14 days into our AI content pipeline: → 48 articles published (6/day) → avg SEO score: 87 → Knowledge Graph: 2,100 entities, 5,400 relationships → AI cost: $5.80 total Still early for ranking data. Internal linking density is up 4x vs manual. Building in public — metrics

SEO check we run before every publish: Count your target keyword in H2/H3 headings. 0 times → heading hierarchy problem 3+ times (no variation) → stuffing Sweet spot: 1-2 times, semantic variants in the rest. This alone moved our avg SEO score from 74 → 86.

Quick SEO win hiding in Google Search Console: Sort by impressions descending. Filter for position 6-15. These articles already rank. They need optimization, not creation. 10 min of analysis beats 5 hours writing a new article on a topic you do not rank for.

3/ Quality variance: SEO scores all over the place. Fix: 24-module analysis as a hard gate. Score under 70 means it does not publish. What is your quality gate?

2/ Keyword cannibalization: 3+ articles targeting the same intent. Fix: entity-level topic clustering BEFORE keyword assignment. Map every article to a unique entity cluster. Zero overlap = zero cannibalization.

1/ Context drift: model forgets earlier article claims. Fix: query a Knowledge Graph of 2000+ entities before every generation. Model starts from what we already know, not what the internet says.

3 things that break AI content pipelines (and how we fixed them):

The metric that changed how we think about content ROI: Not cost-per-article. Cost-per-ranking-article. Pipeline at 4 articles/month @ 20% ranking rate = 0.8 ranking articles/month Pipeline at 24 articles/month @ 30% ranking rate = 7.2 ranking articles/month Volume isn't the g

After 2 weeks of running our AI content pipeline, the most surprising finding: The Knowledge Graph pays for itself in research time. Before: 45 min of manual research per article to avoid repeating ourselves. After: 5 min of graph queries. Entity relationships pre-loaded. Zero

Content cannibalization is silent. It doesn't announce itself. Signs your articles are competing with each other: → 3+ pages ranking for the same query with none above position 5 → GSC shows impression spikes that never convert to clicks → Your best article ranks behind an

Week 2 reflection on building an AI content pipeline: The assumption that was wrong: more articles = more rankings What actually worked: entity cluster coverage 4 articles covering the same Knowledge Graph cluster moved rankings. 20 scattered articles did not. Depth beats

The fastest SEO win hiding in your analytics: Sort articles by impressions, filter for position 5-15. These are already ranking — just not ranking enough. Add 300 words targeting the gap keywords, improve the H2s. They're 80% there. Don't start new articles when these exist.

Content velocity is a moat, not a vanity metric. Here's why: Month 1: 24 articles → limited data on what ranks Month 2: 48 articles → patterns emerge in which clusters gain traction Month 3: 72 articles → you know exactly which topics to double down on The compounding isn't

The part of our content pipeline nobody asks about: deduplication. Before generating any article, we check: → SQLite: exact title + keyword matches → Supabase: published URLs within 90 days → Knowledge Graph: entity overlap > 70% Without it, we'd publish 6 versions of "What

The quick SEO win most content teams miss: Before publishing, Google each of your H2s as standalone queries. If your H2 only makes sense inside the article context — it won't rank for anything. Each H2 should answer a specific question someone actually types into search. 5-mi

The content ROI math changes at month 4. Months 1-3: articles rank for nothing. Looks like a waste. Month 4: domain authority crosses a threshold. Old articles start ranking. Month 6: compounding — each new article ranks faster. Most teams quit in month 2.

Search intent classification matters more than keyword density. Informational: define/explain Commercial: best/top/review Transactional: buy/pricing/trial Mismatched intent = good SEO score, zero rankings. We classify before writing, not after.

Our Knowledge Graph revealed something unexpected: entities mentioned 3+ times across articles predict your next best topic. Not keyword research. Not trends. Entity frequency from existing content.

BYOK changes the unit economics of AI content completely. Typical agency AI content bill: → Jasper/Copy.ai: $49-499/mo (markup on GPT-4) → Clearscope: $170/mo (markup on GPT-3.5 for scoring) → Total: $200-700/mo What you're actually buying: convenience + 60-80% margin for

The entity gap is usually the ranking gap. When you're stuck on page 2, the problem is almost never keyword count. It's entity coverage. Run a quick test: take your top-10 SERP competitors and count the unique named entities in their first 800 words. Now count yours. That