Our editor rejects AI slop — before it's published

The gist, for those who scroll: AI produces filler as fast as it produces text, so our pipeline has deterministic gates that refuse to publish slop to itself — before the deploy.

I once read a batch of generated drafts back to back, and by the third one I could predict the next sentence. That’s the profile Google’s March-2026 core update went after as “scaled content abuse,” and it’s what any pipeline produces if nobody stands at the door — so ours has deterministic gates that block publication of slop.

Why gates are mandatory at scale

If you generate hundreds of pages, the question isn’t “will the AI write it” but “won’t this drag the whole domain under a filter?” Thin, templated, author-less content in volume is exactly the profile that gets penalized, so the filtering has to be automatic, not “we’ll read it later” — automatic according to whom, though?

Four gates that stop the deploy

• Information Gain floor. Every article must add something new, not rehash the SERP. Threshold ≥ 0.30 (≥ 0.40 for YMYL); built on the idea of Google patent US 11,354,342 — information gain per block versus the top results (my subjective take: this one does the heavy lifting).
• Anti-AI-isms. A cliché dictionary + structural markers: boilerplate intros (“in today’s world…”), paragraphs starting only with connectives, lists out of place. Caught in both the writer and the editor.
• Grounding markers. The writer tags unproven claims [unverified] / [citation needed] / [tk]. If such a marker survives to publish — the deploy is blocked. You can’t “accidentally” ship an unverified fact — and accidentally usually does all the work there.
• Content-effort. At least one element of “work you can’t copy”: proprietary data/table, an original visual, first-hand specificity (a tested procedure, a named trade-off). Two for YMYL.

An atomic pre-publish

The final gate, /seo-matrix:pre-publish, checks the batch of ready articles deterministically: frontmatter completeness, JSON-LD validity, internal-link target existence, asset presence, markdown correctness, the IG threshold. Any blocker → exit 1, no partial publish. Either it’s all clean, or we don’t publish — one weak article holds up the whole batch, and yes, that’s the point!

What you get

• At scale the domain doesn’t fall under scaled-content-abuse — because thin and templated never reaches production.
• Content is denser and more original by default (which is exactly what raises the odds of being cited in AI answers).
• No “the AI wrote it, we posted it blind”: the gates don’t depend on a reviewer’s mood.

In short

• Generating at scale with no filtering = filter risk (scaled content abuse).
• 4 gates block slop: information-gain, anti-AI-isms, grounding markers, content-effort.
• pre-publish is atomic: one blocker — nothing ships.
• The loop is closed: writer → editor → pre-publish, all deterministic.

Does it catch everything? Probably not — no gate does, whatever the vendor tells you. But it catches the slop we’d otherwise have published ourselves — arguably the only slop we answer for. Small mercies!

Sources

• seo·matrix’s own quality gates (pre-publish, information-gain self-check, anti-ai-isms, content-quality-checks).
• Google patent US 11,354,342 (“Information Gain”) — information gain per block.
• Google API Content Warehouse leak (2024) — the contentEffort signal (“complexity of reproduction”).
• DrMax, “Evidence-based SEO 2026” — content anti-replicability; context for the scaled-content-abuse core update.