT.pdf input_pw output OUTPUT.pdf; PDFs using actual encryption are trickier but can often be beaten by off-the-shelf password-cracking utilities.)

    I try to set at least title/author/DOI/year/subject, and stuff any additional topics & bibliographic information into the “Keywords” field. Example of setting metadata:

    exiftool -Author="Frank P. Ramsey" -Date=1930 -Title="On a Problem of Formal Logic" -DOI="10.1112/plms/s2-30.1.264" \
        -Subject="mathematics" -Keywords="Ramsey theory, Ramsey's theorem, combinatorics, mathematical logic, decidability, \
        first-order logic,  Bernays-Schönfinkel-Ramsey class of first-order logic, _Proceedings of the London Mathematical \
        Society_, Volume s2-30, Issue 1, 1930-01-01, pg264-286" 1930-ramsey.pdf
  • Public Hosting: if possible, host a public copy; especially if it was very difficult to find, even if it was useless, it should be hosted. The life you save may be your own.

  • Link On WP/Social Media: for bonus points, link it in appropriate places on Wikipedia or Reddit or Twitter; this makes people aware of the copy being available, and also supercharges visibility in search engines.

  • Link Specific Pages: as noted before, you can link a specific page by adding #page=N to the URL. Linking the relevant page is helpful to readers.


Aside from the (highly-recommended) use of hotkeys and Booleans for searches, there are a few useful tools for the researcher, which while expensive initially, can pay off in the long-term:

  • archiver-bot: automatically archive your web browsing and/or links from arbitrary websites to forestall linkrot; particularly useful for detecting & recovering from dead PDF links

  • Subscriptions like PubMed & GS search alerts: set up alerts for a specific search query, or for new citations of a specific paper. ( is not as useful as it seems.)

    1. PubMed has straightforward conversion of search queries into alerts: “Create alert” below the search bar. (Given the volume of PubMed indexing, I recommend carefully tailoring your search to be as narrow as possible, or else your alerts may overwhelm you.)
    2. To create generic GS search query alert, simply use the “Create alert” on the sidebar for any search. To follow citations of a key paper, you must: 1. bring up the paper in GS; 2. click on “Cited by X”; 3. then use “Create alert” on the sidebar.
  • GCSE: a Google Custom Search Engines is a specialized search queries limited to whitelisted pages/domains etc (eg my Wikipedia-focused anime/manga CSE).

    A GCSE can be thought of as a saved search query on steroids. If you find yourself regularly including scores of the same domains in multiple searches search, or constantly blacklisting domains with -site: or using many negations to filter out common false positives, it may be time to set up a GCSE which does all that by default.

  • Clippings: like ⁠/ : regularly making and keeping excerpts creates a personalized search engine, in effect.

    This can be vital for refinding old things you read where the search terms are hopelessly generic or you can’t remember an exact quote or reference; it is one thing to search a keyword like “autism” in a few score thousand clippings, and another thing to search that in the entire Internet! (One can also reorganize or edit the notes to add in the keywords one is thinking of, to help with refinding.) I make heavy use of Evernote clipping and it is key to refinding my references.

  • Crawling Websites: sometimes having copies of whole websites might be useful, either for more flexible searching or for ensuring you have anything you might need in the future. (example: “Darknet Market Archives (2013–2015)”).

    Useful tools to know about: ⁠, ⁠, ⁠; Firefox plugins: NoScript⁠, uBlock origin⁠, Live HTTP Headers⁠, Bypass Paywalls⁠, cookie exporting.

    Short of downloading a website, it might also be useful to pre-emptively archive it by using linkchecker to crawl it, compile a list of all external & internal links, and store them for processing by another archival program (see Archiving URLs for examples). In certain rare circumstances, security tools like can be useful to examine a mysterious server in more detail: what web server and services does it run, what else might be on it (sometimes interesting things like old anonymous FTP servers turn up), has a website moved between IPs or servers, etc.

Web pages

With proper use of pre-emptive archiving tools like archiver-bot, fixing linkrot in one’s own pages is much easier, but that leaves other references. Searching for lost web pages is similar to searching for papers:

  • Just Search The Title: if the page title is given, search for the title.

    It is a good idea to include page titles in one’s own pages, as well as the URL, to help with future searches, since the URL may be meaningless gibberish on its own, and pre-emptive archiving can fail. HTML supports both alt and title parameters in link tags, and, in cases where displaying a title is not desirable (because the link is being used inline as part of normal hypertextual writing), titles can be included cleanly in Markdown documents like this: [inline text description](URL "Title").

  • Clean URLs: check the URL for weirdness or trailing garbage like ?rss=1 or ?utm_source=feedburner&utm_medium=feed&utm_campaign=Feed%3A+blogspot%2FgJZg+%28Google+AI+Blog%29? Or a variant domain, like a URL? Those are all less likely to be findable or archived than the canonical version.

  • Domain Site Search: restrict G search to the original domain with site:, or to related domains

  • Time-Limited Search: restrict G search to the original date-range/years

  • Switch Engines: try a different search engine: corpuses can vary, and in some cases G tries to be too smart for its own good when you need a literal search; (especially for ‘bang’ special searches), ⁠, and Yandex are usable alternatives

  • Check Archives: if nowhere on the clearnet, try the Internet Archive (IA) or the Memento meta-archive search engine:

    IA is the default backup for a dead URL. If IA doesn’t Just Work, there may be other versions in it:

    • misleading redirects: did the IA ‘helpfully’ redirect you to a much-later-in-time error page? Kill the redirect and check the earliest stored version for the exact URL rather than the redirect. Did the page initially load but then error out/redirect? Disable JS with NoScript and reload.

    • Within-Domain Archives: IA lets you list all URLs with any archived versions, by searching for URL/*; the list of available URLs may reveal an alternate newer/older URL. It can also be useful to filter by filetype or substring.

      For example, one might list all URLs in a domain, and if the list is too long and filled with garbage URLs, then using the “Filter results” incremental-search widget to search for “uploads/” on a WordPress blog.11

      Screenshot of an oft-overlooked feature of the Internet Archive: displaying all   for a specific domain, filtered down to a subset matching a string like *uploads/*.
      • wayback_machine_downloader (not to be confused with the internetarchive Python package which provides a CLI interface to uploading files) is a Ruby tool which lets you download whole domains from IA, which can be useful for running a local fulltext search using regexps (a good grep query is often enough), in cases where just looking at the URLs via URL/* is not helpful. (An alternative which might work is⁠.)


      gem install --user-install wayback_machine_downloader
      ~/.gem/ruby/2.5.0/bin/wayback_machine_downloader wayback_machine_downloader --all-timestamps ''
    • did the domain change, eg from to or Entirely different as far as IA is concerned.

    • does the internal evidence of the URL provide any hints? You can learn a lot from URLs just by paying attention and thinking about what each directory and argument means.

    • is this a Blogspot blog? Blogspot is uniquely horrible in that it has versions of each blog for every country domain: a blog could be under any of,,, foo.blogspot.jp12

    • did the website provide RSS feeds?

      A little known fact is that (GR; October 2005–July 2013) stored all RSS items it crawled, so if a website’s RSS feed was configured to include full items, the RSS feed history was an alternate mirror of the whole website, and since GR never removed RSS items, it was possible to retrieve pages or whole websites from it. GR has since closed down, sadly, but before it closed, downloaded a large fraction of GR’s historical RSS feeds, and those archives are now hosted on IA⁠. The catch is that they are stored in mega-⁠, which, for all their archival virtues, are not the most user-friendly format. The raw GR mega-WARCs are difficult enough to work with that I defer an example to the appendix⁠.

    • an IA-like mirror

    • any local archives, such as those made with my archiver-bot

    • Google Cache (GC): GC works, sometimes, but the copies are usually the worst around, ephemeral & cannot be relied upon. Google also appears to have been steadily deprecating GC over the years, as GC shows up less & less in search results. A last resort.



E-books are rarer and harder to get than papers, although the situation has improved vastly since the early 2000s. To search for books online:

  • More Straightforward: book searches tend to be faster and simpler than paper searches, and to require less cleverness in search query formulation, perhaps because they are rarer online, much larger, and have simpler titles, making it easier for search engines.

    Search G, not GS, for books:

    No Books in Google Scholar
    Book fulltexts usually don’t show up in GS (for unknown reasons). You need to check G when searching for books.

    To double-check, you can try a filetype:pdf search; then check LG. Typically, if the main title + author doesn’t turn it up, it’s not online. (In some cases, the author order is reversed, or the title:subtitle are reversed, and you can find a copy by tweaking your search, but these are rare.).

  • IA: the Internet Archive has many books scanned which do not appear easily in search results (poor SEO?).

    • If an IA hit pops up in a search, always check it; the OCR may offer hints as to where to find it. If you don’t find anything or the provided, try doing an IA site search in G (not the IA built-in search engine), eg book title

    • DRM workarounds: if it is on IA but the IA version is DRMed and is only available for “checkout”, you can jailbreak it.

      Check the book out for the full period, 14 days. Download the PDF (not EPUB) version to Adobe Digital Elements version ≤4.0 (which can be run in Wine on Linux), and then import it to with the De-DRM plugin⁠, which will produce a DRM-free PDF inside Calibre’s library. (Getting De-DRM running can be tricky, especially under Linux. I wound up having to edit some of the paths in the Python files to make them work with Wine.) You can then add metadata to the PDF & upload it to LG13⁠. (LG’s versions of books are usually better than the IA scans, but if they don’t exist, IA’s is better than nothing.)

  • : use the same PDF DRM as IA, can be broken same way

  • also hosts many book scans, which can be searched for clues or hints or jailbroken.

    HathiTrust blocks whole-book downloads but it’s easy to download each page in a loop and stitch them together, for example:

    for i in {1..151}
    do if [[ ! -s "$i.pdf" ]]; then
        wget ";orient=0;size=100;seq=$i;attachment=0" \
              -O "$i.pdf"
        sleep 10s
    pdftk *.pdf cat output 1957-super-scientificcareersandvocationaldevelopmenttheory.pdf
    exiftool -Title="Scientific Careers and Vocational Development Theory: A review, a critique and some recommendations" \
        -Date=1957 -Author="Donald E. Super, Paul B. Bachrach" -Subject="psychology" \
        -Keywords="Bureau Of Publications (Teachers College Columbia University), LCCCN: 57-12336, National Science Foundation, public domain, \;view=1up;seq=1" \

    Another example of this would be the Wellcome Library; while looking for An Investigation Into The Relation Between Intelligence And Inheritance, Lawrence 1931⁠, I came up dry until I checked one of the last search results, a “Wellcome Digital Library” hit⁠, on the slim off-chance that, like the occasional Chinese/Indian library website, it just might have fulltext. As it happens, it did—good news? Yes, but with a caveat: it provides no way to download the book! It provides OCR, metadata, and individual page-image downloads all under CC-BY-NC-SA (so no legal problems), but… not the book. (The OCR is also unnecessarily zipped, so that is why Google ranked the page so low and did not show any revealing excerpts from the OCR transcript: because it’s hidden in an opaque archive to save a few kilobytes while destroying SEO.) Examining the download URLs for the highest-resolution images, they follow an unfortunate schema:

    3. etc

    Instead of being sequentially numbered 1–90 or whatever, they all live under a unique hash or ID. Fortunately, one of the metadata files, the ‘manifest’ file, provides all of the hashes/IDs (but not the high-quality download URLs). Extracting the IDs from the manifest can be done with some quick sed & tr string processing, and fed into another short wget loop for download

    fgrep '@id' manifest\?manifest\=https\ | \
       sed -e 's/.*imageanno\/\(.*\)/\1/' | egrep -v '^ .*' | tr -d ',' | tr -d '"' # "
    # bf23642e-e89b-43a0-8736-f5c6c77c03c3
    # 334faf27-3ee1-4a63-92d9-b40d55ab72ad
    # 5c27d7de-6d55-473c-b3b2-6c74ac7a04c6
    # d514271c-b290-4ae8-bed7-fd30fb14d59e
    # f85ef645-ec96-4d5a-be4e-0a781f87b5e2
    # a2e1af25-5576-4101-abee-96bd7c237a4d
    # 6580e767-0d03-40a1-ab8b-e6a37abe849c
    # ca178578-81c9-4829-b912-97c957b668a3
    # 2bd8959d-5540-4f36-82d9-49658f67cff6
    # ...etc
    for HASH in $HASHES; do
        wget "$HASH/full/2212,/0/default.jpg" -O $I.jpg

    And then the 59MB of JPGs can be cleaned up as usual with gscan2pdf (empty pages deleted, tables rotated, cover page cropped, all other pages binarized), compressed/OCRed with ocrmypdf, and metadata set with exiftool, producing a readable, downloadable, highly-search-engine-friendly 1.8MB PDF.

  • remember the works for papers/books too:

    if you can find a copy to read, but cannot figure out how to download it directly because the site uses JS or complicated cookie authentication or other tricks, you can always exploit the ‘analogue hole’—fullscreen the book in high resolution & take screenshots of every page; then crop, OCR etc. This is tedious but it works. And if you take screenshots at sufficiently high resolution, there will be relatively little quality loss. (This works better for books that are scans than ones born-digital.)


Expensive but feasible. Books are something of a double-edged sword compared to papers/theses. On the one hand, books are much more often unavailable online, and must be bought offline, but at least you almost always can buy used books offline without much trouble (and often for <$10 total); on the other hand, while paper/theses are often available online, when one is not unavailable, it’s usually very unavailable, and you’re stuck (unless you have a university ILL department backing you up or are willing to travel to the few or only universities with paper or microfilm copies).

Purchasing from used book sellers:

  • Sellers:

    • used book search engines: Google Books/ a good starting point for seller links; if buying from a marketplace like AbeBooks/Amazon/Barnes & Noble, it’s worth searching the seller to see if they have their own website, which is potentially much cheaper. They may also have multiple editions in stock.

    • bad: eBay & Amazon are often bad, due to high-minimum-order+S&H and sellers on Amazon seem to assume Amazon buyers are easily rooked; but can be useful in providing metadata like page count or ISBN or variations on the title

    • good: AbeBooks⁠, Thrift Books⁠, Better World Books⁠, B&N⁠, Discover Books⁠.

      Note: on AbeBooks, international orders can be useful (especially for behavioral genetics or psychology books) but be careful of international orders with your credit card—many debit/credit cards will fail on international orders and trigger a fraud alert, and PayPal is not accepted.

  • Price Alerts: if a book is not available or too expensive, set price watches: AbeBooks supports email alerts on stored searches, and Amazon can be monitored via CamelCamelCamel (remember the CCC price alert you want is on the used third-party category, as new books are more expensive, less available, and unnecessary).


  • Destructive Vs Non-Destructive: the fundamental dilemma of book scanning—destructively debinding books with a razor or guillotine cutter works much better & is much less time-consuming than spreading them on a flatbed scanner to scan one-by-one14⁠, because it allows use of a sheet-fed scanner instead, which is easily 5x faster and will give higher-quality scans (because the sheets will be flat, scanned edge-to-edge, and much more closely aligned), but does, of course, require effectively destroying the book.

  • Tools:

    • cutting: For simple debinding of a few books a year, an X-acto knife/razor is good (avoid the ‘triangle’ blades, get curved blades intended for large cuts instead of detail work).

      Once you start doing more than one a month, it’s time to upgrade to a guillotine blade paper cutter (a fancier swinging-arm paper cutter, which uses a two-joint system to clamp down and cut uniformly).

      A guillotine blade can cut chunks of 200 pages easily without much slippage, so for books with more pages, I use both: an X-acto to cut along the spine and turn it into several 200-page chunks for the guillotine cutter.

    • scanning: at some point, it may make sense to switch to a scanning service like 1DollarScan (1DS has acceptable quality for the black-white scans I have used them for thus far, but watch out for their nickel-and-diming fees for OCR or “setting the PDF title”; these can be done in no time yourself using gscan2pdf/exiftool/ocrmypdf and will save a lot of money as they, amazingly, bill by 100-page units). Books can be sent directly to 1DS, reducing logistical hassles.

  • Clean Up: after scanning, crop/threshold/OCR/add metadata

    • Adding metadata: same principles as papers. While more elaborate metadata can be added, like bookmarks, I have not experimented with those yet.
  • File format: PDF.

    In the past, I used for documents I produce myself, as it produces much smaller scans than gscan2pdf’s default PDF settings due to a buggy Perl library (at least half the size, sometimes one-tenth the size), making them more easily hosted & a superior browsing experience.

    The downsides of DjVu are that not all PDF viewers can handle DjVu files, and it appears that G/GS ignore all DjVu files (despite the format being 20 years old), rendering them completely unfindable online. In addition, DjVu is an increasingly obscure format and has, for example, been dropped by the IA as of 2016. The former is a relatively small issue, but the latter is fatal—being consigned to oblivion by search engines largely defeats the point of scanning! (“If it’s not in Google, it doesn’t exist.”) Hence, despite being a worse format, I now recommend PDF and have stopped using DjVu for new scans15 and have converted my old DjVu files to PDF.

  • Uploading: to LibGen, usually. For backups, filelockers like Dropbox, Mega, MediaFire, or Google Drive are good. I usually upload 3 copies including LG. I rotate accounts once a year, to avoid putting too many files into a single account.

    Do Not Use Google Docs/Scribd/Dropbox/etc

    ‘Document’ websites like Google Docs (GD) should be strictly avoided as primary hosting. GD does not appear in G/GS, dooming a document to obscurity, and Scribd is ludicrously user-hostile. Such sites cannot be searched, scraped, downloaded, clipped, used on many devices, or counted on for the long haul.

    Such sites may be useful for collaboration or surveys, but should be moved to clean static HTML/PDF hosted elsewhere as soon as possible.
  • Hosting: hosting papers is easy but books come with risk:

    Books can be dangerous; in deciding whether to host a book, my rule of thumb is host only books pre-2000 and which do not have Kindle editions or other signs of active exploitation and is effectively an ‘’.

    As of 2019-10-23, hosting 4090 files over 9 years (very roughly, assuming linear growth, <6.7 million document-days of hosting: 3763 × 0.5 × 8 × 365.25 = 6722426), I’ve received 4 takedown orders: a behavioral genetics textbook (2013), The Handbook of Psychopathy (2005), a recent meta-analysis paper (Roberts et al 2016), and a CUP DMCA takedown order for 27 files. I broke my rule of thumb to host the 2 books (my mistake), which leaves only the 1 paper, which I think was a fluke. So, as long as one avoids relatively recent books, the risk should be minimal.

Case Studies

Below are >13 case studies of difficult-to-find resources or citations, and how I went about locating them, demonstrating the various Internet search techniques described above and how to think about searches.

  • Missing Appendix: asked:

    Does anybody know where the online appendix to Nordhaus’ “Two Centuries of Productivity Growth in Computing” is hiding?

    I look up the title in Google Scholar; seeing a friendly PDF link (CiteSeerx), I click. The paper says “The data used in this study are provided in a background spreadsheet available at”. Sadly, this is a lie. (Sandberg would of course have tried that.)

    I immediately check the URL in the IA—nothing. The IA didn’t catch it at all. Maybe the official published paper website has it? Nope, it references the same URL, and doesn’t provide a copy as an appendix or supplement. (What do we pay these publishers such enormous sums of money for, exactly?) So I back off to checking, to check Nordhaus’s personal website for a newer link. The Yale personal website is empty and appears to’ve been replaced by a Google Sites personal page. It links nothing useful, so I check a more thorough index, Google, by searching Nothing there either (and it appears almost empty, so Nordhaus has allowed most of his stuff to be deleted and bitrot). I try a broader Google: nordhaus appendix.xls. This turns up some spreadsheets, but still nothing.

    Easier approaches having been exhausted, I return to the IA and I pull up all URLs archived for his original personal website:*/* This pulls up way too many URLs to manually review, so I filter results for xls, which reduces to a more manageable 60 hits; reading through the hits, I spot from 2014-10-10; this sounds right, albeit substantially later in time than expected (either 2010 or 2012, judging from the filename).

    Downloading it⁠, opening it up and cross-referencing with the paper, it has the same spreadsheet ‘sheets’ as mentioned, like “Manual” or “Capital_Deep”, and seems to be either the original file in question or an updated version thereof (which may be even better). The spreadsheet metadata indicates it was created “04/09/2001, 23:20:43, ITS Academic Media & Technology”, and modified “12/22/2010, 02:40:20”, so it seems to be the latter—it’s the original spreadsheet Nordhaus created when he began work several years prior to the formal 2007 publication (6 years seems reasonable given all the delays in such a process), and then was updated 3 years afterwards. Close enough.

  • Misremembered Book: A Redditor asked:

    I was in a consignment type store once and picked up a book called “Eat fat, get thin”. Giving it a quick scan through, it was basically the same stuff as Atkins but this book was from the 50s or 60s. I wish I’d have bought it. I think I found a reference to it once online but it’s been drowned out since someone else released a book with the same name (and it wasn’t Barry Groves either).

    The easiest way to find a book given a corrupted title, a date range, and the information there are many similar titles drowning out a naive search engine query, is to skip to a specialized search engine with clean metadata (ie. a library database).

    Searching in WorldCat for 1950s–1970s, “Eat fat, get thin” turns up nothing relevant. This is unsurprising, as he was unlikely to’ve remembered the title exactly, and this title doesn’t quite sound right for the era anyway (a little too punchy and ungrammatical, and ‘thin’ wasn’t a desirable word back then compared to words like ‘slim’ or ‘sleek’ or ‘svelte’). People often oversimplify titles, so I dropped back to just “Eat fat”.

    This immediately turned up the book: 1958 Eat Fat and Grow Slim—note that it is almost the same title, with a comma serving as conjunction and ‘slim’ rather than the more contemporary ‘thin’, but just different enough to screw up an overly-literal search.

    With the same trick in mind, we could also have found it in a regular Google search query by adding additional terms to hint to Google that we want old books, not recent ones: both "Eat Fat" 1950s or "Eat Fat" 1960s would have turned it up in the first 5 search results. If we didn’t use quotes, the searches get harder because broader hits get pulled in. For example, Eat fat, get thin 1950s -Hyman excludes the recent book mentioned, but you still have to go down 15 hits before finding Mackarness, and Eat fat, get thin -Hyman requires going down 18 hits.

  • Missing Website: ⁠, on the phenomenon of quotes striking transcripts from a major example of a disappearing crystal, when ~1998 Abbott suddenly became unable to manufacture the anti-retroviral drug (Norvir™) due to a rival (and less effective) crystal form spontaneously infecting all its plants, threatening many AIDS patients, but notes:

    The transcripts were originally published on the website42 of the International Association of Physicians in AIDS Care [IAPAC], but no longer appear there.

    A search using the quotes confirms that the originals have long since vanished from the open Internet, turning up only quotes of the quotations. Unfortunately, no URL is given. The Internet Archive has comprehensive mirrors of the IAPAC, but too many to easily search through. Using the filter feature, I keyword-searched for “ritonavir”, but while this turned up a number of pages from roughly the right time period, they do not mention it and none of the quotes appear. The key turned out to be to use the trademark name instead which pulls up many more pages, and after checking a few, the IAPAC turned out to have organized all the Norvir material into a single subdirectory with a convenient index.html⁠; the articles/transcripts, in turn, were indexed under the linked ⁠.

    I then pulled the Norvir subdirectory with a ~/.gem/ruby/2.5.0/bin/wayback_machine_downloader wayback_machine_downloader '' command and hosted a mirror to make it visible in Google.

  • Speech → Book: Nancy Lebovitz asked about a citation in a Roy Baumeister speech about sex differences:

    There’s an idea I’ve seen a number of times that 80% of women have had descendants, but only 40% of men. A little research tracked it back to this⁠, but the speech doesn’t have a cite and I haven’t found a source.

    This could be solved by guessing that the formal citation is given in the book, and doing keyword search to find a similar passage. The second line of the speech says:

    For more information on this topic, read Dr. Baumeister’s book Is There Anything Good About Men? available in bookstores everywhere, including here.

    A search of Is There Anything Good About Men in Libgen turns up a copy. Download. What are we looking for? A reminder, the key lines in the speech are:

    …It’s not a trick question, and it’s not 50%. True, about half the people who ever lived were women, but that’s not the question. We’re asking about all the people who ever lived who have a descendant living today. Or, put another way, yes, every baby has both a mother and a father, but some of those parents had multiple children. Recent research using DNA analysis answered this question about two years ago. Today’s human population is descended from twice as many women as men. I think this difference is the single most under-appreciated fact about gender. To get that kind of difference, you had to have something like, throughout the entire history of the human race, maybe 80% of women but only 40% of men reproduced.

    We could search for various words or phrase from this passage which seem to be relatively unique; as it happens, I chose the rhetorical “50%” (but “80%”, “40%”, “underappreciated”, etc all would’ve worked with varying levels of efficiency since the speech is heavily based on the book), and thus jumped straight to chapter 4, “The Most Underappreciated Fact About Men”. (If these had not worked, we could have started searching for years, based on the quote “about two years ago”.) A glance tells us that Baumeister is discussing exactly this topic of reproductive differentials, so we read on and a few pages later, on page 63, we hit the jackpot:

    The correct answer has recently begun to emerge from DNA studies, notably those by Jason Wilder and his colleagues. They concluded that among the ancestors of today’s human population, women outnumbered men about two to one. Two to one! In percentage terms, then, humanity’s ancestors were about 67% female and 33% male.

    Who’s Wilder? A C-f for “Wilder” takes us to pg286, where we immediately read:

    …The DNA studies on how today’s human population is descended from twice as many women as men have been the most requested sources from my earlier talks on this. The work is by Jason Wilder and his colleagues. I list here some sources in the mass media, which may be more accessible to laypersons than the highly technical journal articles, but for the specialists I list those also. For a highly readable introduction, you can Google the article “Ancient Man Spread the Love Around,” which was published September, 20, 2004 and is still available (last I checked) online. There were plenty of other stories in the media at about this time, when the research findings first came out. In “Medical News Today,”⁠, on the same date in 2004, a story under “Genes expose secrets of sex on the side” covered much the same material.

    If you want the original sources, read Wilder, J. A., Mobasher, Z., & Hammer, M. F. (2004). “Genetic evidence for unequal effective population sizes of human females and males”⁠. Molecular Biology and Evolution, 21, 2047–2057. If that went down well, you might try Wilder, J. A., Kingan, S. B., Mobasher, Z., Pilkington, M. M., & Hammer, M. F. (2004). “Global patterns of human mitochondrial DNA and Y-chromosome structure are not influenced by higher migration rates of females versus males”⁠. Nature Genetics, 36, 1122–1125. That one was over my head, I admit. A more readable source on these is Shriver, M. D. (2005), “Female migration rate might not be greater than male rate”⁠. European Journal of Human Genetics, 13, 131–132. Shriver raises another intriguing hypothesis that could have contributed to the greater preponderance of females in our ancestors: Because couples mate such that the man is older, the generational intervals are smaller for females (i.e., baby’s age is closer to mother’s than to father’s). As for the 90% to 20% differential in other species, that I believe is standard information in biology, which I first heard in one of the lectures on testosterone by the late James Dabbs, whose book Heroes, Rogues, and Lovers remains an authoritative source on the topic.

    Wilder et al 2004, incidentally, fits well with Baumeister remarking in 2007 that the research was done 2 or so years ago. And of course you could’ve done the same thing using Google Books: search “Baumeister anything good about men” to get to the book, then search-within-the-book for “50%”, jump to page 53, read to page 63, do a second search-within-the-book for “Wilder” and the second hit of page 287 even luckily gives you the snippet:

    Sources and References 287

    …If you want the original sources, read Wilder, J. A., Mobasher, Z., & Hammer, M. F. (2004). “Genetic evidence for unequal effective population sizes of human females and males”. Molecular Biology and Evolution

  • Connotations a commenter who shall remain nameless wrote

    I challenge you to find an example of someone saying “this den of X” where X does not have a negative connotation.

    I found a positive connotation within 5s using my Google hotkey for "this den of ", and, curious about further ones, found additional uses of the phrase in regard to dealing with rattlesnakes in Google Books.

  • Rowling Quote On Death: Did say the Harry Potter books were about ‘death’? There are a lot of Rowling statements, but checking WP and opening up each interview links (under the theory that the key interviews are linked there) and searching for ‘death’ soon turns up a relevant quote from 2001:

    Death is an extremely important theme throughout all seven books. I would say possibly the most important theme. If you are writing about Evil, which I am, and if you are writing about someone who is essentially a psychopath, you have a duty to show the real evil of taking human life.

  • Crowley Quote: Scott Alexander posted a piece linking to an except titled “ on Religious Experience”.

    The link was broken, but Alexander brought it up in the context of an earlier discussion where he also quoted Crowley; searching those quotes reveals that it must have been excerpts from Magick: Book 4

  • Finding The Right ‘SAGE: Phil Goetz noted that an anti-aging conference named “SAGE” had become impossible to find in Google due to a LGBT aging conference also named SAGE.

    Regular searches would fail, but a combination of tricks worked: SAGE anti-aging conference combined with restricting Google search to 2003–2005 time-range turned up a citation to its website as the fourth hit, (which has ironically since died).

  • UK Charity Financials: The Future of Humanity Institute (FHI) doesn’t clearly provide charity financial forms akin to the US Form 990s, making it hard to find out information about its budget or results.

    FHI doesn’t show up in the CC, NPC, or GuideStar, which are the first places to check for charity finances, so I went a little broader afield and tried a site search on the FHI website: budget This immediately turned up FHI’s own documentation of its activities and budgets, such as the 2007 annual report; I used part of its title as a new Google search: future of humanity institute achievements report

  • Nobel Lineage Research: John Maxwell referred to a forgotten study on high correlation between Nobelist professors & Nobelist grad students (almost entirely a selection effect, I would bet). I was able to refind it in 7 minutes.

    I wasted a few searches like factor predicting Nobel prize or Nobel prize graduate student in Google Scholar, until I search for Nobel laureate "graduate student"; the second hit was a citation, which is a little unusual for Google Scholar and meant it was important, and it had the critical word mutual in it—simultaneous partners in Nobel work is somewhat rare, but temporally separated teams don’t work for prizes, and I suspected that it was exactly what I was looking for. Googling the title, I soon found a PDF like “Eminent Scientists’ Demotivation in School: A symptom of an incurable disease?”, Viau 2004 which confirmed it (and Viau 2004 is interesting in its own right as a contribution to the Conscientious vs IQ question). I then followed it to a useful paragraph:

    In a study conducted with 92 American winners of the Nobel Prize, Zuckerman (1977) discovered that 48 of them had worked as graduate students or assistants with professors who were themselves Nobel Prize award-winners. As pointed out by Zuckerman (1977), the fact that 11 Nobel prizewinners have had the great physicist Rutherford as a mentor is an example of just how significant a good mentor can be during one’s studies and training. It then appears that most eminent scientists did have people to stimulate them during their childhood and mentor(s) during their studies. But, what exactly is the nature of these people’s contribution.

    • Zuckerman, H. (1977). Scientific Elite: Nobel Laureates in the United States. New York: Free Press.

    GS lists >900 citations of this book, so there may well be additional or followup studies covering the 40 years since. Or, also relevant is “Zuckerman, H. (1983). The scientific elite: Nobel laureates’ mutual influences. In R. S. Albert (Ed.), Genius and eminence (pp. 241–252). New York: Pergamon Press”, and “Zuckerman H. ‘Sociology of Nobel Prizes’, Scientific American 217 (5): 25& 1967.”

  • Too Narrow: A failure case study: The_Duck looked for but failed to find other uses of a famous Wittgenstein anecdote. His mistake was being too specific:

    Yes, clearly my Google-fu is lacking. I think I searched for phrases like “sun went around the Earth,” which fails because your quote has “sun went round the Earth.”

    As discussed in the search tips, when you’re formulating a search, you want to balance how many hits you get, aiming for a sweet spot of a few hundred high-quality hits to review—the broader your formulation, the more likely the hits will include your target (if it exists) but the more hits you’ll return. In The_Duck’s case, he used an overly-specific search, which would turn up only 2 hits at most; this should have been a hint to loosen the search, such as by dropping quotes or dropping keywords.

    In this case, my reasoning would go something like this, laid out explicitly: ‘“Wittgenstein” is almost guaranteed to be on the same page as any instance of this quote, since the quote is about Wittgenstein; LW, however, doesn’t discuss Wittgenstein much, so there won’t be many hits in the first place; to find this quote, I only need to narrow down those hits a little, and after “Wittgenstein”, the most fundamental core word to this quote is “Earth” or “sun”, so I’ll toss one of them in and… ah, there’s the quote!’

    If I were searching the general Internet, my reasoning would go more like “‘Wittgenstein’ will be on, like, a million websites; I need to narrow that down a lot to hope to find it; so maybe ‘Wittgenstein’ and ‘Earth’ and ‘Sun’… nope, nothing on the first page, so toss in 'goes around' OR 'go around'—ah there it is!”

    (Actually, for the general Internet, just Wittgenstein earth sun turns up a first page mostly about this anecdote, several of which include all the details one could need.)

  • Dead URL: A link to a research article in a post by Morendil broke, he had not provided any formal citation data, and the original domain blocks all crawlers in its robots.txt so IA would not work. What to do?

    The simplest solution was to search a direct quote, turning up a Scribd mirror; Scribd is a parasite website, where people upload copies from elsewhere, which ought to make one wonder where the original came from. (It often shows up before the original in any search engine, because it automatically runs OCR on submissions, making them more visible to search engines.) With a copy of the journal issue to work with, you can easily find the official HP archives and download the original PDF⁠.

    If that hadn’t worked, searching for the URL without /pg_2/ in it yields the full citation, and then that can be looked up normally. Finally, somewhat more dangerous would be trying to find the article just by author surname & year.

  • Description But No Citation: A 2013 Medical Daily on the effects of reading fiction omitted any link or citation to the research in question. But it is easy to find.

    The article says the authors are one Kaufman & Libby, and implies it was published in the last year. So: go to Google Scholar, punch in Kaufman Libby, limit to ‘Since 2012’; and the correct paper (“Changing beliefs and behavior through experience-taking”) is the first hit with fulltext available on the right-hand side as the text link “[PDF] from” & many other domains.

  • Finding Followups: Is soy milk bad for you as one study suggests? Has anyone replicated it? This is easy to look into a little if you use the power of reverse citation search!

    Plug Brain aging and midlife tofu consumption into Google Scholar, one of the little links under the first hit points to “Cited by 176”; if you click on that, you can hit a checkbox for “Search within citing articles”; then you can search a query like experiment OR randomized OR blind which yields 121 results⁠. The first result shows no negative effect and a trend to a benefit, the second is inaccessible, the second & third are reviews whose abstract suggests it would argue for benefits, and the fourth discusses sleep & mood benefits to soy diets. At least from a quick skim, this claim is not replicating, and I am dubious about it.

  • How Many Homeless?: does NYC really have 114,000+ homeless school children? This case study demonstrates the critical skill of noticing the need to search at all, and the search itself is almost trivial.

    Won’t someone think of the children? In March 2020, as centered in Manhattan (with a similar trend to Wuhan/Iran/Italy), NYC Mayor refused to take social distancing/quarantine measures like ordering the NYC public school system closed, and this delay until 16 March contributed to the epidemic’s unchecked spread in NYC; one justification was that there were “114,085 homeless children” who received social services like free laundry through the schools. This number has been widely cited in the media by the NYT, WSJ, etc, and was vaguely sourced to “state data” reported by “Advocates for Children of New York”. This is a terrible reason to not deal with a pandemic that could kill tens of thousands of New Yorkers, as there are many ways to deliver services which do not require every child in NYC to attend school & spread infections—but first, is this number even true?

    Basic numeracy: implausibly-large! Activists of any stripe are untrustworthy sources, and a number like 114k should make any numerate person uneasy even without any or fact-checking; “114,085” is suspiciously precise for such a difficult-to-measure or define thing like homelessness, and it’s well-known that the population of NYC is ~8m or 8,000k—is it really the case that around 1 in every 70 people living in NYC is a homeless child age ~5–18 attending a public school? They presumably have at least 1 parent, and probably younger siblings, so that would bring it up to >228k or 1 in every <35 inhabitants of NYC being homeless in general. Depending on additional factors like transiency & turnover, the fraction could go much higher still. Does that make sense? No, not really. This quoted number is either surprising, or there is something missing.

    Redefining “homeless”. Fortunately, the suspiciously-precise number and attribution make this a good place to start for a search. Searching for the number and the name of the activist group instantly turns up the source press release⁠, and the reasons for the bizarrely high number are revealed: the statistic actually redefines ‘homelessness’ to include living with relatives or friends, and counts any experience of any length in the previous year as rendering that student ‘homeless’ at the moment.

    The data, which come from the New York State Education Department, show that in the 2018-2019 school year, New York City district and charter schools identified 114,085, or one in ten, students as homeless. More than 34,000 students were living in New York City’s shelters, and more than twice that number (73,750) were living ‘doubled-up’ in temporary housing situations with relatives, friends, or others…“This problem is immense. The number of New York City students who experienced homelessness last year—85% of whom are Black or Hispanic—could fill the Barclays Center six times,” said Kim Sweet, AFC’s Executive Director. “The City won’t be able to break the cycle of homelessness until we address the dismal educational outcomes for students who are homeless.”

    The WSJ’s article (but not headline) confirms that ‘experienced’ does indeed mean ‘at any time in the year for any length of time’, rather than ‘at the moment’:

    City district and charter schools had 114,085 students without their own homes at some point last year, topping 100,000 for the fourth year in a row, according to state data released in a report Monday from Advocates for Children of New York, a nonprofit seeking better services for the disadvantaged. Most children were black or Hispanic, and living “doubled up” with friends, relatives or others. But more than 34,000 slept in city shelters at some point, a number larger than the entire enrollment of many districts, such as Buffalo, Rochester or Yonkers.

    Less than meet the eye. So the actual number of ‘homelessness’ (in the sense that everyone reading those media articles understands it) is less than a third the quote, 34k, and that 34k number is likely itself a loose estimate of how many students would be homeless at the time of a coronavirus closure. This number is far more plausible and intuitive, and while one might wonder about what the underlying NYS Education Department numbers would reveal if fact-checked further, that’s probably unnecessary for showing how ill-founded the anti-closure argument is, since even by the activists’ own description, the relevant number is far smaller than 114k.

  • Citation URL With Typo: ⁠, discusses the limits to the intelligence of increasingly large primate brains due to considerations like increasing latency and overheating. One citation attempting to extrapolate upper bounds is “Biological limits to information processing in the human brain”, Cochrane et al 1995.

    The source information is merely a broken URL: which stands out for looking doubly-wrong: “.phd” is almost certainly a typo for “.php” (probably muscle memory on the part of Hofman from “PhD”), but it also gives a hint that the entire URL is wrong: why would an article or essay be named anything like archive/articles.php? That sounds like an index page listing all the available articles.

    After trying and failing to find Cochrane’s paper in the usual places, I returned to the hint. The Internet Archive doesn’t have that page under either possible URL, but the directory strongly hints that all of the papers would exist at URLs like archive/brain.php or archive/information-processing.php, and we can look up all of the URLs the IA has under that directory—how many could there be? A lot⁠, but only one has the keyword “brain” in it, providing us ⁠.

    If that hadn’t worked, there was at least one other version hiding in the IA. When I googled the quoted title “Biological limits to information processing in the human brain”, the hits all appeared to be useless citations repeating the original Hofman citation—but for a crucial difference, as they cite a different URL (note the shift to an ‘’ subdomain rather than the subdirectory, and change of extension from .html to .php):

    • hit 5:

      Biological Limits to Information Processing in the Human Brain. Retrieved from:

    • hit 7:

      Biological Limits to Information Processing in the Human Brain. Available online at:; Da Costa …

    Aside from confirming that it was indeed a ‘.php’ extension, that URL gives you a second copy of the paper in the IA⁠. Unfortunately, the image links are broken in both versions, and the image subdirectories also seem to be empty in both IA versions, though there’s no weird JS image loading badness, so I’d guess that the image links were always broken, at least by 2004. There’s no indication it was ever published or mirrored anywhere else, so there’s not much you can do about it other than to contact Peter Cochrane (who is still alive and actively publishing although he leaves this particular article off his publication list). -Really, Try It: Someone asked on IRC: “anybody here know that one artist with the really creepy art sytle that starts with a z?”

    I googled: ‘that one artist with the really creepy art sytle that starts with a z’. It was hit #2, ⁠. (DuckDuckGo, incidentally, buries Beksiński several pages in, and I didn’t find him in Bing at all.)

  • Comics: Julia Galef tweeted:

    I read a webcomic ~15 years ago that I’ve been unable to find since, even with my best google-fu. It involved a robot living a bleak life as a working stiff. At the end he cracked open his “skull” and there was a small dying creature inside. The art style was less cartoony, and more like Moebius, I think? And maybe it was wordless? And, sorry, it wasn’t a “webcomic” in the sense of a long-running thing. It was a self-contained story, maybe 15 pages long?

    Ultimately rediscovering that

    The comic was called and it was by Sam Chivers.

    Unfortunately, no mirrors of it appeared online or on Chivers’s current website, and discussions of it mentioned that it was interesting for being an webcomic. Worse still, nothing useful appeared in the Internet Archive for the original website—somehow the IA appeared to have missed any relevant .swf files, and ‘head’/‘case’ turned up no relevant looking filenames. It might have been buried in the opaquely-named images, and my usual next step would be to download the IA archives and inspect every image, but in other hits, I found that an obscure comics publisher had published an anthology involving Chivers, and confirmed that “Headcase” was in fact published in their (long out of print) 2004 anthology Prophecies: Volume 1. (Not a prophetic name inasmuch as there was no volume 2.)

    In one of the usual ironies of linkrot, Chivers presumably taking down “Headcase” for print publication in Prophecy may have preserved it, as while I am unable to find any digital copies, the paper version is easily obtained as a used book & scanned at modest cost.

  • Beating PDF Passwords: a physics article mentioned they had been unable to get an old 1973 interview in a popular magazine; as is usually the case for non-scholarly magazines, after looking thoroughly, I could find no trace of it anywhere (not even in libraries or used-magazine sellers) other than an expensive DVD collection of backissues 1970–2010 still being sold by the publisher. Reasoning that if they had digitized the archives and were even selling it as a DVD collection, they ought to provide subscribers access to them as well, I signed up—they didn’t! So I resorted to the DVD, as, worst-case, I should be able to get it running under if nothing else, and can screenshot the interview.

    The DVDs turned out to store all the PDFs as encrypted PDFs and the metadata in an ancient opaque database format I’d never heard of. Despite WINE’s claims, the viewing software only partially worked, and I set about attacking the PDFs directly. They used actual encryption, so pdftk couldn’t strip the passwording. Given the viewing software, I hypothesized that there was either a single master password or per-PDF passwords stored in the database.

    In the hopes of it being a single short master password, I installed (JtR) jumbo edition and extracted the hash of a random file to attack: /snap/john-the-ripper/current/run/ *.pdf > ~/hash. (Note: pdf2john is not in the default JtR, and it depends on JtR internal files so you can’t easily just copy it out of the Github repo & run it, as I discovered the hard way. You need to install the jumbo edition.) The password hashes of all the PDFs indeed turned out to be the same, so it used a master password. A simple attack with default password-space could be executed as john-the-ripper ~/hash. While I waited for all of the DVDs to copy, I saw that JtR was getting something like only a hundred thousand hashes/s on my 16 Threadripper CPU cores, and did not have any success up to 5-character passwords.

    If the password wasn’t really short, CPU wouldn’t be enough. I decided to switch to to put my 2×1080ti Nvidia GPUs to good use, as they ought to run hundreds of times faster than JtR. (To convert the JtR hash format to Hashcat hash format, you delete the colon-separated filename field at the beginning of each line.) Hashcat uses a powerful but confusing DSL of specifying the exact password-space, and I made a reasonable guess that if the original programmer was so lazy as to use a single master password, he would also use a simple alphanumeric password (uppercase + lowercase + decimal numbers), and nothing harder to type or read. To specify the PDF hash type and an attack starting at 1-character alphanumeric & increasing, I wound up with the incantation hashcat -m 10500 ~/ -w 3 --force -a 3 --increment -1 '?l?u?d' ?1?1?1?1?1?1?1?1?1?1?1.

    Hashcat worked much better and within an hour had bruteforced on the order of 170 billion hashes and up somewhere around 8 characters. This did not succeed either. At this point, another programmer thought it’d be fun to participate and, while reverse-engineering the executable to see how it decrypted PDFs, suggested that the master password was probably hardcoded as a string literal inside the viewer executable. One could just dump all the strings inside it with the CLI utility strings *.exe > strings.txt, and then use it as a Hashcat password list. To my chagrin, when I finally got around to trying cat strings.txt | hashcat -m 10500 ~/ -w 3, it finished within 1s.

    The password turned out to be B775tO11dQvu74. I was right that it was alphanumerical, but at a length of 14 characters, I doubt I would have brute-forced it. (He successfully reverse-engineered it and discovered the viewer had been used for several other magazine archives as well, apparently, and simply switched master passwords to decrypt each one; the other passwords left in the executable were PbS19LuXd2pTXw, 1386r8wRrH01, & mfU33QQNlAFGI1.)

    I then decrypted the PDF (for PDF in *.pdf; do pdftk "$PDF" input_pw "B775tO11dQvu74" output foo.pdf && mv foo.pdf "$PDF"; done), extracted & uploaded the interview, and archived the collection elsewhere.

See Also


Searching the Google Reader archives

A tutorial on how to do manual searches of the 2013 archives on the ⁠. Google Reader provides fulltext mirrors of many websites which are long gone and not otherwise available even in the IA; however, the Archive Team archives are extremely user-unfriendly and challenging to use even for programmers. I explain how to find & extract specific websites.

A little-known way to ‘undelete’ a blog or website is to use Google Reader (GR). Unusual archive: Google Reader. GR crawled regularly almost all blogs’ RSS feeds; RSS feeds often contain the fulltext of articles. If a blog author writes an article, the fulltext is included in the RSS feed, GR downloads it, and then the author changes their mind and edits or deletes it, GR would redownload the new version but it would continue to show the version the old version as well (you would see two versions, chronologically). If the author blogged regularly and so GR had learned to check regularly, it could hypothetically grab different edited versions, even, not just ones with weeks or months in between. Assuming that GR did not, as it sometimes did for inscrutable reasons, stop displaying the historical archives and only showed the last 90 days or so to readers; I was never able to figure out why this happened or if indeed it really did happen and was not some sort of UI problem. Regardless, if all went well, this let you undelete an article, albeit perhaps with messed up formatting or something. Sadly, GR was closed back in 2013 and you cannot simply log in and look for blogs.

Archive Team mirrored Google Reader. However, before it was closed, Archive Team launched a major effort to download as much of GR as possible. So in that dump, there may be archives of all of a random blog’s posts. Specifically: if a GR user subscribed to it; if Archive Team knew about it; if they requested it in time before closure; and if GR did keep full archives stretching back to the first posting.

AT mirror is raw binary data. Downside: the Archive Team dump is not in an easily browsed format, and merely figuring out what it might have is difficult. In fact, it’s so difficult that before researching Craig Wright in November–December 2015, I never had an urgent enough reason to figure out how to get anything out of it before, and I’m not sure I’ve ever seen anyone actually use it before; Archive Team takes the attitude that it’s better to preserve the data somehow and let posterity worry about using it. (There is a site which claimed to be a frontend to the dump but when I tried to use it, it was broken & still is in December 2018.)


Find the right archive. The 9TB of data is stored in ~69 opaque compressed WARC archives. 9TB is a bit much to download and uncompress to look for one or two files, so to find out which WARC you need, you have to download the ~69 CDX indexes which record the contents of their respective WARC, and search them for the URLs you are interested in. (They are plain text so you can just grep them.)


In this example, we will look at the main blog of Craig Wright, (Another blog,, appears to have been too obscure to be crawled by GR.) To locate the WARC with the Wright RSS feeds, download the the master index⁠. To search:

for file in *.gz; do echo $file; zcat $file | fgrep -e 'gse-compliance' -e 'security-doctor'; done
# com,google/reader/api/0/stream/contents/feed/http:/\
# archiveteam&comments=true&hl=en&likes=true&n=1000&r=n 20130602001238\
# api/0/stream/contents/feed/\
# likes=true&comments=true&client=ArchiveTeam unk - 4GZ4KXJISATWOFEZXMNB4Q5L3JVVPJPM - - 1316181\
# 19808229791 archiveteam_greader_20130604001315/greader_20130604001315.megawarc.warc.gz
# com,google/reader/api/0/stream/contents/feed/http:/\
# alt=rss?client=archiveteam&comments=true&hl=en&likes=true&n=1000&r=n 20130602001249\
# com/reader/api/0/stream/contents/feed/\
# %3Falt%3Drss?r=n&n=1000&hl=en&likes=true&comments=true&client=ArchiveTeam unk - HOYKQ63N2D6UJ4TOIXMOTUD4IY7MP5HM\
# - - 1326824 19810951910 archiveteam_greader_20130604001315/greader_20130604001315.megawarc.warc.gz
# com,google/reader/api/0/stream/contents/feed/http:/\
# client=archiveteam&comments=true&hl=en&likes=true&n=1000&r=n 20130602001244\
# reader/api/0/stream/contents/feed/\
# r=n&n=1000&hl=en&likes=true&comments=true&client=ArchiveTeam unk - XXISZYMRUZWD3L6WEEEQQ7KY7KA5BD2X - - \
# 1404934 19809546472 archiveteam_greader_20130604001315/greader_20130604001315.megawarc.warc.gz
# com,google/reader/api/0/stream/contents/feed/http:/\
# &comments=true&hl=en&likes=true&n=1000&r=n 20130602001253\
# /feed/\
# &client=ArchiveTeam text/html 404 AJSJWHNSRBYIASRYY544HJMKLDBBKRMO - - 9467 19812279226 \
# archiveteam_greader_20130604001315/greader_20130604001315.megawarc.warc.gz

Understanding the output: the format is defined by the first line, which then can be looked up:

  • the format string is: CDX N b a m s k r M S V g; which means here:

    • N: massaged url
    • b: date
    • a: original url
    • m: mime type of original document
    • s: response code
    • k: new style checksum
    • r: redirect
    • M: meta tags (AIF)
    • S: ?
    • V: compressed arc file offset
    • g: file name


?client=archiveteam&comments=true&hl=en&likes=true&n=1000&r=n 20130602001238\
&n=1000&hl=en&likes=true&comments=true&client=ArchiveTeam unk - 4GZ4KXJISATWOFEZXMNB4Q5L3JVVPJPM\
- - 1316181 19808229791 archiveteam_greader_20130604001315/greader_20130604001315.megawarc.warc.gz

Converts to:

  • massaged URL: (com,google)/reader/api/0/stream/contents/feed/ http:/ client=archiveteam&comments=true&hl=en&likes=true&n=1000&r=n
  • date: 20130602001238
  • original URL: r=n&n=1000&hl=en&likes=true&comments=true&client=ArchiveTeam
  • MIME type: unk [unknown?]
  • response code:—[none?]
  • new-style checksum: 4GZ4KXJISATWOFEZXMNB4Q5L3JVVPJPM
  • redirect:—[none?]
  • meta tags:—[none?]
  • S [? maybe length?]: 1316181
  • compressed arc file offset: 19808229791 (19,808,229,791; so somewhere around 19.8GB into the mega-WARC)
  • filename: archiveteam_greader_20130604001315/greader_20130604001315.megawarc.warc.gz

Knowing the offset theoretically makes it possible to extract directly from the IA copy without having to download and decompress the entire thing… The S & offsets for gse-compliance are:

  1. 1316181/19808229791
  2. 1326824/19810951910
  3. 1404934/19809546472
  4. 9467/19812279226

So we found hits pointing towards archiveteam_greader_20130604001315 & archiveteam_greader_20130614211457 which we then need to download (25GB each):

wget ''
wget ''

Once downloaded, how do we get the feeds? There are a number of hard-to-use and incomplete tools for working with giant WARCs⁠; I contacted the original GR archiver, ivan, but that wasn’t too helpful.


I tried using warcat to unpack the entire WARC archive into individual files, and then delete everything which was not relevant:

python3 -m warcat extract /home/gwern/googlereader/...
find ./ -type f -not \( -name "*gse-compliance*" -or -name "*security-doctor*" \) -delete
find ./

But this was too slow, and crashed partway through before finishing.

Bug reports:

A more recent alternative library, which I haven’t tried, is warcio⁠, which may be able to find the byte ranges & extract them.


If we are feeling brave, we can use the offset and presumed length to have directly extract byte ranges:

dd skip=19810951910 count=1326824 if=greader_20130604001315.megawarc.warc.gz of=2.gz bs=1
# 1326824+0 records in
# 1326824+0 records out
# 1326824 bytes (1.3 MB) copied, 14.6218 s, 90.7 kB/s
dd skip=19810951910 count=1326824 if=greader_20130604001315.megawarc.warc.gz of=3.gz bs=1
# 1326824+0 records in
# 1326824+0 records out
# 1326824 bytes (1.3 MB) copied, 14.2119 s, 93.4 kB/s
dd skip=19809546472 count=1404934 if=greader_20130604001315.megawarc.warc.gz of=4.gz bs=1
# 1404934+0 records in
# 1404934+0 records out
# 1404934 bytes (1.4 MB) copied, 15.4225 s, 91.1 kB/s
dd skip=19812279226 count=9467 if=greader_20130604001315.megawarc.warc.gz of=5.gz bs=1
# 9467+0 records in
# 9467+0 records out
# 9467 bytes (9.5 kB) copied, 0.125689 s, 75.3 kB/s
dd skip=19808229791 count=1316181 if=greader_20130604001315.megawarc.warc.gz of=1.gz bs=1
# 1316181+0 records in
# 316181+0 records out
# 1316181 bytes (1.3 MB) copied, 14.6209 s, 90.0 kB/s
gunzip *.gz


Success: raw HTML. My dd extraction was successful, and the resulting HTML/RSS could then be browsed with a command like cat *.warc | fold --spaces -width=200 | less. They can probably also be converted to a local form and browsed, although they won’t include any of the site assets like images or CSS/JS, since the original RSS feed assumes you can load any references from the original website and didn’t do any kind of or mirroring (not, after all, having been intended for archive purposes in the first place…)

  1. For example, the info: operator is entirely useless. The link: operator, in almost a decade of me trying it once in a great while, has never returned remotely as many links to my website as Google Webmaster Tools returns for inbound links, and seems to have been disabled entirely at some point.↩︎

  2. WP is increasingly out of date & unrepresentative due to increasingly narrow policies about sourcing & preprints, part of its overall deletionist decay⁠, so it’s not a good place to look for references. It is a good place to look for key terminology, though.↩︎

  3. Most search engines will treat any space or separation as an implicit AND, but I find it helpful to be explicit about it to make sure I’m searching what I think I’m searching.↩︎

  4. This probably explains part of why no one cites that paper, and those who cite it clearly have not actually read it, even though it invented racial admixture analysis, which, since reinvented by others, has become a major method in medical genetics.↩︎

  5. University ILL privileges are one of the most underrated fringe benefits of being a student, if you do any kind of research or hobbyist reading—you can request almost anything you can find in ⁠, whether it’s an ultra-obscure book or a master’s thesis from 1950! Why wouldn’t you make regular use of it‽ Of things I miss from being a student, ILL is near the top.↩︎

  6. The complaint and indictment are not necessarily the same thing. An indictment frequently will leave out many details and confine itself to listing what the defendant is accused of. Complaints tend to be much richer in detail. However, sometimes there will be only one and not the other, perhaps because the more detailed complaint has been sealed (possibly precisely because it is more detailed).↩︎

  7. Trial testimony can run to hundreds of pages and blow through your remaining PACER budget, so one must be careful. In particular, testimony operates under an interesting & system related to how report—who are not necessarily paid employees but may be contractors or freelancers—intended to ensure covering transcription costs: the transcript initially may cost hundreds of dollars, intended to extract full value from those who need the trial transcript immediately, such as lawyers or journalists, but then a while later, PACER drops the price to something more reasonable. That is, the first “original” fee costs a fortune, but then “copy” fees are cheaper. So for the US federal court system⁠, the “original”, when ordered within hours of the testimony, will cost <$7.25/page but then the second person ordering the same transcript pays only <$1.20/page & everyone subsequently <$0.90/page, and as further time passes, that drops to <$0.60 (and I believe after a few months, PACER will then charge only the standard $0.10). So, when it comes to trial transcript on PACER, patience pays off.↩︎

  8. I’ve heard that LexisNexis terminals are sometimes available for public use in places like federal libraries or courthouses, but I have never tried this myself.↩︎

  9. Curiously, in historical textual criticism of copied manuscripts, it’s the opposite: ⁠. But with memories or paraphrases, longer = truer, because those tend to elide details and mutate into catchier versions when the transmitter is not ostensibly exactly copying a text.↩︎

  10. I advise prepending, like instead of appending, like because the former is slightly easier to type but more importantly, Sci-Hub does not have SSL certificates set up properly (I assume they’re missing a wildcard) and so appending the Sci-Hub domain will fail to work in many web browsers due to HTTPS errors! However, if prepended, it’ll always work correctly.↩︎

  11. To further illustrate this IA feature: if one was looking for Alex St. John’s “Judgment Day Continued…”⁠, a 2013 account of organizing the wild 1996 Doom tournament thrown by Microsoft, but one didn’t have the URL handy, one could search the entire domain by going to*/* and using the filter with “judgment”, or if one at least remembered it was in 2013, one could narrow it down further to*/* and then filter or search by hand.↩︎

  12. If any Blogspot employee is reading this, for god’s sake stop this insanity!↩︎

  13. Uploading is not as hard as it may seem. There is a web interface (user/password: “genesis”/“upload”). Uploading large files can fail, so I usually use the FTP server: curl -T "$FILE" ↩︎

  14. Although flatbed scanning is sometimes destructive too—I’ve cracked the spine of books while pressing them flat into a flatbed scanner.↩︎

  15. My workaround is to export from gscan2pdf as DjVu, which avoids the bug, then convert the DjVu files with ddjvu -format=pdf; this strips any OCR, so I add OCR with ocrmypdf and metadata with exiftool.↩︎