Writing that lasts ten years

On 27 June 2011, Justice Samuel Alito filed a concurring opinion in the United States Supreme Court. Footnote 14 cites a web page: Inside the Sick Site of a School Shooter Mod, at ssnat.com.

In April 2011 that address held what he said it held. By March 2012 the domain had lapsed. In June 2012 somebody else bought it and put up a note: if you had cited this page in the Supreme Court Reporter, “the original content would have long since disappeared”.

I checked it this month. The note has gone, and the page now argues about the size of the Supreme Court.

The opinion is printed, bound, and shelved in law libraries in several countries. It cannot be edited. The footnote is still there, still pointing, with a supreme court’s authority behind it.

A dead address on the back of three hundred zine copies fails the same way: the paper outlives the thing it points at, and the paper cannot be corrected.

The one that still works is the one that lies

Footnote 14 holds two links, and the footnote after it a third. I checked all three on 15 August 2026.

The Time article is gone: the address answers with an error. The CNN one is worse off, because the computer it used to live on no longer exists at all. And ssnat.com loads perfectly well, instantly, with nothing on it that has anything to do with what Alito read.

Librarians and archivists separate these. Link rot is when the address stops answering. Reference rot is when it answers and what you cited is no longer there.

Two Harvard studies measured both. Zittrain, Albert and Lessig found in 2014 that more than 70% of the links in law journal articles, and half of those in Supreme Court opinions, no longer led to the material being cited. Then in 2021 Bowers, Stanton and Zittrain went through the New York Times archive: 553,693 articles, 2.3 million links pointing off the site, a quarter of them completely inaccessible. Of 4,500 surviving links they checked by hand, 13% had drifted to something significantly different.

Anything that tallies broken links catches only link rot. Reference rot is invisible to a machine and worse for a person, because nothing tells you that you are reading the wrong thing.

How long a page actually lasts

Two studies have run long enough to be useful, and most of what gets quoted is not either of them. A web page lives 44 days, or 75, or 100, depending which article you read, and all three trace back to one man quoted in magazines across six years with no study underneath any of it. The Library of Congress in Washington, DC untangled that in 2011. Its own source link for the hundred-day version is now dead.

Two hold up. Wallace Koehler followed 361 randomly chosen pages from December 1996 to May 2003, the longest continuous study of a single set of addresses. He put the half-life of a random web page at about two years. Pew Research sampled just under a million pages and reported in 2024 that 38% of the pages that existed in 2013 had gone ten years later, with a quarter of everything they looked at across the whole decade gone.

The shape of Koehler’s loss matters more than its size. His sample fell from every page present in December 1996 to 34.4% by February 2001. Two years later, in May 2003, it stood at 33.8%.

Two further years cost it six-tenths of a percentage point. Almost all the dying happens early, and what gets through the first few years mostly keeps going. The question was never whether your work will still be there in ten years. It is whether it survives its first two, because after that the odds turn sharply in your favour.

Koehler also tabulated figures from other people’s studies, and the spread across them is the useful part. A random web page, two years. A citation in a scholarly article, eighteen months. A legal citation, seventeen. An object held in a digital library, twenty-four and a half years.

The difference there is not technical. It is whether keeping the thing is somebody’s job.

One more detail from his paper. He lost most of 2001, partly because the software he collected the data with would not run on anything newer than Windows 95. The tool he was measuring persistence with had not persisted.

What actually rots

The Movable Type announcement in Part Two left 236 blog addresses attached to it, one from every blog that replied on the day. I checked each one twice: is the name still theirs, and is the post still there?

Ninety-nine names are. Nineteen posts are.

The gap between those two numbers is the thing worth looking at. A name is cheap to keep and hard to lose by accident. The writing behind it is neither. Of the ninety-nine, nine redirect to wherever the writer went next, and three carry a note where the blog used to be: asmallvictory.net says “the end”, and gives a forwarding address. Of the rest, thirty-five belong to somebody else now, parked or for sale or filled with spam, and a hundred and two answer nothing at all.

The announcement is still quoted, and so is the best response to it, Mark Pilgrim’s “Freedom 0”, written the next day. Neither is still at the address where it was published. Both survive because other people kept copies.

Pilgrim’s piece was a reaction, the same as the 236 from the day before. The difference is what is inside it. He sets out what the old license allowed, what the new one allowed, what it would cost him, and why he objected. You can read it today knowing nothing about Movable Type and follow every step.

I should be careful about what that proves, though. The posts that died were mostly on personal sites that lapsed, and the ones that survived had somebody else keeping them. I tested whether addresses answer, not whether the writing was any good. All I can fairly claim is this: Pilgrim’s essay is readable by someone who has never heard of Movable Type, and a post saying “I can’t believe they’ve done this, see Mena’s announcement” would not be, even if its address had held.

Write for somebody who wasn’t there

That is the whole craft of it, as far as I can tell, and it costs nothing.

Name the date. Say what the thing said, rather than pointing at it. Quote the number instead of calling it high. Write “in August 2026” rather than “recently”, because a reader in 2036 cannot tell whether you meant last week or last decade.

If you print, this matters more rather than less. The copy in somebody’s flat is the one that outlives your address, and it is the one you cannot go back and fix. Everything in it has to work without you standing next to it.

It is also the same discipline that makes a claim checkable. A sentence with a date in it can be verified, argued with, or corrected. A sentence that gestures can only be believed or ignored. Researching the other parts I found several widely repeated dates that were simply wrong, and they were correctable only because somebody had written a specific claim down. Nobody can correct “a few years ago”.

Hosting is not preservation

I sell hosting, and I want to be plain about what it is not.

A host keeps your site answering while you pay. It is not an archive, it has no duty to outlive your subscription, and mine will not. The Internet Archive might have a copy of your work or might not, and owes you nothing either way: it is a charity with its own legal problems, doing a job nobody assigned it.

So the person who survives their supplier is the one holding their own copy. Not a backup inside somebody’s control panel. A folder of files, on a disk you can see, that opens without anybody’s permission. Run the export today and open what comes out, because the version of you who needs it will be having a much worse week.

That, and the printed copies. Paper turns out to have the longest support window anybody ever shipped.

Write it, so a stranger in ten years can follow it without you, and keep the address so they have somewhere to arrive. The copy that depends on neither of us is the one I would bet on.