Summary
Agency article production runs $800 to $2,000 per piece, and every engagement I’d seen rebuilt the same pipeline by hand. I built one factory where Claude drafts, code verifies and a human edits, running for $12 to $15 a month per client. Then I built a second system whose only job is to block the first when it’s wrong.
The two problems with AI content
The first problem is cost, and it’s less obvious than it sounds. AI made drafting cheap. It didn’t make the rest cheap: keyword research, briefs, link checking, images, publishing, measuring what ranked, and deciding what to write next.
At three articles a week, agency rates run $10,000 to $26,000 a month per client. Most of that isn’t the writing.
The second problem is trust. Most “AI content” is a prompt and a prayer. The model writes something plausible, nobody checks the links, a fabricated statistic ships under the company’s name, and six months later a prospect finds it.
The fix isn’t a better prompt. It’s a system where the model’s output is treated as a draft to be verified, not a deliverable to be published.
The factory
I built my Article Factory, a client-agnostic system that runs the whole line, not just the drafting.
Research. Keywords pulled and prioritized against what the client’s buyers search.
Drafting. A Claude agent runs a six-stage pipeline per article: brief, draft, structure for AI search, optimize, self-grade, revise. The prompts carry the client’s brand context, a 30-point scorecard across six categories (content, SEO, AI-search readiness, entity clarity, structured data, conversion), and a kill list of roughly 34 words and 12 phrases that make writing sound like a machine.
Verification. This is where trust lives, and it’s deterministic. After the agent finishes and grades itself, code re-checks every external link with a live request and overwrites the agent’s grade with the result.
A dead link fails the article regardless of what the model claimed. Images, calls to action and lead-magnet wiring are attached by code, not by the model.
Publishing and monitoring. The system detects when an article goes live, then watches it in Search Console, analytics, the CRM and an AI-visibility probe. A weekly audit reads what moved and feeds the next batch’s research. The line learns.
Multi-tenant. One queue, one database with strict per-client isolation, and each client gets its own live dashboard and an operator copilot. Provisioning a new client is one command and about fifteen minutes. Five are live.
The build: 42 shared modules, 30 scheduled and on-demand tasks, 28 database tables, 21 operator tools. Queue capacity is 500 to 1,000 articles a week per client.
Actual cadence is whatever the client approves; for the home-services company it was three a week. The full build, from context layer to publishing, is in How I Built My Article Factory.
The gate that blocks us
The factory grades its own work, and self-grading is the industry’s disease. Every marketing deliverable is scored by the person who made it. So I built a second system, a ship-or-block gate for any SEO or AI-search deliverable, that runs in a fresh session against the original sources rather than against the finished document.
Its rules are simple and non-negotiable. Findings grade from F1 to F4. One F1, a fabricated, false or unverifiable claim, is an automatic block. No exceptions.
It runs five passes: confirm the data connections point at this client before spending anything; a scripted check specific to the deliverable type; live re-verification, which means re-querying analytics, re-pulling rank data, re-crawling the site, and re-running a sample of AI answers with at least one positive and one negative per engine; an adversarial human read; then a verdict. Twelve check classes, ten claim classes with explicit tolerances.
Only the named account lead can override a block, in writing, logged in the report. In April 2026 a client-side catch on one batch became a hard rule inside the production prompt the same week.
The gate ratchets. It doesn’t loosen.
Claude writes, code verifies
That’s the principle both systems share. The model isn’t trusted; the checks around it are, and they’re deterministic so they can’t be talked out of a verdict.
What it produced
- Cost to run: $12 to $15 a month per client, against $800 to $2,000 per article at agency rates.
- For a 59-market home-services company: 158 articles refreshed and 14 new in one quarter, every one through the scorecard and a human edit. Over the same window the company’s organic leads went from zero to 134 a month, organic collected $32,538 in 90 days, and monthly leads doubled while ad spend fell 96%.
- For a pre-FDA medical device company: nine articles at 24 or 25 out of 25, six cleared clinical review on the first pass.
- A private equity firm’s Head of AI walked through the per-client isolation model looking for a hole. He didn’t find one.
The factory makes the content affordable. The gate makes it publishable. Neither is worth much without the other.
At a glance
- The niche: Any company publishing SEO and AI-search content; five clients live.
- The problem: Agency content at $800 to $2,000 per article, and AI content with no verification behind it.
- What was built: A multi-tenant article factory where Claude drafts, code verifies and a human edits, plus a separate ship-or-block quality gate for any deliverable.
- Implementation: 42 shared modules, 30 tasks, 28 tables, 21 operator tools; one-command client provisioning; a 30-point scorecard with a deterministic link audit that overrides the model; one unverifiable claim blocks a deliverable.
- Results: $12 to $15 a month per client; 172 pieces in a quarter for one client that collected $32,538 from organic in 90 days; six of nine medical articles cleared review on the first pass.
- Built with: my Article Factory and the ship-or-block QA gate.
