Twenty thousand articles later

Before Mastheads was a product, it was my problem. I run my own portfolio of news sites — general news, finance, lifestyle, local editions in five languages — and over the past years the engine behind them has drafted, checked, illustrated, and published more than twenty thousand articles. Not demos. Live pages, on real domains, with my name standing behind them.

That number is the honest answer to the question people actually ask me, which is some version of "does this hold up?" One article proves nothing. Anyone can cherry-pick a good generation. The interesting failures — the ones that decide whether you'd put your real name on the output — only show up in the thousands. Here's what showed up in mine.

What breaks at scale

Everything converges on one skeleton. This is the failure nobody warns you about. Article one looks great. Article forty looks suspiciously like article one: same opening move, same paragraph rhythm, same "In conclusion" landing. No single piece is bad. The body of work is a tell. Readers feel it before they can name it, and Google's scaled content abuse policy is aimed at exactly this shape: many pages, little added value, one mold.

The model is confident about things that aren't true. At volume, hallucination isn't an edge case — it's a certainty. If one article in fifty invents a specific, that's four hundred invented specifics across twenty thousand articles. You will not catch them by spot-checking. Either verification runs on every article, or errors ship on schedule.

Images fail differently than text. Text degrades gracefully; images fail absurdly — the wrong person, the wrong city, a generic stock photo under a specific local headline. And the failure that actually costs money is legal, not aesthetic: a scraped image with someone's rights attached is a problem no amount of good prose fixes.

Source material bites back. When your pipeline reads the open web, the open web occasionally writes back — links you never chose, sneaking from a source document toward your published page. At scale, "occasionally" is a standing threat, not an anecdote.

Silence is the scariest failure. The pipeline that errors loudly is fine. The one that quietly publishes something mediocre — or quietly stops — is the one that ruins a publication. Cadence without oversight is just a faster way to embarrass yourself.

What twenty thousand articles taught us to build

Each of those failures became a design rule, and the rules became Mastheads.

• No two pieces share a skeleton. Every article draws its own seeded profile — length target, structure, opening move, reading level — before a word is written. The variation isn't a style request to the model; it's enforced by the engine. It's the difference between a body of work and a template on repeat.
• Every claim is checked against the sources, every time. Articles are written from real, cited material, and independent quality gates verify the claims before delivery. Any gate can hold any article back for human review. Nothing ships silently.
• An article with no fitting image doesn't publish. It's held for review instead. A gradient or a random stock photo never fills the gap, and web-sourced images can't go live until their rights risk is acknowledged — recorded, not assumed.
• Links that don't trace to a source get stripped. Before publish, any outbound link that isn't one of the article's own cited sources is removed. The open web doesn't get a vote on what your page links to.
• Everything leaves a trail. Every article carries an authenticity trace — the sources it used, the checks it passed, the named editor behind it — and AI content published without human review carries the disclosure the EU AI Act's transparency article expects. Google's own guidance has said since 2023 that AI content is fine and manipulative content is not; the trail is how you stay legibly on the right side.

The lesson under the lessons

The real finding from twenty thousand articles isn't a trick. It's that the writing was never the hard part. Models write competent prose on demand and have for years. The hard part is everything a newsroom wrapped around writers a century ago: verification, editing, standards, accountability, someone whose name is on the door. Scale doesn't remove the need for that structure — scale is precisely what makes the structure non-negotiable, because at volume every rare failure becomes a scheduled one.

That's why Mastheads is built as a newsroom and not a text generator. The twenty thousand articles weren't the product. They were the tuition.