By Remco Livain in AI — 02 Sep 2026

I shipped a magazine section in a day. The work was deciding what not to index.

Generation is cheap now, so the scarce input is refusal. What I noindexed, what I labelled, and what I took back out of an AI-assisted editorial launch.

A directory I run now has an editorial section. A magazine hub, four long articles, an archive page, its own design system, thirty-four illustrations, structured data on every page. Concept to live in about a day.

The building is not the story. That part is unremarkable in 2026 — it is simply what the tooling does now, and anyone telling you otherwise is selling something. The story is that every decision worth anything that day was a decision not to publish something.

Four things I took back out

The concept included a product comparison — a sortable table of robotic mowers with specs, scores and a recommendation. The data was placeholder. Invented models, invented numbers, entirely convincing. It went live behind noindex, and it stays there until the testing is real.

The concept also included an archive page listing eight previous issues. There is one issue. Eight back issues is perfectly good design fiction in a mock-up and a straight lie in production, so that page is noindexed too, with the download links relabelled "PDF to follow" rather than pointing at nothing.

The design carried a publication date three weeks in the future, because that is when the designer imagined it shipping. Live content does not get a future date. Changed to the actual day.

And the design confidently stated the directory holds 831 verified businesses. It holds 829. The build script now pulls the real figure at publish time, so the number cannot rot into a stale boast the moment someone opts out.

None of that was caught by a tool

This is the bit I keep coming back to. Nothing in the stack objected. There was no warning, no lint failure, no red box. The comparison table rendered beautifully. The archive page was arguably the nicest layout of the six. Every one of those pages would have published cleanly and looked professional doing it.

When the AI is very good, humans have no reason to work hard and pay attention.

That is Ethan Mollick in Co-Intelligence, and it is the whole problem in a sentence. The failure mode of cheap generation is not bad output. Bad output gets caught. The failure mode is plausible output arriving faster than anyone's appetite to interrogate it — which is the same dynamic I wrote about in the twenty-page document that makes its problem your problem, except here I would have been shipping it to the public rather than to a colleague.

Google's own guidance on generative content lands in the same place: the production method is not the issue, and using AI to make many pages without adding value for users is. A comparison table of fictional products against a real buying question is not a grey area. It is the example.

Labelling, before anyone makes me

Every generated image carries a caption saying it is generated. Not because a regulator has arrived — although the EU's transparency rules on synthetic content are heading exactly here — but because of what the alternative does to a directory.

The whole product is a claim that these businesses are real and checked. Illustrating that with photorealistic fiction and no label teaches a reader that they cannot tell which parts of the page are true. You do not get to be trusted about the register data and casual about the pictures. It is one impression, and it is the same reason I deleted my own testimonials and watched my rating go down.

I left one tension unresolved and I would rather say so than pretend the day was tidy. My own rule is that AI imagery does not belong on pages where a reader is choosing between real businesses. The new buyer's guide is such a page and it carries generated images anyway, clearly labelled. It is flagged. It is not fixed. It is the kind of thing that gets quietly normalised if you don't write it down.

The unglamorous failure that nearly buried all of it

The pages went live and stayed invisible. Published, reachable by direct link, absent from the sitemap. The SEO plugin caches its sitemaps as files on disk and does not invalidate that cache when pages arrive through the API rather than through the admin screen. Flush the cache, and the pages appear.

File that next to the schema my plugin was quietly emitting wrong. Every publishing stack has a layer that assumes a human clicked something, and it will only ever fail silently.

The scarce input changed

Two years ago the constraint on a section like this was production: who writes it, who designs it, who builds the templates. That constraint is gone, and it is not coming back.

What is left is editorial judgement — knowing that a table of invented specs is worse than no table, that a fictional archive costs more trust than it buys authority, that a number in the design has to become a number from the database. None of that is a prompt. All of it is knowing what a reader is entitled to assume when they land on your page.

Generation got cheap. Refusal is the job now.

Sources & further reading

External
Google Search Central: guidance on AI-generated content
Google Search Central: spam policies, including scaled content abuse
Google Search Central: blocking indexing with noindex
EU AI Act, Article 50 — transparency obligations for synthetic content