Inside the Quality Gates: How AutoArticle AI Prevents Bad Articles

Why Quality Gates Matter

When you're publishing 100 articles a day across 8 agents, even a 1% error rate means one broken article every day. Over a month, that's 30 articles with truncated paragraphs, duplicated content, or SEO-killing focus keywords.

We rebuilt our pipeline with 12 independent quality gates — each one catching a specific class of failure. Here's what they are and how they work.

The 12 Gates

Gate 1: Topic Fragment Filter

Rejects headlines that are too short or contain only stop-words. "for Season" — a 2-word fragment — can never become an article.

Gate 2: Review Rewrite Blocker

Rejects headlines in the review format ("X Review: verdict", "Hands-On: …"). These are someone else's copyrighted original work — rewriting them is plagiarism. Structural review format only; news stories mentioning "review" still pass.

Gate 3: Focus Keyword Validation

Seven sub-checks on the focus keyword:

• All-stopword rejection ("Already Has" is never a valid keyword)
• Verb-leading rejection ("change fifa" → re-derived from headline entities)
• Non-ASCII transliteration (Tōkon → Tokon — the ō character breaks downstream tools)
• Maximum word count (9-word keywords are sentence fragments, not keywords)
• Trailing stopword rejection (keywords ending with "the" or "in")
• Cannibalization auto-differentiation (extends a short keyword if another article recently targeted it)
• Quote-aware differentiator extraction (strips quoted speech before picking differentiating entities)

Gate 4: Tag Quality Sweep

Four filters on the final tag list:

• Verb fragment rejection ("Demands", "Hiked", "Slams" — these are actions, not tags)
• Spec fragment rejection ("000mah", "Inch HD" — hardware table columns, not tags)
• Trailing stopword rejection ("Revolt The" — ending with "The" is never right)
• Entity-variant dedup ("Man United" / "Manchester United" / "United" — keep the longest, drop the rest)

Gate 5: Truncation Detection

Scans every paragraph (and every BR-separated segment) for mid-sentence cuts. Articles with truncated paragraphs are routed to draft, never published. Catches:

• Paragraphs ending without terminal punctuation
• Mid-word cuts ("pledged to keep coope")
• Short question headings cut mid-phrase ("Will the " at 8 chars)

Gate 6: Paragraph Deduplication

The writer sometimes outputs the same content block twice — once as an outline leak, once in the proper section. The dedup pass normalizes every paragraph (≥80 chars), groups by key, and keeps only the occurrence inside a proper H2 section.

Gate 7: Save Gate (The Last Line)

Four checks immediately before the article is saved to WordPress:

• Title garbage — rejects "deepseek-v4-flash Generated Content", "HEADLINE: …", "DOMAIN: …"
• Title length — rejects titles under 15 chars or ending with a preposition
• Near-duplicate — checks if an identical title was created in the last 60 minutes
• Stub content — rejects articles with fewer than 100 real words

Gate 8: SEO Label Scrubber

Strips leaked SEO metadata from body text. Catches five variants:

• Standard labels ("FOCUS_KEYWORD: …", "META_DESCRIPTION: …")
• Bold labels ("SEO_TITLE: …")
• Agent-name variants ("Rahul Dravid SEO …")
• END markers ("END_RANKMATH_SEO")
• Orphaned values (the label was stripped but the value survived)

Gate 9: Evidence Sanitizer

Removes internal metadata from the writer's fact-verification prompt. ISO timestamps ("2026-08-24T19:34:45") and "published on X's Y site" phrases are stripped before the writer sees them, preventing leaks into body prose.

Gate 10: Heading Validator

• Removes stray single-word headings ("Now", "More", "Update")
• Deduplicates identical H2s (the same heading appearing twice)
• Promotes H3s to H2s when the body has zero H2s

Gate 11: 3-Tier External Link Strategy

Non-exclusive articles drop all Tier 3 links — no free link equity to news rivals.

Gate 12: Internal Link Booster

Extracts capitalized entity names from the article's topic (brands, products, people), searches your last 365 days of posts for matches, and injects them as high-priority internal links — creating the hub-and-spoke entity structure Google rewards.

The Result

Since deploying these gates, we've seen:

• Zero published truncated articles (previously ~2 per day)
• Zero duplicate publishes (previously ~1 per week)
• Zero metadata leaks (previously ~3 per day)
• Tag quality complaints down 90% (from 8+ garbage tags to 1-2 clean entity tags)

Every gate logs its decision, so you can see exactly what was caught and why in the agent logs.

Want to See It In Action?

Watch the plugin logs after any article generates — each gate prints its verdict. Or visit the Agents page and check the quality trend metrics on each agent card.