Why an AI SEO draft should refuse invented facts and still not publish itself

A founder sits at her desk at eleven at night. The landing page is due tomorrow morning. She has three things she actually knows about her product. She knows what it does. She knows who it serves. She knows one specific detail about how it was built. She pastes a prompt into an auto-publish SEO writer and watches a full article assemble itself. Somewhere in paragraph four, a customer named Maria describes how the product saved her team six hours a week. There is no Maria. There was never a team. The tool filled the gap between what the founder knew and what the word count demanded.

Nothing about that scene is exotic. It is the default behavior of an entire category of tools. The behavior follows from two design choices that compound each other. One choice is to fill every gap. The other is to treat publication as part of generation. A trustworthy AI SEO draft must refuse both. It should refuse to invent facts. It should refuse to publish itself. Those refusals are related. How they work together is the difference between a draft you can read and a liability you have to audit.

The default is fill every gap, then publish

The failure modes hiding inside auto-publish SEO writers are mechanical, not accidental. Fabrication has a precise technical cause. A language model asked to produce a complete article has no internal sense of what it actually knows versus what it is pattern matching toward. Give it a thin brief and demand a finished piece, and the path of least resistance is invention. The model needs a transition, so it supplies an anecdote. It needs social proof, so it supplies a customer. It needs a statistic to make a paragraph feel authoritative, so it supplies a number. None of this requires malice or a bug. It is what the objective rewards. The model optimizes for token probability, not truth. A complete-looking article scores higher than a hesitant, incomplete one.

Consider a different kind of fabrication. A founder provides a brief about a new encryption feature. The brief mentions the algorithm but omits the key length. The statement is plausible. It sounds correct. It is entirely made up. If the actual implementation uses a different protocol, the published page contains a factual error that could mislead technical buyers or violate compliance standards. The model did not know the fact. It guessed the most likely sounding fact. That guess becomes permanent on the web.

Auto-publish removes the only safety step

Auto-publish writers stack a second failure on top of this fabrication risk. When generation and publication are the same action, the fabricated anecdote never stops in a draft folder where someone might catch it. It goes live. The human review step, the only place a founder would notice that Maria does not exist, has been engineered out of the loop in the name of speed. The pitch is that you skip the boring part. The boring part was the safety mechanism.

This is the same operating-file failure described in why agent work needs receipts: “done” becomes a feeling instead of a checkable record. A page that went live without a human decision is not a draft. It is an unreviewed claim.

Volume pressure makes thin briefs normal

The two failures reinforce each other in a way that matters for anyone evaluating tools. A tool that publishes automatically creates pressure to generate more content faster and at volume. Volume pressure makes thin briefs normal. Nobody writes a rich brief for thirty articles a week. Thin briefs make fabrication normal. The model has less real material to work from, so it relies more heavily on pattern matching and guessing. The result is a site that looks productive while quietly filling itself with plausible fiction. And because the founder never read the drafts, she cannot tell you which sentences on her own site are true.

Page Draft uses only the facts you give it

Page Draft runs on a different mechanism. It is simple enough to state in one sentence. It uses only the facts and audience you give it. You supply what you actually know about your product and who the page is for. The draft gets assembled from that material. When your inputs run out, the draft gets thinner or the gap stays visible. It does not get filled with a story about a customer who never existed.

That is a claim ceiling applied to drafting. The trust architecture of AI agent work treats a claim ceiling as a boundary, not a budget. Page Draft applies the same idea to sentences: if you did not hand over the fact, the draft does not invent it.

The refusal is the feature

Most writing tools are judged by the opposite standard. They are judged by how much they produce from how little you give them. Page Draft inverts that metric. The less you give it, the less it claims. A draft built on three real facts is a draft where every sentence traces back to something you handed over. That traceability is what makes it safe to read. You review it quickly because you are checking emphasis and wording. You are not fact-checking your own website for inventions you never approved.

When you give a rough idea without specific details, the draft reflects that vagueness. It does not invent specifics to fill the void. Where the draft has a hole, the hole tells you something: the claim you have not earned yet, the proof still missing. An auto-publish writer hides that information by patching over it with fiction. Page Draft surfaces it.

A visible gap is cheap to fix before publication and expensive to discover after.

Grounding does not replace judgment

Some founders push back here. If the draft is fully grounded in facts you supplied, why does the tool also need to refuse publishing? Grounding solved the fabrication problem. Why not let the pipeline run all the way through?

The flaw in that reasoning is that grounding removes fabrication risk while leaving judgment risk untouched. A draft can be entirely accurate and still wrong for the page. A company pivoting its positioning can supply facts about old features; the draft accurately reflects them and still misses the current priority. Those are decisions. Decisions belong to the person whose name goes on the page.

That is why Page Draft still never publishes for you. A human decision checkpoint sits between the draft and the live page. The product shape supports this: one paid draft at a time. Medium and long stay one draft. A pipeline producing thirty articles a week produces a review queue nobody actually reads. An unread queue is functionally the same as auto-publish.

Generic agent workflow tools fail the same way when they optimize for autonomy without a founder gate. Agent memory without approval fails when an invented sentence is retained as company truth. Page Draft is the drafting instance of that refusal, not a clone of those essays. The same refusal on a live query-page pair is the 14-day search repair loop: patch from owned evidence, do not invent the advice.

What search usually offers instead

Search for guidance on this topic and you mostly find two things, neither of which helps you decide. The first is generic workflow advice: always review AI output. Correct and useless if the tool's incentives outpace review. The second is auto-publish SEO writers themselves. Their optimization target is volume. That target is where fabrication lives.

Page Draft is the deliberate opposite of auto-publish SEO writers. Distinct from generic workflow tools and from journal posts about agent architecture. The mechanism is a constraint built into generation. No invented stories. No publication without you.

What you can actually try

On Page Draft the visible mechanics are bounded. Free preview of outline plus a short sample. Writer is $19 a month for short drafts, three per day. Medium is $19 one-time. Long is $29 one-time. Social stays free. It starts from what you already know and never invents facts. It still never publishes for you. One paid draft at a time.

That is the claim ceiling. No guaranteed traffic. No invented case studies. No ranking or lift metrics.

A composite scene: a founder has three real facts. An auto-publish writer fills the rest and ships. Input-bounded drafting comes back thinner where knowledge runs out. The gap sits in a draft, before anything is live.