The New York Times set its last hot-metal page in July 1978.
The Linotype had been in newsrooms since 1884. Hand-set type went back five centuries before that. Both ended inside a decade, replaced not by something more beautiful but by something good enough for a fraction of the price. “Good enough” had been the part the compositors owned. Once it became table stakes the rest of what a typesetter knew was overhead. Newspapers that had paid for a room full of compositors for a century rationally stopped inside a fiscal year.
The same shift is happening to documentation platforms now.
For most of the last decade, paying a docs platform was the rational answer to the same question every printer’s customer answered in 1900: who is going to make this look like a real publication, and what is that worth. The bundle was a renderer, a theme, search, a sidebar tree, a few clever components, a domain. It was not trivial to assemble in 2018; it was, plainly, the moat. Mintlify won the category by closing the gap cleanly, and the rest of the field built variants of the same wager. They were charging rent on quality.
The moats around typesetting collapsed between 1978 and 1988 under two pressures. The same two are hitting docs platforms now.
The polish floor moved
Compugraphic was the largest typesetting equipment manufacturer in the world in 1985. It was absorbed into Agfa in 1988 and wound down through the decade that followed. The thing that took its market was not a more beautiful machine. It was Aldus PageMaker plus an Apple LaserWriter, producing a page that was good enough for a fraction of the price. The bar Compugraphic built itself around moved underneath it.
Open source documentation tooling has had its PageMaker moment without anyone declaring one. Starlight, Astro, Docusaurus, and Nextra ship at the bar the paid platforms set. AI-assisted theming closes whatever sliver of gap remains. You can ask a model for a Stripe-grade sidebar treatment against your existing pages and it produces one in an afternoon. The bespoke version of every component a paid platform sells now lands as a pull request from a model that bills in cents per request.
The work of clearing the polish bar went from a budget line to a config commit.
The platforms’ response has been to ship more features around the moat, on the implicit assumption that the moat is still doing the work. It is not. What teams who migrate off Mintlify or GitBook in 2026 notice immediately is that the thing they were paying to keep was never the polish. It was the proprietary MDX dialect that held the polish in place.
The dialect is the lock-in. Callouts, tabbed code groups, accordions, cards, the frontmatter conventions controlling the navigation: every one is a stake driven into the docs that has to be pulled out before the project can leave. A hundred pages in vendor MDX is a hundred migrations. What looked like quality was a contract with a renewal clause embedded in the syntax of every page.
The second reader arrives
What killed Linotype in two decades was not just PageMaker. It was also that the audience moved partly to screens, where the typographic finesse the compositor had been trained in stopped being decisive. The micro-adjustments at the top of the craft, leading, hyphenation, optical kerning, were not what a screen reader was registering, if they were registering them at all. To the new audience, the difference between a master compositor and a passable one flattened.
Every URL on a documentation site now serves a reader past even that threshold. An LLM fetches the page with a model-shaped user agent and reads what comes back. The hero animation, the carefully chosen colour palette, the typographic spacing the design team spent six weeks on: none of it is there. A token stream, the text of the page, the substrate underneath what the human evaluator saw. “The Documentation Problem” is the longer version of that shift. What matters here is the shape: the reader the platforms have spent a decade designing for is no longer the reader who decides most of the API calls landing against the project.
A decade of polish optimisation turns out to have been optimising for the audience that no longer dominates the load.
The new signal
The interesting part is not that the platforms lost the moat. It is what replaces the moat as the visible signal of a serious project, and whether anything that fits inside a subscription can be the replacement.
In 1990 the surviving design profession split. At the top, design moved from execution to identity systems, custom type, the work no teenager with PageMaker could do, ever. At the bottom, rendering pages was absorbed into office software and stopped being a profession at all. The middle, the typesetting bureau that charged you for one page of presentable type, disappeared. There was no longer a market for renting access to a craft anyone could buy outright.
Documentation is heading the same way. At the top, good docs will mean editorial judgement, a writer’s voice, examples that earn their place in production, a contract surface that survives an agent grounding fifty calls in a session against it without getting one wrong. None of that is rentable. You write it or you don’t have it. At the bottom, rendering Markdown into a static site at the field’s bar will ship with any modern stack, no decision involved, no fee attached. The middle is the typesetting bureau in 1995, still in business, unable to explain to a new customer why anyone would sign on now.
The new signal of a serious project is not how its docs look. Every project’s docs are about to look fine. The signal is the substrate underneath the hero: the surfaces an agent can read, words written rather than assembled, a URL the project owns. These are not features. They are conditions, true before the marketing surface gets built or never true at all. No platform sells them, because no platform can.
After the moat
Open the last invoice from whichever platform the project is on. The line items are real. The world they made sense in is not. Cancel or don’t; the question every team will answer this year is no longer which platform looks best. Every platform’s output looks fine. The question is what the docs are doing underneath while the rendering chain becomes a commodity.
What ended typesetting did not stop the day the last hot-metal page was set. The equipment plants kept shipping their catalogues. The buyers kept signing renewals. The receivers came at the end of the next decade and decided nobody could plausibly explain who was supposed to keep buying. The bills got paid until somebody decided to stop sending them.
The moat is gone. Nobody has told the invoicing system.