Somewhere in the past two years, you stopped clicking. Not entirely, and not all at once. But if you are anything like the people I keep having this conversation with, a growing share of your questions now end the same way: you ask, something answers, and you move on with your day. The answer arrives as a confident paragraph. It does not arrive as a list of links in blue.
For twenty-five years, the web ran on a particular deal. You typed a few words into a box, a search engine handed you ten results, and you did the choosing. The deal had plenty wrong with it. Those ten positions were fought over, gamed, and sold against relentlessly. But the final act of judgement belonged to you. You looked at what came back and decided whom to believe.
That deal ended, and nobody announced it. No terms changed, and the box looks the way it always did. But the choosing has moved. It now happens before you see anything at all, inside a system that reads on your behalf and reports back. The paragraph you receive was assembled from a handful of sources, and the handful was drawn from thousands of candidates you will never know existed.
I wanted to understand how that handful gets picked. The answers are not the problem; most of them are very good, which is exactly why we stopped clicking. The problem is that who decides what I am told is the kind of question a person ought to be able to answer about their own life. This piece is what I found, and none of it requires a technical background.
Your Question Is Not the Question
The first surprise is that the machine does not go looking for what you asked. It goes looking for several things you did not ask.
Suppose you ask an assistant whether a heat pump makes sense in an old house. You typed one question. The system, before answering, splits it into a small family of searches: how heat pumps perform in poorly insulated buildings, what retrofits cost, what subsidies exist where you live, what owners of old houses complain about a year after installing one. Each of those searches returns its own material. The answer you eventually read is written from the whole pile.
The industry calls this query fan-outMechane definition: The way an AI search breaks one question into several smaller searches, gathers passages from all of them, and writes a single answer — so no page is ever competing for just one phrase. Link opens the full glossary entry., and it changes something fundamental. Under the old deal there was one list, so a source's job was to win a position on it. Now there is no list to win. Sources enter the answer through different doors. One is there because it explained costs plainly, another because it was the only page that mentioned insulation at all, and no two of them were competing for the same slot. The old contest was not made fairer or harsher. It was replaced.
Machines Don't Read Pages
The second surprise is what the machine is actually reading when it reads. Not your favourite writer's argument, unfolding across three thousand words. Machines split pages into passages, a few paragraphs at most, and consider each passage on its own, a practice known as chunkingMechane definition: The way machines break a page into short passages before reading it — so what gets quoted is rarely the whole page, but one paragraph standing on its own. Link opens the full glossary entry.. Whatever gets quoted in your answer is a single passage, standing alone, stripped of everything that was built around it.
Think about what that means for how writing works. A conclusion that lands on page four because pages one through three earned it is, to this way of reading, nearly invisible: the passage containing it does not explain itself. A paragraph that states its point, supports it, and can be lifted whole into a conversation without losing its meaning is a strong candidate. An article earns its citations one self-sufficient paragraph at a time.
You may have noticed a certain style creeping across the web: short declarative openings, the answer first and the reasoning after it, headings that read like questions. That is what adaptation looks like, whether or not the writers could name what they were adapting to.
How the Shortlist Gets Made
So the machine holds a pile of candidate passages, gathered by several searches it invented on your behalf. Most of the pile will not be used. What decides?
No lab publishes its exact recipe, but the visible behaviour is consistent. It favours passages that actually answer instead of gesturing toward an answer. It favours pages that show signs a real, identifiable someone with experienceMechane definition: Experience, Expertise, Authoritativeness and Trustworthiness — the shorthand for whether whoever wrote a page has any standing to say what they are saying. Link opens the full glossary entry. wrote them: a name, a date, evidence of first-hand knowledge. It favours pages where you can tell whether any of this is still true, which mostly means visible dates. And it favours what researchers call information gainMechane definition: How much a page adds that is not already said everywhere else — the measure of whether a machine has any reason to read it rather than the twenty pages saying the same thing. Link opens the full glossary entry., the quality of adding something the other twenty candidates had not already said. The web is full of pages that restate each other. The machine has read all of them, and it has no reason to quote the echo.
Two further facts about the shortlist are stranger, and they matter more than the criteria. The first is that it does not hold stillMechane definition: The tendency of AI assistants to give different answers to the same question asked twice — which makes any single check of what they say close to meaningless. Link opens the full glossary entry.. Ask the same question twice and you will often be handed different sources, because the machine is not consulting a ranking; it is drawing from a distribution. There is no stable position one anymore, no top spot a source could occupy and defend. Anyone who tells you they know where a site ranks in an AI's answers is describing a photograph of a river.
The second is that some of the web simply is not in the pile. A great many modern sites are built so that the page arrives at the browser nearly empty, and the text is assembled afterwards by code running on the visitor's machineMechane definition: A way of building websites where the page arrives nearly empty and its text is assembled by code running in the visitor's browser — code that many machines never run. Link opens the full glossary entry.. Human visitors never notice the difference. But many of the systems doing the readingMechane definition: An automated visitor sent by a search or AI company to read web pages — GPTBot, ClaudeBot, PerplexityBot and their kin — so their systems can answer questions about what they found later. Link opens the full glossary entry. never run that code. What they see is the empty page. The owners of those sites appear in no answers, and no one tells them. To the machines now summarising the world for everyone else, they are blank.
Gatekeeping Without a Gate
Add these mechanics up and something larger comes into view. Being findable by people used to be about popularity, the slow compounding of links, clicks and reputation over years. Being usable by answer machinesMechane definition: A tool that responds to a question with a direct answer instead of a list of links to click — like ChatGPT, Perplexity, or the AI summaries now sitting at the top of Google. Link opens the full glossary entry. is about something else: legibilityMechane definition: Whether the meaning of a page survives being read by software — the plainest test of whether an AI can understand what is actually on it. Link opens the full glossary entry.. Can the machine reach your words, parse them, tell who wrote them, and lift a passage that stands on its own? Legibility, not fame, is becoming the precondition for being part of the public record.
What makes this new form of gatekeeping remarkable is that it has no gate. Nobody is refused. There is no rejection letter, no delisting notice, no moment where you learn that the systems answering everyone's questions have stopped mentioning you. Exclusion just looks like silence, and silence looks like nothing at all.
I want to be fair to what this rewards, because some of it is discipline writers should have had anyway: say it plainly, put your name on it, date it, build paragraphs that can survive being quoted. But some of it deserves suspicion. Prose engineered to be lifted has a flatness you can taste. The answer-shaped block opens with its conclusion because a machine prefers it that way, and it sits on a page built for the quoting rather than for the reader. A web that converges on whatever machines find convenient to quote will be easier to summarise and poorer to inhabit. Both of those things are true at once, and I don't think we get to pick only one.
What We Do About It
It would be evasive to write all of this from a distance, because Mechane is a publication on that same web, read by those same machines. So, in the interest of the plainness this article keeps recommending, here is our own position.
We make ourselves legible on purpose. This site serves a plain-text map written for machines rather than people, and every article carries a summary written under one strict rule: it must faithfully restate what the piece argues, and never embellish it. If a machine quotes Mechane, we want the quote to be true to the article it came from. What we refuse to do is write the articles themselves for the quoting. The essays here open where the argument needs to open, not where a passage-lifter would prefer. If that costs us citationsMechane definition: When an answer engine names the source of what it just told someone — a business, a site, a particular set of words — instead of leaving the answer unattributed. The AI-era version of being recommended by name. Link opens the full glossary entry., the price is acceptable. The body of a piece is for you.
One more thing belongs in the open rather than in a footer. The consequences described here fall hardest on people who will never read an article like this one: the physiotherapist, the roofer, the family hotel, businesses that are recommended by name or not at all and will never learn which. Temu, the Swiss firm that publishes Mechane, builds a small tool called Limen for exactly those businesses, and I own both. That is a conflict of interest on this subject, which is why it is stated here in the body of the piece, and why this article has explained the machinery instead of pointing you at a product.
This piece began with a question: who chose what you just read? For twenty-five years it had a comfortable answer. You did. Ten blue links, and the judgement was yours. That arrangement is gone.
You will not get the choice back. What you can keep is the question, worth asking of every confident paragraph, this one included: where did that come from? A reader who still asks it is a very different reader from one who never thought to.
