Case Study · Content
Case Study: SEO Agency Ships 200 Programmatic Pages in One Sprint
How a 4-person content agency used a batch AI pipeline with human review gates to deliver a client's programmatic SEO project — on deadline, without slop.
Results
Pages produced
200
Human review pass rate (first draft)
62%
Production calendar
12 days
Pages indexed (6 weeks)
180
Client cost vs manual quote
-58%
Background
ClearRank Agency had a legal-tech client needing 200 "State A vs State B compliance" comparison pages for long-tail search. Manual writing: 10 weeks, $48k. The client had 6 weeks and a $25k budget.
The agency had shipped smaller batch projects but never at this scale without quality collapse.
The problem
| Constraint | Risk |
|---|---|
| 200 unique pages | Template repetition |
| Legal accuracy | Hallucinated statutes |
| Client brand voice | Generic AI tone |
| 6-week deadline | Review bottleneck |
They adapted the AI Content Batch Pipeline workflow.
Architecture
Topics CSV (state pair, verified facts, sources)
→ Per topic (parallel, cap 8):
Outline call (Sonnet) — H2s, thesis, fact placement
Draft call (Sonnet) — voice guide + outline + facts only
Fact-check call (Haiku + search) — flag unsupported claims
→ Queue in Postgres (status: draft | needs_edit | approved)
→ Editor UI: diff vs outline, approve/reject
→ Astro static publish (15 pages/day max)
Critical rule: the LLM could only use facts present in the CSV verified_fact column. Inventing new legal claims was a pipeline failure, not a prompt tweak.
Implementation
Week 1 — CSV prep (human work)
Researchers built the 200-row CSV:
state_a,state_b,primary_keywordverified_facts(3–5 bullet points with statute links)sources(URLs)
This took 5 researcher-days. It was the highest-ROI work in the project.
Week 2 — Pipeline + review tooling
Trigger.dev orchestrated batch jobs. Editors got a simple review UI:
- Outline on the left, draft on the right
- Fact-check flags highlighted in yellow
- One-click approve → publish queue
First-draft pass rate: 38% (single-shot) → 62% (outline-then-draft).
Weeks 3–4 — Review + publish
- 2 senior editors, 1 fact-checker
- 15 pages published/day (throttled after day 3 crawl warnings)
- Rejected drafts recycled with editor notes fed into prompt changelog
Results
| Metric | Manual estimate | Batch pipeline |
|---|---|---|
| Calendar | 10 weeks | 12 days production + 10 days review |
| Agency cost to deliver | $48k quote | $20k (profitable at $25k client fee) |
| First-draft usable | — | 62% |
| Pages indexed (6 wk) | — | 180/200 |
| Organic traffic (90 d) | — | +34% vs baseline |
Quality gates that mattered
- Verified facts column — constrained generation to seeded truth.
- Outline-then-draft — structure before prose.
- Search-backed fact-check — caught stale statistics.
- Human review on 100% — no exceptions.
- Publish throttle — 15/day beat 20/day for indexing.
What they'd do differently
- Build the review UI before the pipeline — editors worked in spreadsheets for 3 days and hated it.
- Throttle publish from day one — the crawl budget scare cost a weekend of panic.
- Charge more for CSV prep — clients underestimate research cost; it's half the value.
Takeaway
Programmatic SEO at scale isn't a writing problem — it's a fact supply chain problem. AI generates fast; humans verify facts; throttled publishing keeps Google happy. The batch pipeline workflow is the engine; the CSV is the steering wheel.
Lessons learned
- 01
Outline-then-draft beat single-shot generation — first-draft pass rate jumped from 38% to 62% when they split into two LLM calls.
- 02
A 200-row topic CSV with mandatory 'verified_fact' column forced researchers to front-load truth — the LLM couldn't invent what wasn't seeded.
- 03
Fact-check pass with search retrieval caught 94% of stale stats; the remaining 6% were caught by human reviewers who knew the domain.
- 04
Publishing 20 pages/day triggered a Google crawl budget warning — they throttled to 15/day and indexing improved.
- 05
Client approval sped up when reviewers saw side-by-side diffs vs the outline, not raw AI output.
Frequently asked questions
Related on AIKnowHub
Workflow
Build an AI Content Batch Pipeline
Generate, edit, and publish 50+ pieces of content in one run — programmatic content production with quality controls and human-in-the-loop review.
Concept
Evals Explained
Evals are the unit tests of AI systems. Without them you're flying blind. Here's how to build a useful eval set without going overboard.
Concept
RAG Explained
Retrieval-Augmented Generation is how you give LLMs access to your own data without retraining them. Here's how it actually works.
Comparison
Best AI Model for Writing
Which model writes prose that doesn't read like AI? A look at voice, restraint, and which models you can actually trust with words.
Prompt
Blog Post Outline
Generate a tight outline before you draft.
Prompt
Conference Talk Outline
Outline a talk that doesn't put the audience to sleep.