The Cheapest Signal

pen any blog in your browser today. The column stretches to whatever thescreen happens to be. The body face is system-ui, or Inter, or a web-safe stackthat terminates in sans-serif. The line spacing was left at the browser’s default.The lines break wherever the text runs out of room. None of these outcomes werechosen. They were arrived at through neglect — or, increasingly, through a templatethat ten thousand other sites are also using right now, assembled from the samedefaults in the same thirty seconds, producing pages that are indistinguishable notbecause they are bad but because they were never decided.

This was always true, to some degree. The web has never been a hospitable medium fortypography. It was designed for documents, not pages, and it has always favoured thefluid over the composed. A stylesheet can name a typeface and set a size. It cannotchoose where a line ends, or hold the column to the 65-character measure that fivecenturies of print practice established as the range where prose becomes easy to readand hard to put down. Those decisions have always been declined by default. What is dif-ferent in 2026 is not the negligence. It is the cause.

AI-generated content tends toward a visual median. This is not a flaw in the models: it isa direct expression of how they work. A language model trained on the web reproducesthe statistical centre of the web. The statistical centre of the web is Inter at 16px, a line-height of 1.5, a column that runs to the edges of a 1280-pixel container, a palette drawnfrom a Tailwind preset, and an em-dash appearing in every third sentence because em-dashes signal a kind of breezy authority in the prose the models were trained on, and themodels reproduce the signal without the authority.1 The result is not ugliness, exactly. Apage assembled from these defaults is not offensive. It is merely indistinguishable fromevery other page assembled from the same defaults this afternoon.

Ten years ago, indistinguishability was the floor: the minimum viable web presence. In2026 it is becoming the ceiling, because the default is free and the floor keeps rising. If acredible-looking page can be generated in thirty seconds, then a credible-looking pagecommunicates thirty seconds of effort. The reader does not calculate this consciously.But they feel it at the level of surface, before a single sentence has landed. What the sur-face communicates is that the words were produced and not placed.

Against this, the signal that matters is not beauty. Beauty is downstream of taste, andtaste is hard to fake but also hard to read at a glance. The signal that is immediatelylegible — that survives even a two-second encounter — is care. And the cheapest, mostlegible evidence of care in written work is the way the type is set.

A reader may not know what a 65-character measure is, or why a page set in Fraunces at19 pixels reads differently from the same text in Inter at 16, or what it means that theline breaks were computed before the page loaded rather than delegated to the browserat render time. None of that knowledge is required. What they notice — what theyregister within a few seconds of landing on the page — is that this text was arranged forthem. That the column was held at a deliberate width. That the face was chosen for thissize and this purpose. That the lines end where they should rather than where theyhappened to. The page communicates, at the level of surface before content: a personmade these decisions. That noticing is the signal.

It is not decorative. A drop cap and a pull quote and a warm accent colour can bedecorative. Setting the measure at 65 characters is not decorative: it is a decision takenin the reader’s interest, because decades of reading research and five centuries of printpractice agree on where prose becomes readable and where it becomes a strain.2 It isthe kind of decision the typographic tradition has been accumulating since Aldus Manu-tius cut his first roman types in Venice at the close of the fifteenth century — quiet calls,made over generations, refined against the evidence of how human beings actually read.The web, remarkably, is the first medium that found it difficult to apply any of them.

Print survived the introduction of movable type without losing the decision point:someone still had to compose the forme. It survived the typewriter, which handed everytypist Courier and nothing else, and still managed to carry Tschichold’s argumentsabout proportion from one generation to the next. It survived the laser printer, whichgave every office a copy of Times New Roman and Helvetica — serious typefaces, both ofthem, however mistreated. It survived desktop publishing, which briefly introducedCaslon and Garamond alongside Comic Sans, and eventually sorted itself. In each case, aperson still had to decide. A decision point existed. The web, by making its defaults closeenough to adequate, removed that point almost entirely: there was no embarrassment ofbadness to force a reconsideration, and so the reconsideration never came.

What is new is not the problem. What is new is the contrast. When every automaticallygenerated page looks the same, the page that was set stands out — not by shouting, butby being quiet in a different register. The reader may not be able to name what isdifferent. They will feel it. The column holds. The lines end where they should. Thetypeface earns its place at this scale. The page belongs to its writer the way a printedbook belongs to its printer: not through ornament, but through the evidence ofparticular choices, made deliberately, for the reader’s sake.

You can generate an em-dash at scale. You can instantiate a template in thirty seconds.What you cannot generate is the accumulation of care that shows in a typeset page,because that accumulation is the evidence of someone having thought, paragraph byparagraph, about whether this is right. The template cannot do that. The thirty-secondpage cannot do that. Thinking takes time, and time spent is exactly what the readerreads in the surface of the page.

This essay was set with a static site generator I built for this purpose — one thatcomputes line breaks at build time and locks them before the page reaches the reader. ¶

The question worth asking is not how to make your writing look better. It is whether youwant your writing to visibly belong to you. A well-set page is the evidence of a personhaving thought about a reader. That thinking accumulates in the leading and themeasure, in the typeface chosen and held, in the column that does not drift with theviewport. It is quiet. It persists. And in a year when every other page was assembledfrom defaults, that evidence — for the reader who arrives and pauses and stays — isenough.

Footnotes

  1. The em-dash observation is testable: a corpus comparison of AI-generated blog prose againsthuman-written magazine prose will show a consistently higher ratio of em-dashes to total punc-tuation in the AI output. The models reproduce the surface feature of authoritative prosewithout the underlying reasoning that earns it. ↩

  2. Robert Bringhurst places the readable range at 45–75 characters per line, with 66 as the ideal, inThe Elements of Typographic Style §2.1.2. Studies in digital reading (Dyson, 2004; Ling & vanSchaik, 2006) support moderate line lengths for both reading speed and comprehension. Pil-crow’s default of 65ch sits at the centre of that range. ↩