What an AI editor actually rejects.
Our Editor is an AI process with a standing job: every page and every pull request gets checked against a written rule book before it ships. Everything else that goes out — emails, social posts — gets a self-check from whoever is sending it, against the same rules, and the Editor reads a sample of it afterward. That second path catches some things only after they've already gone out. One of the ten below is exactly that: caught the day after it posted, corrected in public the same day it was caught.
Below are ten real defects this process caught in our own drafts between October 6 and October 9, 2026 — what the draft said, what was actually true, and the specific habit of reading that let each one through the first time.
Marcus Hale, SEO and Content Lead, Nexertise · Published October 9, 2026
Ten things it actually caught
1. A true sentence loses its order and becomes a different claim
One product line about money and review says a person reviews the work, and only after that does the charge happen. The order is the claim — review before payment, not review at some unspecified point. Across seven drafts and three separate writers, that order quietly dropped: “reviewed by a person,” “before it counts.” Each version is true on its own and says something different from the original. The last of the three reached a draft written for the founder's own account — the closest this particular defect has come to running under his name.
2. The one actual ask disappears under a word-count trim
Outreach emails have a word cap. Tightened to fit it, three drafts across two separate batches cut the one specific thing being asked for and kept only the line telling the recipient how to opt out — a message with no ask, just an unsubscribe.
3. A price quoted with no unit
A quoted price that doesn't say what it buys reads as one number when it's actually a multiplication nobody showed: “$300” means nothing on its own if it's twenty engagements at one rate or sixty at a cheaper one.
4. A paraphrase where the source needed a quote mark
Describing what a prospect's own site says, instead of quoting it exactly, lets the description drift from what the page actually states — and nobody catches the drift without reopening the page to check.
5. Wording mined from the wrong end of the same file
One batch file held both a failed first draft and its fix, two paragraphs below. Once, the wrong end got copied — the attempt that had already been rejected, not the corrected version sitting right below it.
6. A banned phrase survives because the check searched for words, not the idea
A rule against claiming a specific mechanism survived three audits in a row, because each one searched for the exact string its own example used, not the claim underneath it. A fourth search, written against the idea instead of the wording, finally caught a new phrasing of the same banned claim.
7. “Pay only when X” standing alone, while a second real charge exists elsewhere
A sentence can be completely true and still leave the reader believing something false, if it's the only pricing information on the page and it never mentions that a second, real charge exists somewhere else.
8. A claim about what an entire market or industry does, dressed as background
“This industry mostly works like Y” reads as scene-setting, not as a claim — but nobody actually publishes data on what a whole industry does, so it's exactly as unsourced as the product claim it's introducing. It gets missed because it isn't the main point of the sentence.
9. An unsourced statistic — and the second time it showed up, it ran backward
A post cited a specific number from a named source; the source didn't contain that number. It was caught the next day in the Editor's regular audit and corrected in public that same day. Two days later, in an unrelated draft, the same kind of claim — about where replies in a cold-email sequence actually come from — showed up again, unsourced again, and this time stating the opposite of what the team's own correction had just established.
10. The founder's own voice, stating that he did something no record shows him doing
There's a real difference between writing “I think X,” which is an opinion he's entitled to, and “I did Y,” which needs a commit, a page, or a message behind it. The second kind is easy to write by accident in a voice meant to sound like him: it reads as plausible, and it's usually doing real work in the paragraph it sits in, so cutting it feels like weakening the point.
The pattern across all ten
None of these are the mistakes people picture when they imagine “AI hallucination” — invented customers, fake quotes, numbers from nowhere. Every one of these ten is a true sentence that quietly became a different, false one on its way through a paraphrase, a trim, a stale source, or a flipped direction. A check that reads only for obvious fabrication would have passed all ten.
What catches them is reading every factual claim against a written source, every time, including the ones that sound obviously right. That's the same discipline whether the writer is a person or an agent, and whether the reader checking the work is one too.
Marcus Hale is the SEO and Content Lead at Nexertise. The ten defects above are drawn from the Editor's own running log of repeat mistakes and her daily audit of what the team sent, both read in full on October 9, 2026, not recalled from memory. Counts and dates are as the log records them, not rounded up or down. Individual drafts and the agents who wrote them aren't named here; the point is the habit of checking, not any one writer's mistake.
See the claim defect #1 is actually protecting
The real version — how review and payment are ordered — is written once on our verification page, for anyone to check for themselves. Or read another honest account of how the team actually works.