Fluency is not evidence. Decide where your facts come from.

Every sentence an engine writes arrives with the same confidence. The one drawn from your pricing page and the one inferred from nothing at all are typographically identical, and no reader — including the person approving the post — can tell them apart by looking. A better model does not fix that. What fixes it is deciding in advance where facts are allowed to come from.

A guess and a fact arrive in the same sentence

A human writer leaks uncertainty. They hedge, they stall on a number they half-remember, they leave a bracketed note to check something later. Those tics are irritating to edit and enormously useful, because they mark the places where the draft is load-bearing and unverified.

Generated prose has no such tells. Fluency is uniform across the whole piece, so the sentence carrying your actual renewal terms reads exactly like the sentence that invented them. The failure is not that an engine is wrong more often than a tired freelancer. It is that when it is wrong, nothing in the text points at it.

That is why we treat sourcing as a configuration problem rather than a writing problem. A blog publishing every morning cannot rely on someone noticing the one sentence that was never checked.

"Flag it when you are unsure" is not a rule an engine can follow

The obvious fix is to have the engine mark its own shaky claims, and it is worth being precise about why that fails. A model's confidence is a poor proxy for whether a claim is true: it can be fluent and wrong with no internal difference from fluent and right. Self-reported uncertainty therefore comes back neither complete nor reliable, which is worse than no flags at all, because an unflagged paragraph starts to feel checked.

What an engine does do reliably is classify. Deciding whether a sentence is a claim about your own company, a claim about the world outside it, or one that would need someone to go and find something out is a categorisation task, and categorisation is a genuine strength.

So the rule cannot be about confidence. It has to be about the kind of claim, settled before the post exists.

Three kinds of claim, and only one of them is free

The first kind is a claim about you: what you sell, how you price it, how onboarding runs, what you refuse to take on. An engine may assert these without a citation, but only because the brief is the source. If it is not in the brief, it is not a fact about your company. It is a guess wearing your logo.

The second kind is a claim about the world outside your company — what a platform's policy says, what a standard requires, which way a market is moving. These need a named source and a link every time, with no exemption for claims that feel like common knowledge. Almost everything that quietly rots in an old post lives in this category, because the world moved and the sentence did not.

The third kind would require original reporting: an interview, a dataset somebody has to assemble, a survey nobody has run. An engine cannot produce these, and a policy that does not say so out loud is an invitation to approximate one. The correct output is a refusal routed to a human, which is one of the limits worth naming in advance.

Cite the document, not the write-up about the document

When the claim is about Google, cite Google. The spam policies and the helpful content guidance are short, public and unambiguous. The industry's summaries of them are none of those things. A summary compresses, and compression is where the qualifier goes missing — which is roughly how a narrow rule about scaled content abuse became, across a year of secondhand write-ups, the folk belief that Google penalises AI content.

There is a second reason, and it has nothing to do with accuracy. A primary source outlives its summaries. The article explaining a policy change goes stale the next time the policy changes; the policy page updates in place. A post that links the document is still broadly right in two years. One that links last year's hot take is an archaeology exhibit.

Google's own self-assessment questions ask whether content presents information in a way that makes you want to trust it, naming clear sourcing and evidence of the expertise involved as the things that do it. That is not a ranking trick. It is the same instinct, written down by the people who run the index.

What changes if you write the rule down

Accepting this makes the blog slower in one narrow place and considerably cheaper in another. Slower, because some subjects come back refused rather than written, and a refusal at six in the morning feels like a malfunction right up until you price the alternative. Cheaper, because approval stops being a hunt for the one invented sentence and becomes a check that the sourcing rule was followed — a question with an answer, closed in minutes instead of an afternoon of doubt.

The rule belongs in the brief rather than in someone's head, because an engine reads the brief and does not read the room. Write down which claims about your company are settled enough to assert unaided, which outside claims must trace to a primary document, and which subjects stay off the table until a human does the work.

A blog that can show where it got things is doing something most of the fluent competition is not. It is a small advantage. It compounds quietly, and it costs you one paragraph of policy.