Microsoft Word did not win on the writing

I started out writing manuals, and then online help, and then the systems the help lived in. The tools we used for that were built for the job and they were very good. They lost anyway, and not on the writing.

Geert Valère Allegaert · October 2026

Support for long documents stored across multiple files, while still supporting shared formatting.

David J. Murray, Frame Technology and FrameMaker, IEEE Annals of the History of Computing, 2019

That sentence is from the man who founded the company, listing what his program could do. It reads like nothing. It is the whole argument of this note. A manual is not a long letter; it is a set of files that have to behave as one book, and almost everything the profession knew was packed into that distinction.

Frame Technology was founded in 1986 and FrameMaker 1.0 shipped in March 1987, on Sun workstations. WordPerfect came out of Satellite Software International in 1980, reached MS-DOS in 1982, and by 1986 was the best-selling word processor in the United States, with a share that reportedly passed seventy per cent of the DOS market before the decade was out. Those were the tools. Then Microsoft Word took the market, and the specialist knowledge that the specialist tools encoded went quiet.

What the specialist tool knew

Adobe, which bought Frame Technology in 1995, still publishes a document comparing FrameMaker with Word, and the feature list is an education in itself. A book file that combines many documents hierarchically and renumbers the whole thing when one chapter moves. Cross-references in which any paragraph can be a target, updated on build, so that “see page 143” is never wrong. Tables of contents and indexes generated from the text with no practical limit on depth. Conditional text, which shows and hides marked passages so that one source produces the manual for all six models of the machine. Separate catalogues for paragraph and character styles, so that structure and appearance stay apart.

Read that list again as a claim about the world rather than as a list. It says: a document is a hierarchy, not a stream. Its parts have names and can be referred to. Its appearance is a separate question from its structure. One source can produce many outputs. Those four propositions are the intellectual content of technical documentation, and in the late eighties you could buy them in a box.

A general word processor makes the opposite claim, mostly by not making it. A document is one flow of text with formatting attached to the characters. That is a perfectly good model for a letter, and it is why Word won the letters. It is also why a four-hundred-page manual in Word becomes a thing that two people maintain by remembering where the problems are.

Two numbers from 1996

Here is the part that changed how I think about this, and it took thirty years and a court record to find it. In the week Corel bought WordPerfect, the Associated Press printed two sets of figures from 1995. Among word processors, WordPerfect still held 47.9 per cent against Word’s 44.5. Among office suites, Novell’s WordPerfect package held 5.4 per cent against Microsoft Office’s 86.5.

Those two lines sit in the same article. As a product, WordPerfect was still ahead. As a bundle, it had already been annihilated. The competition had moved to a different unit of sale, and the better word processor was not in that competition at all.

The prices tell the same story faster. Novell bought WordPerfect in June 1994 for about 1.4 billion dollars. Nineteen months later it sold it to Corel for something in the order of 185 million, counting royalty commitments. Nothing happened to the writing in those nineteen months. By January 1997 Microsoft was quoting IDC in its own press release: ninety per cent of the suite category, fifty-five million users of Office applications.

FrameMaker went the other way and survived, by retreating. Adobe still sells it. It is Windows-only now, subscription-only, and there has been no new major release since September 2022. It is a specialist tool for a specialism, which is a decent outcome and a long way from being the thing everybody used.

What was never proven

It would be easy to end the story there, and it would be wrong, so this section is the correction to my own instinct.

Novell did sue. In 2004 it took Microsoft to court over Windows 95, alleging, in the Tenth Circuit’s own summary, that after sharing the interfaces in the June 1994 beta Microsoft “reversed course in October, indicating to ISVs that they could no longer rely on the previously published APIs.” The case ran to an eight-week trial in 2011 and the jury hung. The judge then entered judgment for Microsoft as a matter of law, the Tenth Circuit affirmed in September 2013 in an opinion by Neil Gorsuch, and the Supreme Court declined to hear it. The holding was that a monopolist has no antitrust duty to help a rival: whatever happened, it was not unlawful.

And there is a serious argument that bundling is not the explanation at all. Stan Liebowitz and Stephen Margolis went through the contemporaneous software reviews and concluded that Word won because reviewers judged it the better program, not because of lock-in. I do not find that fully convincing against the two 1996 numbers, but it is argued from evidence rather than from grievance, and anyone writing about this owes it a mention.

So the honest version is narrower than the one I would tell in a bar. That WordPerfect and FrameMaker were displaced is a fact. Why is contested. What needs no theory is the shape of it: by 1995 the question had stopped being which word processor is better.

Where the craft actually went

It did not die with the tools. It moved out of the application and into markup, which in hindsight was the only place it could survive.

SGML became an ISO standard in 1986, the same year Frame Technology was founded, and it took the four propositions above and made them independent of anybody’s program. Then in March 2001 three people at IBM — Don Day, Michael Priestley and Dave Schell — published an architecture called DITA, built by the user-assistance teams of IBM, Lotus and Tivoli, in which the unit of information is not the document but the topic: “a unit of information that describes a single task or concept or reference item.” OASIS approved it as a standard in June 2005.

There is a coincidence worth noticing and not overstating. The same company that produced John Carroll’s work on minimal instruction, a decade earlier, produced the markup in which instruction gets cut into task-sized pieces. I have found nothing claiming one descends from the other, and I am not going to invent a lineage. But both came out of IBM’s user-assistance world, and both say the same thing from different ends: people do not read the manual, they come for one task, so the task is the unit.

That is also roughly when the job stopped being about documents and started being about systems, which is how I ended up building online help rather than writing it. Not a promotion. A relocation. What became of that architecture is its own story, and I have told it separately: it is still here, still has paying customers, and has not had a new version since 2015.

What follows is what I take from all of this when choosing a tool now.

It keeps happening

The reason this is not nostalgia is that the same trade is on offer every year, and the web version is the one I work in now. A page builder will let anybody lay out a page, and in exchange your content stops being content: it becomes a tangle of nested containers that mean nothing outside the builder that made them. That is the Word bargain exactly — convenience now, in return for a format only one program can read.

What makes it bearable this time is that the open format won the layer underneath. HTML is the thing everybody has, and it is a structure language: a heading is marked as a heading, a list as a list, whatever drew it. Keep the content in that and you keep the four propositions. You can lose a tool and not lose the work, which is the only durable version of this whole argument.

And one honest correction to the mood. Miles Kimball has argued that technical communication is in a golden age precisely because it escaped the institution — that the departments and the standards were a cage as much as a home, and that the work is now everywhere instead of in one building. I think he is largely right, and I am an instance of it: nobody employs me to write a manual, and I am still doing a version of the same job. What was lost was not the craft. It was the agreement that the craft needed its own tools, and that is the thing nobody has got back.

Read the originals

• David J. Murray, Frame Technology and FrameMaker — IEEE Annals of the History of Computing, 2019. The founder’s own account, and the best place to see what the tool was for. Read this one if you read one.
• Corel buying WordPerfect — Associated Press, 1 February 1996. The two numbers, in the same paragraph, reported without anybody noticing what they meant.
• Adobe, FrameMaker vs Word for Technical Documentation — a sales document, and still the clearest statement of what a long-document tool does that a word processor does not.
• Novell, Inc. v. Microsoft Corp. — Tenth Circuit, 23 September 2013. The allegation in the court’s own words, and the reason it failed. Unexpectedly good prose.
• Stan Liebowitz and Stephen Margolis, the word processor chapter of Winners, Losers & Microsoft — the argument that Word won on merit. Read it before you settle on the bundling explanation.
• The announcement of DITA — Cover Pages, 16 March 2001, on the IBM article by Day, Priestley and Schell. Where the topic becomes the unit.
• SGML, ISO 8879:1986 — Library of Congress format registry. The standard that made the ideas independent of any program, and the ancestor of HTML.
• Microsoft announces Office 97 — 16 January 1997. Ninety per cent of the suite category, in Microsoft’s own words.
• Miles A. Kimball, The Golden Age of Technical Communication — Journal of Technical Writing and Communication, 2017. The case that leaving the institution was a gain. The best argument against this note.