Won is the right verb
The first thing to establish, before anything else, is that there was a contest. This sounds trivial. It is not. The way the history of the web is usually told — in introductions to programming books, in keynote addresses, in the official mythology that the web’s own institutions produce — presents the web as a kind of inevitable arrival, a thing that simply happened when the conditions were ripe, with no opponents worth naming and no losers worth mourning. This is wrong. The web arrived into a room full of working systems, several of which were older than it, several of which had more users than it had for the first three years of its existence, several of which had richer feature sets than it would have for another fifteen years, and several of which were specifically designed to do better the things the web was about to do worse. None of them survived as primary infrastructure. Each was beaten. This chapter is about why the verb is “won.”
The empty-room story and why it is false
The empty-room story goes like this. There were computers. They were connected by the internet. People wanted to share information. Tim Berners-Lee proposed the web. The web spread because it was useful. End of history. Inside this story, the web is not so much an invention as a recognition: the recognition that information could be linked, displayed, and shared in a uniform way. The story has a clean shape because it is a story about a single hero solving a problem nobody else had thought to solve. It is wrong on every clause. Many people had thought to solve it. Several had built working solutions. Two of them, Gopher and WAIS, were widely deployed across the academic internet by the time the web was a year old. One of them, HyperCard, had millions of users on Macintoshes. Another, the various academic hypermedia systems descended from Engelbart’s NLS, had been used in production research environments since the 1970s. And one of them, Xanadu, had been proposed in 1960 and was, by 1990, the subject of a software company funded by Autodesk with a staff of working engineers.
This matters because the empty-room story makes the web’s victory look like discovery, whereas in fact it was selection. Discovery has no losers. Selection does. To say the web won is to say that some specific properties of the web — properties chosen by specific people for specific reasons — beat specific properties of specific competitors that were chosen by other people for other reasons. Recovering the names of those competitors and the names of those properties is the first step in being able to talk seriously about what was traded away. If the web simply arrived from nowhere there is nothing to trade. If it beat real alternatives, there is.
What was actually in the room
A partial list of what was running on networked machines in 1991, the year the first web server came online at CERN:
Gopher had been released in April 1991 at the University of Minnesota and was being adopted, quickly, across the academic internet. Its model was a hierarchy of menus. You navigated by selecting items. Each item could be a document, another menu, a search interface, or a connection to a remote service. There was no markup. There were no inline links. There was a clean, almost severe abstraction: the world is a tree of things you can look at, and you look at them by walking the tree. This was understood, at the time, as a feature and not a limitation. The web’s later association of “navigation” with “following an arbitrary link inside a paragraph” was one specific choice among several plausible ones, and Gopher had made the other choice with care.
WAIS — the Wide Area Information Server — had been released in 1991 by Thinking Machines, Apple, and Dow Jones, with Brewster Kahle as its principal designer. WAIS was a federated search system built on the Z39.50 protocol. Any machine could expose its content as a WAIS index. Any client could query multiple indices in parallel and get ranked results. Search was a peer service: there was no central place. Anyone running a server contributed to the searchable universe. The current internet’s relationship to search — a small number of centralized indices, each owned by a company, each a chokepoint — was specifically not the relationship WAIS proposed.
HyperCard had shipped on every Macintosh since 1987. By 1991 it had been used to build commercial software, educational stacks, games (including the original Myst), reference works, business tools, and personal note-taking systems. Its scripting language, HyperTalk, was natural-language enough that schoolteachers used it. The cultural fact of HyperCard — that ordinary computer users routinely made interactive media for themselves and each other — has no direct equivalent on the web. We rebuilt some of its surface in the web. We have not rebuilt the underlying cultural pattern.
Usenet, the federated discussion system using NNTP, had been operational since 1980. By 1991 it carried tens of thousands of newsgroups, was the locus of most online community life, and had a working governance model based on group creation votes and a hierarchy of moderators. FidoNet, its dial-up cousin, had been founded by Tom Jennings in 1984 and connected tens of thousands of bulletin board systems through store-and-forward messaging. Both worked. Neither had a single owner. The current web’s relationship to community — a small number of platforms, each operating its own walled garden — was specifically not the relationship Usenet and FidoNet proposed.
Behind these were older and stranger things. The Memex, Vannevar Bush’s 1945 thought experiment, was the founding image of hypermedia and had remained a touchstone for forty-five years. Engelbart’s NLS, demonstrated in the 1968 “Mother of All Demos,” had been a working system for two decades. Xanadu had been in development since the early 1960s under various sponsors. Brown University’s hypermedia work, beginning with Andries van Dam’s Hypertext Editing System in 1967, had produced Intermedia by 1985 and was being used to teach undergraduate literature courses. Southampton’s Microcosm system, built by Wendy Hall’s group, was being used for industrial documentation. NoteCards at Xerox PARC, KMS at Carnegie Mellon, Storyspace at Eastgate Systems, Guide on the Macintosh — each was a serious, deployed hypermedia system with users, papers, and a roadmap.
The room was not empty. It was crowded.
The decisive years, 1991 to 1995
In the four years between the first web server at CERN and the Netscape IPO, all of this was decided. It is worth being specific about the timing because the legend foreshortens it. The web was not “released” in any one moment. It became, in stages, the dominant system.
In 1991, the web existed as a single server, info.cern.ch, accessible mostly to physicists. Berners-Lee’s first browser ran only on NeXT machines. The total number of web pages in the world was small enough to enumerate. Gopher had more users that year than the web did. So did WAIS. So did HyperCard, by orders of magnitude.
In 1992, ViolaWWW, Erwise, and MidasWWW gave the web cross-platform browsers. The number of web servers grew from one to about fifty. Gopher’s user base also grew, faster, and Gopher’s developers introduced Gopher+ to add features that addressed early criticisms. The competition was live.
In 1993, two things happened that mattered more than anything else. The University of Minnesota announced in March that it would charge licensing fees for Gopher in commercial settings. The fees were modest and applied only to commercial use, but they were enough to spook institutions and developers, who had been operating under the assumption that internet protocols were free in the sense of unencumbered. CERN, that same year, announced that the web’s specifications would remain public-domain and royalty-free. The shift in momentum was immediate. The other thing that happened in 1993 was Mosaic. Released by NCSA in late January, Mosaic was a graphical browser that ran on Unix, then on Mac and Windows, and that displayed inline images. Marc Andreessen and Eric Bina, the principal authors, made the web look like a magazine where Gopher looked like a card catalog. The number of web servers rose from fifty to about five hundred over the course of the year.
In 1994, Andreessen co-founded what became Netscape Communications. The web servers crossed ten thousand. The W3C was founded at MIT under Berners-Lee. HotWired and the first commercial web banner ads appeared in October. Gopher’s growth flattened and reversed. The hypermedia academic conferences began to absorb the fact that the web was not, technically, a serious hypermedia system, but that it was nevertheless the system that was going to win.
In 1995, Netscape went public on August 9th, valuing the company at $2.9 billion on its first day despite never having shown a profit. The IPO is the conventional marker for the moment the contest ended. By the end of 1995 the web had more than twenty thousand servers, more users than Gopher had ever had, and a commercial gravity that made it the obvious target for any new investment in networked services. Gopher continued to exist; it still does, in a small archival way. WAIS continued to exist for several more years before being absorbed into commercial search products. HyperCard continued to ship with Macintoshes until 1998 and was formally discontinued in 2004. Xanadu’s first public release, ZigZag, came in 1999, after the contest was already over. The academic hypermedia community redirected its work toward the web, often grudgingly.
Four years is a short time. The standard narrative tends to compress them further, into a kind of single instant in which the web simply was, but the people who lived through those years remember the contest as live. Andreessen remembers it. Berners-Lee remembers it. The Gopher team at Minnesota remembers it with particular sharpness, because they had reason to think they were going to win until they weren’t.
What winning looked like at the protocol level
The web’s protocol-level victories were specific and worth naming.
HTTP was stateless. Each request stood alone; the server kept no per-client memory between requests. This was, by the standards of the systems it competed with, almost shockingly thin. Gopher was also stateless, but most of the hypermedia systems were not — Intermedia, NoteCards, KMS, and others all maintained rich session state about who was reading what and where. Statelessness was a feature for scale and a liability for almost everything else. The features the hypermedia systems had — robust two-way linking, version tracking, awareness of other readers — depended on state the web threw away. The web traded those features for the ability to handle, eventually, billions of clients on commodity hardware.
URLs were uncoordinated. Anyone could mint one without asking anyone. This was a deliberate rejection of the rich addressing systems that Xanadu and the academic hypermedia systems had proposed, in which a link was a two-way relationship registered with some authority. The cost of uncoordinated URLs is link rot: the web’s links break, constantly, because nobody promised to keep them working. The benefit is that the system has no chokepoint and can grow without permission. The bet was that the cost of broken links was lower than the cost of a permission system, and over a span of decades the bet has mostly paid off, though it has also been the cause of nearly every recovery effort discussed in Part V.
Links were one-way and lossy. You could link to a page without that page knowing. The linked page could move, change, or vanish, and the link would silently rot. Xanadu had specifically proposed two-way linking and had built parts of the machinery to support it. The academic hypermedia systems had proposed the same. The web rejected both, partly because two-way linking does not scale without coordination and partly because the people building the web found the engineering of one-way links easier to ship. The cost of lossy linking is the loss of citation structure: a link no longer reliably means “this thing exists and says this.” The benefit is that you can link to anything without asking and without negotiation.
HTML was permissive. The original specification was small; browsers were forgiving of malformed input; the layout model was forgiving of authors who did not know what they were doing. This permissiveness was the opposite of what most hypermedia systems had offered. NoteCards, KMS, and Intermedia were strict; their authoring tools enforced structure; you could not produce malformed content because the tools would not let you. HTML let anyone produce anything that vaguely looked like HTML, and the browser would do its best. This was a major reason for the web’s growth. It was also the seed of the layered complexity that would become CSS, then JavaScript, then the modern browser-as-operating-system.
View-source. This was a feature, not a protocol, but it shaped the social pattern more than any single protocol decision. A web page’s source was visible to any reader; the View Source menu let anyone open any page’s HTML and read it. The hypermedia systems mostly did not do this. They presented content; they did not invite you to inspect its construction. View-source meant that the early web was a school. People learned HTML by reading other people’s HTML, then changing it, then publishing. This was the recovery of a teaching pattern that had been familiar from BASIC programs in computer magazines a decade earlier, and it shaped a generation of authors.
What winning looked like at the social level
The protocol-level story is half the story. The other half is social. The web won, at the social level, for reasons that have less to do with technology than with how the technology was governed.
It was royalty-free, definitively, from 1993 onward. CERN’s release of the web specifications into the public domain that year was a specific governance decision and an answer to the Gopher licensing scare. The promise that no one would ever owe money for using HTTP or HTML was a precondition for institutional adoption. The institutions adopting the web were universities, government agencies, and early commercial publishers — entities with lawyers who needed to know whether they were taking on a liability. The web’s answer was no.
It was permissionless. There was no central registry of web servers, no application to fill out, no naming authority other than DNS, which already existed. If you had a machine with a public IP, you could be on the web in an afternoon. None of the hypermedia systems had this property. Xanadu by design required participation in its docuverse, which implied coordination. Intermedia required institutional infrastructure. HyperCard was permissionless to author but had no native network distribution; sharing a stack meant copying a file. The web’s setup cost for a publisher was very nearly zero and that mattered enormously.
It was rendered, badly, on cheap machines. This is sometimes treated as a minor technical detail, but it was load-bearing. Mosaic and the browsers that followed it ran on consumer hardware over consumer modem connections and produced something legible. The hypermedia systems were mostly built on workstation-class machines or required specific operating systems. The web, by contrast, ran wherever there was a TCP/IP stack and a screen. This let it grow into the consumer internet of the mid-1990s, where Gopher and the hypermedia systems would have remained academic curiosities.
It was promiscuous about content. You could put anything on a web page. The hypermedia systems mostly had stronger opinions about what hypermedia was for, and stronger structural commitments — links between propositions, links between argument structures, links between annotations and source. The web treated content as opaque. This was a real loss in the hypermedia sense; it was also, in retrospect, a precondition for adoption. The web did not care what you were saying. It only required that you say it in HTML.
The honest accounting
The chapters that follow do not pretend the web won by accident or by malign luck. The web won because it picked a set of trade-offs that turned out to be correct for the scale it was going to need to reach. The cost of those trade-offs is the subject of the rest of this book, but the trade-offs themselves were not foolish. The people who made them — Berners-Lee at CERN, Andreessen and Bina at NCSA, the early IETF working groups — were not unaware that they were trading away features the competing systems had. They were aware. They made the trade anyway because they correctly understood that the systems they were building had to be lighter, more permissive, and less coordinated than the alternatives if they were going to scale across a decentralized network of millions of independent operators.
But “won” still implies losers. The systems in Parts II and III lost things they were specifically designed to do well. Linking, in the rich sense of two-way addressable persistent connection between named pieces of writing, lost. Authorship, in the sense that the system’s job is to make a reader into a writer, lost. Federation, in the sense that the unit of community is a peer rather than a tenant, lost. Hierarchy, in the sense that information has a place and you find it by following the structure, lost. Search-as-peer-service lost. Stack-based interactive media as a popular form lost. Literary nonlinear writing as a recognized form lost.
Each of those losses is the subject of a chapter. Each is real. Each happened because the web took a different set of trade-offs and the world chose the web.
This is the frame for everything that follows. The web won. It is the right verb because there was a contest, and the things on the other side of the contest were not vapor — they were systems that worked, that had users, that proposed coherent answers, and that, in many specific ways, did things better than what replaced them. Naming what those things were, and what they did, and why they lost, is the work of the next twenty-four chapters.
The contest had a starting point. To see it clearly we have to go back to 1945, to an essay in The Atlantic, written by an engineer who had run American science through the Second World War and was thinking, in the months after the war ended, about what was going to happen to the documents.