We publish articles no human will ever read. On purpose.

There is a piece of writing on dev.to with our name on it. It was drafted carefully, edited, proofread and published in the full expectation that no developer would ever read it. That was not a failure of the piece; it was the specification. It exists so a crawler would record that one domain linked to another, and so a language model ingesting that page would file our name next to a topic. If a person reads it, that is a rounding error, and a pleasant one. None of this is a confession of something unusual. It is a description of how a large and growing share of the web is now produced, and almost everybody producing it is entirely clear-eyed about what they are doing.

The inventory

Every row is something we have built, bought, been asked for, or deliberately refused. None of it is hypothetical and none of it is rare — most of it is standard practice at agencies that would describe themselves, accurately, as reputable.

The majority reader stopped being a person

For most of the web’s history, writing for machines was a small tax on writing for people. You added a title tag, you wrote a meta description, you kept the URLs tidy, and the rest of the page belonged to a reader. That proportion has inverted, and it inverted quickly enough that most of the vocabulary has not caught up — we still say “content” and “audience” as though the audience were the point.

Thales and Imperva put automated traffic at more than half of all web traffic in 2025: 53%, against 47% human, and still rising. That counts requests rather than attention, and it is still the most clarifying number in this business. For a typical page, the median visitor is a program. Not an unfortunate side effect of publishing — the majority case.

The industry’s response to that fact was the rational one. If most of what arrives is machinery, write for the machinery. What is worth noticing is not that it happened, but how completely it stopped being controversial while it did.

Nobody in this industry is confused about what they are doing

It would be more comfortable to describe machine-directed writing as a side effect — you write for people, and the crawler happens to read it too. That is not what is happening, and pretending otherwise gets every incentive in the business wrong. When a marketing team commissions two hundred city pages, nobody in the room believes two hundred towns are waiting for them. When a founder pays for a guest post, the word in the brief is “do-follow”, not “readership”.

The scale is easiest to see from the far end. Ahrefs ran its full index — around 14 billion pages — and found that 96.55% of them get no search traffic from Google at all, with a further 1.94% getting between one and ten visits a month. Some of that is ordinary failure. A great deal of it is text published by someone who already knew the odds, because the page was never the point: the link on it was, or the keyword was, or the entity mention was.

That is the honest frame, and it is worth stating before any of the tactics. Not an accident, not a technicality, not an emergent property of the algorithm — a deliberate, budgeted, staffed industry manufacturing text whose intended reader is a program. Ours included, which is why the inventory above has us in it.

Three reasons text gets written for a machine

The inventory looks like a grab bag until you sort it by what the writing is trying to achieve, at which point it collapses into three motives. To be found: robots.txt, the sitemap, internal links, programmatic pages — everything whose job is to get a URL into an index. To be understood: schema, llms.txt, the FAQ shape, the comparison table, naming your own brand inside a section instead of only in the heading. To be vouched for: links, mentions, anchor text, and every word you publish on a website that is not yours.

The first two are plumbing. They are cheap, they are mostly harmless, and a competent developer does them once. Nobody has ever been misled by a sitemap.

The third motive is where the budget goes and where the ethics live, because it is the only one that operates on other people’s property — their publications, their forums, their comment sections, their attention. Everything uncomfortable in this article is downstream of that distinction.

The backlink post is the purest form of it

Take the cleanest example, which is ours. A short technical piece goes up on dev.to or Medium. It is written properly, because we do not publish rubbish with our name on it, and it carries one or two links back to a page we want ranking. Realistic readership: a handful of people, none of whom will become customers. The piece is not failing to reach them — it was never aimed at them. It is aimed at the crawler that records an edge in the link graph, and at the model that will later be asked who does this kind of work.

Once that is said out loud, the obvious question is why anyone bothers writing it well. The answer is that the machines got better at telling. Thin keyword-shaped filler on a free publishing platform is worth close to nothing now, the platforms increasingly nofollow or strip it, and the ranking systems have spent a decade learning what it looks like. What still works is a piece a publication would have run on merit — which is an expensive sentence, because it means the only reliable way to game the system is to stop gaming it.

That is the whole argument for doing the paid version properly instead of cheaply. A sponsored article on a real trade publication is four things at once — a link, a branded mention, a branded anchor, and a block of text sitting in a retrieval index that describes what you do in somebody else’s voice — and it is the one row on the inventory that also, incidentally, gets read. That is what our advertorials and link building work is for, and SEO Advertorials is the packaged version: a publisher shortlist matched to your niche, anchor text agreed before a word is written, three placed pieces, indexation checked afterwards.

We will also say what it is not. It will not make a bad website rank. It does not work in volume. And anybody offering you fifty placements a month is selling you the row above this one, at the price of the row below it.

The machines reward other people’s websites, which is why the writing left yours

This is the finding that explains the shape of the whole industry better than any manifesto. When Ahrefs correlated brand visibility in AI Overviews against every signal it could measure across 75,000 brands, the three strongest were all off your own site: branded web mentions at 0.664, branded anchor text at 0.527, and brand search volume at roughly 0.33 to 0.39. Backlinks — the thing an entire market sells by the unit — came in at 0.218. Brand mentions on YouTube correlated harder than anything else measured, at about 0.737.

Correlation is not causation, and that deserves saying plainly rather than in a footnote: brands that get written about are usually brands doing several other things right, so a mention is partly just evidence that you exist. But the direction is consistent across every independent study of it, and it points away from your domain. If the strongest available signal about you lives on pages you do not own, then the rational place to put a content budget is somebody else’s website — and that is precisely where it went.

It also explains why the machine-directed writing moved off-site rather than disappearing. Ten years ago the keyword-stuffed paragraph lived at the bottom of your own category page, where it embarrassed only you. Now it lives on a publishing platform, a forum, a listicle and a press wire, where it is somebody else’s problem and everybody’s corpus.

Where it stops being harmless

Follow the loop all the way round. A model is trained on pages, and retrieves from pages. Marketers learn which pages get retrieved. They produce pages shaped for retrieval. Those pages become part of the corpus the next model is trained on. Roughly half of all new articles on the web are already machine-generated — Graphite classified 65,000 English-language URLs out of Common Crawl and dated the crossover to the end of 2025 — so a growing share of what the machines read was written by machines, for machines, about topics chosen by machines.

Meanwhile the human end of the pipe keeps narrowing. SparkToro’s reading of Similarweb clickstream data puts 68.01% of US Google searches ending without a click in early 2026, up from 60.45% in 2024; per thousand searches, the open web went from 374 clicks to 276. We are producing more text than ever, for fewer readers than ever, in a format optimised for the thing standing between us and them.

That is the dystopian reading and it is not paranoid — it is the arithmetic of the last two years extended by exactly one step. Where the pessimists overshoot is the conclusion. This is not the web dying; the web is busier than it has ever been. It is the web acquiring a second audience that is larger, more literal and far less forgiving than the first, and an industry quietly reorganising itself around that audience without ever announcing it.

What survives the loop

The finding that keeps this from being hopeless comes out of the same Graphite work. About half of new articles are machine-written, and in that same dataset 86% of the articles actually ranking in Google Search were written by people. The volume is synthetic. The selection, so far, is not — which means the loop has a filter in it, and the filter is still pointed roughly where you would want it pointed.

The technical baseline underneath all of this is also boringly cheap. Vercel and MERJ instrumented a very large network of sites and found that none of the major AI crawlers execute JavaScript at all: GPTBot, OAI-SearchBot, ClaudeBot, PerplexityBot, Meta-ExternalAgent and Bytespider download your script files and never run them, with Gemini the exception because it rides Google’s own infrastructure. So the single highest-value technical act in this entire article is server-rendering the sentences you want quoted — and most sites still get it wrong.

After that, the scarce input is a claim that is true, specific to you, and not already inside the model. Everything else on the inventory is distribution, and distribution without such a claim is the machine-written half of the corpus with a human name on it. What AI crawlers actually send back has the traffic side of this, and how long it takes to be cited at all has the timeline.

Our rules for this, since we are in it

We do most of what is on that inventory. It would be dishonest to write this article from the outside, and the list is not a survey of other people’s behaviour. So here is where we actually draw the lines, in case it is useful as a template for asking your own agency.

Every piece has to be defensible to whichever human does find it, because one occasionally does, and because a piece written to be worthless usually is. Nothing goes on a third-party site that we would not put on our own — the byline is the same, so the standard is the same. No volume plays: if a tactic only works at fifty units a month, it is not a tactic, it is a bet on not being caught, and that bet has been losing for about three years.

And one more, which is the reason this article exists at all. Say what the thing is. A client paying for a sponsored article is buying a branded mention in a retrieval index and a link in a graph, not “exposure” — and the moment an agency will not put it that plainly, you are being sold the cheap row rather than the expensive one. Our GEO guide covers the technical half of this honestly, and the llms.txt question covers the file everyone now asks about, including whether it does anything yet.

Sources

Checked September 2026. Two of these are correlation studies, which show direction rather than cause, and one measures requests rather than people — each is described as what it is where it is quoted.

• Imperva / Thales — Bad Bot Report 2026: bots in the agentic age ↗

  Automated traffic at 53% of all web traffic in 2025 against 47% human, up from 51% the year before. A measure of requests across Imperva’s monitored network, not of reading.

• Ahrefs — 96.55% of content gets no traffic from Google ↗

  The full Content Explorer index, around 14 billion pages: 96.55% with no Google search traffic and a further 1.94% receiving one to ten visits a month. Published 2023; the ratio has not improved since.

• Search Engine Land — Google zero-click searches reach 68% in early 2026 ↗

  The 68.01% figure and the 374 → 276 clicks per thousand searches comparison, from SparkToro’s analysis of Similarweb clickstream data for January to April 2026 against a 2024 Datos baseline.

• Ahrefs — An analysis of AI Overview brand visibility factors (75,000 brands studied) ↗

  Spearman correlations between brand visibility in AI Overviews and every measurable signal: branded web mentions 0.664, branded anchors 0.527, brand search volume around 0.33–0.39, backlinks 0.218, YouTube mentions about 0.737.

• Graphite — More articles are now created by AI than humans ↗

  Classification of 65,000 English-language URLs from Common Crawl, the crossover at the end of 2025, and the finding that 86% of articles ranking in Google Search were human-written.

• Vercel — The rise of the AI crawler ↗

  The Vercel and MERJ measurement that no major AI crawler executes JavaScript, including the share of each crawler’s requests that download script files without running them.