The future of SEO is invisible, and writers shouldn't have to see it

Before you have written a single line, every CMS editor asks you questions. Title, slug, description, tags, category, language. Empty fields waiting, which basically ask you to know what the text is before it exists.

For years I found that normal. Then, while working on Sandy, the CMS I’m building, I realised those fields aren’t questions for the writer. They are questions one machine asks another, and they pass through a person’s hands only because until now there was nobody else to ask.

Metadata is a conversation between machines

Almost nobody reads the meta description. Google reads it, and decides whether to use it in the snippet or rewrite it. The slug is for a crawler and a shared link. Tags feed related articles, the category feeds a menu, and structured data feeds whatever system comes along to interpret the page.

In 2011 Google, Bing and Yahoo launched schema.org together for exactly this reason: a common vocabulary that sites could use to describe themselves to search engines. It was an agreement between machines. The job of translating a text into that vocabulary, though, stayed with us, and for the most part technical SEO has been just that: acting as interpreters.

Social algorithms already do it

There is a place where nobody fills in metadata and content still reaches the right people: feeds. On TikTok you don’t choose what to watch, and those who publish don’t really choose who sees it. The system looks at the video, looks at how you watch it, and decides. There, SEO doesn’t exist, because the machine has absorbed it.

The mechanism works extremely well. The trouble is what it is optimised for: the time you spend in the app. It finds you interesting things as a means of keeping you there, and the relationship with your attention becomes parasitic.

I’m interested in the same mechanism with a different purpose: helping someone find what they’re looking for, and then letting them go.

It’s already happening on the web

Search is moving in that direction, whether we like it or not. Google answers with AI Overviews before the links, and in July 2025 the Pew Research Center measured that people who see an AI summary click on a traditional result in 8% of visits, against 15% for those who don’t. ChatGPT and Perplexity read pages on behalf of whoever asks the question, and the person asking often doesn’t even open the site.

Something is moving on the sites’ side too. In 2024 Jeremy Howard proposed llms.txt, a file that explains to language models how a site is built. In 2025 Microsoft introduced NLWeb, a project led by R.V. Guha, the same person behind RSS and schema.org: the idea is that every site can answer in natural language to people and agents.

Put in a row, these pieces sketch a web in which what we used to call SEO becomes an exchange between systems. The site describes itself, the engine or the agent interprets, the person receives an answer. Humans stay at the two ends, the one who writes and the one who searches, and the machines work in between.

I don’t know how long it will take, or whether it will go exactly like this. What I do know is that asking writers to fill in by hand fields designed for machines is already an old habit.

The “Next” button

The latest feature of Sandy came from this idea. The first version was neither fish nor fowl: the AI filled in half the fields while you wrote, with little buttons scattered across the page that suggested a title or a tag. The AI did its job. The problem was that it crept into the writing, and the writer ended up doing two jobs at the same time.

Writing and publishing call for different states of mind. The first wants space and no judgement, the second wants order. Mixing them makes both worse.

Now, when you create a text, Sandy opens a full-screen blank page. No sidebar, no mandatory title, no language to choose. If you want a title, you write it as the first line. The assistant stays closed in a bubble in the corner and speaks only if you call on it.

When the text feels ready you press “Next” and the page gives way to Sandy. It reads what you wrote, understands what kind of content it is and does the editorial work: title (if you haven’t written one), page address, description, category, tags, cover. Next to the conversation you see the real preview of the page, not a table of fields, and at the end Sandy explains in a few lines what it decided and why. If something doesn’t feel right you write it in the chat (“a shorter title”, “remove that tag”) as you would say it to a colleague. If the site is in more than one language, it asks you whether to translate.

That button is the border between work for people and work for machines.

Two guarantees so you can delegate with peace of mind

The body of the text stays yours. Sandy decides everything around it, but it doesn’t touch the words you wrote. That is the part meant for a human being, and a human being writes it. The rest is metadata, things that talk to other systems, and it is right that a system takes care of it.

Nothing is final. Until you confirm, everything is a draft. If you cancel, what Sandy did disappears, translations included, and you find your text exactly as it was. Even after you confirm, nothing goes online until you decide to publish.

The freedom to let go needs the certainty of being able to go back.

Why I removed autosave

This was the least obvious detail. It looks like a step backwards, since every modern editor saves at every keystroke. But that “saved at 2:32 PM” flashing while you write reminds you that the system records every word, including those you will delete in a minute. It makes you feel watched.

In Sandy you save when you want, with ⌘S. The draft stays in the site, outside the published content, and you find it in the list of articles in a “to finish” section.

These are small choices, and they all go in the same direction: the machine works a lot and the writer hardly notices.

What is left for the writer

If SEO becomes invisible, what remains visible is the text. Those who know search keep being useful, but their work moves further upstream, to what is worth writing and for whom, and further downstream, to how the site presents itself to the systems that will read it. The middle step disappears: the person who fills in a form on behalf of an algorithm.

The risk is the same as with social media: that those who build these systems optimise them to retain rather than to help people find things. That is why I like building them myself, at least the piece that is mine to build, the one where a person writes and a machine takes care of the rest.

If you want to understand how to take your site in this direction, tell us about your project.