The AI Content Pipeline That Publishes 30 SEO Articles a Month Without Burning Out #

Automating content creation with AI means wiring a staged pipeline—brief → draft → human edit → schema/FAQ → publish—so models do the volume work and humans own judgment, claims, and brand. Thirty SEO articles a month is not a prompt problem. It is an operations problem: queue management, QA gates, model routing, and a cadence your team can sustain for a year.

I'm William Spurlock — AI Solutions Architect and Fractional AI CTO. I run a daily publishing engine on this site and build the same class of systems for founders and growth teams who need content output without hiring a ten-person editorial department. This pillar is the mechanism-level playbook: architecture, tools, stages, QA gates, and the math behind ~30 posts/month.

If you already run a sibling loop for email, see how to build an AI-powered newsletter that writes and sends itself. For tool host choice, see n8n vs Make vs Zapier in 2026. For the content strategy layer that feeds this pipeline, see the AI visibility content strategy and the question-first content model.

---

How do I automate content creation with AI? #

You automate content creation by treating each article as a ticket that moves through fixed stages with machine work and human gates—not by pasting a keyword into ChatGPT and hoping the CMS publish button feels lucky. The winning shape in 2026 looks like a factory: inputs (questions, keywords, cluster maps), transforms (briefs, drafts, schema), inspection (edit + claims), and shipping (CMS/git deploy).

The factory metaphor that actually maps to ops #

Factory concept	Content pipeline equivalent
Raw materials	Question bank, keyword list, SERP notes, internal links, brand voice rules
Work order	Airtable (or Notion) Post row with status, slug, date, cluster, primary query
Station 1	Brief generation (outline, H2s, FAQ list, entities, claims checklist)
Station 2	Draft generation (model write into markdown/CMS)
Station 3	Human edit + claims gate
Station 4	Schema / FAQ / meta packaging
Shipping dock	Publish + Airtable status flip + internal link audit
Quality lab	Rank/citation checks, refresh queue, failure alerts

If any station is "whoever feels like it this week," you do not have a pipeline. You have a vibe.

Reference architecture (mid-2026) #

Layer	Job	Typical tools
Source of truth	Schedule, status, questions, claims	Airtable (WS Blog style base)
Orchestration	Move tickets, call models, write files	n8n (primary), Make for lighter stacks
Authoring IDE	Long-form draft + repo edits	Cursor + skill/rules files
Draft models	Volume prose	Claude Sonnet 5, Gemini 3.5 Flash, GPT-5.4 mini
Edit / hard sections	Brand, claims, architecture	Claude Opus 4.8, GPT-5.5, Gemini 3.1 Pro
Research assist	SERP notes, source pulls	Firecrawl, search APIs, internal docs
Publish	Site artifact	Markdown under content/blog/, CMS API, or headless
Observe	Rankings, citations, failures	GSC, analytics webhooks → Airtable

The five stages you must not skip #

1. Brief — Primary query, H2 questions, FAQ list, entities, internal link targets, claims to source.
2. Draft — Full article from the brief, answer-first H2s, tables where comparisons exist.
3. Human edit — Voice, accuracy, banned AI-tells, no invented stats, link verification.
4. Schema / FAQ packaging — FAQ H3s that render FAQPage JSON-LD, meta, OG, entity clarity.
5. Publish — Ship the artifact, flip Airtable to Published, schedule the next ticket.

Skip stage 1 and drafts wander. Skip stage 3 and you ship confident nonsense. Skip stage 4 and you leave citation-ready structure on the table. Skip stage 5 discipline and your "pipeline" becomes a folder of drafts nobody ships.

What "automation" owns vs what humans own #

Owns	Machine	Human
Pull next Queued post from Airtable	Yes	No
Generate brief from TargetQuestions	Yes	Spot-check
First draft markdown	Yes	No (unless brand is ultra-sensitive)
Claims sourcing / hedge decisions	Assist	Final call
Voice pass + banned-word scrub	Assist	Final call
Internal link disk-verification	Script + human	Confirm
Cover image generation	Yes	Approve aesthetic
CMS/git publish	Yes (after approve)	Approve button
Strategy: which clusters to prioritize	Suggest	Decide

Opinion: If your "AI content automation" has no human edit gate, you are not building a publishing system. You are building a spam cannon with better grammar.

Minimum viable pipeline (2 weeks to first ship) #

Week 1:

1. Stand up Airtable tables: Posts, Questions, Schedule, Claims (or equivalent).
2. Define statuses: Queued → In Progress → Edit → Published → Refresh.
3. Write a brief prompt + draft prompt that force JSON/markdown schemas.
4. Wire n8n: Cron or webhook → fetch next Queued row → call model → write draft to Notion/Drive/git branch.

Week 2:

5. Add Slack/email approval with Approve / Revise / Kill.
6. Add FAQ + meta generation step.
7. Add publish step (CMS API or PR merge).
8. Add failure alerts + idempotency keys so retries do not double-publish.

That is enough to ship. Scale comes from queue depth and editor hours, not from a prettier prompt.

---

Can AI write my blog posts automatically? #

Yes—AI can draft complete posts on a schedule—but "automatically" only works when you define auto as "machine-generated draft + required human approve," not "model presses Publish alone." Fully unattended generative publishing is how brands get hallucinated product claims, thin SERP clones, and Google quality problems.

What automatic actually means in production #

Mode	What runs without a human	When I use it
Draft-auto	Brief + first draft	Default for most B2B / SEO programs
Assist-auto	Research pull + outline only	Regulated niches, legal, medical-adjacent
Ship-auto	Publish after approve token	Mature pipelines with trusted editors
Full-auto	Model publishes with no gate	Almost never for public SEO brand content

Full-auto is acceptable for a few narrow cases: changelog digests that only restate committed release notes, internal knowledge base stubs, or sandbox blogs you treat as experiments. It is not acceptable for client-facing SEO pillars with revenue claims.

The human minutes budget (this is the burnout math) #

Thirty posts/month sounds impossible until you stop writing from a blank page.

Role	Minutes per post (mature pipeline)	Notes
Brief review	5–10	Mostly accept/amend H2s
Draft skim + structural fix	15–25	Headings, tables, missing sections
Claims / accuracy pass	10–20	Kill unsourced numbers
Voice + banned-word pass	10–15	Faster with lint scripts
Link + schema check	5–10	Automate what you can
Cover approve	2–5	GenerateImage then glance
Total human	~50–85 min	vs 4–8 hours blank-page writing

At 60 minutes average × 30 posts = 30 human hours/month. That is one half-time editor equivalent—not a five-writer agency. Burnout happens when teams keep the blank-page process and bolt a chatbot onto it.

Prompt contract that keeps drafts shippable #

I force drafts into a fixed contract. Example prompt skeleton for an n8n AI node or Cursor skill:

You are drafting a blog post for {{brand}}.
Register: AI Solutions Architect — mechanism-level, specific, no fluff.
Primary query: {{primary_query}}
Target H2 questions (answer each; bold lead answer first):
{{h2_list}}
FAQ H3 questions (2–4 sentence answers; bold lead fact):
{{faq_list}}
Internal links allowed (ONLY these slugs; format (/blog/<slug>)):
{{verified_slugs}}
Banned words (never use in prose): delve, leverage, seamless, cutting-edge,
state-of-the-art, game-changer, robust solution, holistic, streamline, utilize,
empower, unleash, "to summarize", "in conclusion", standalone "dynamic".
Models you may mention: Claude Opus 4.8, Claude Sonnet 5, Gemini 3.1 Pro,
Gemini 3.5 Flash, GPT-5.5, GPT-5.4 mini, Llama 4. Never use retired 2024–2025 model names.
Claims rule: no naked statistics without a source URL or an explicit hedge.
Output: full markdown with camelCase frontmatter fields as specified.

That contract is what makes "AI wrote it" compatible with "humans can edit it in under an hour."

Failure modes I see every quarter #

1. SERP cloning — Draft mirrors the top 5 results with no original mechanism. Fix: require one opinion + one architecture table per H2.
2. Invented benchmarks — Model invents "47% CTR lift." Fix: Claims table + reject gate.
3. Orphan drafts — Beautiful markdown that never hits the CMS. Fix: Airtable status ownership + weekly publish SLO.
4. Link rot inside drafts — Links to posts that do not exist. Fix: disk-verify or allowlist slugs only.
5. Model name drift — Prompts still list retired mid-decade model names. Fix: pin the current model table in the skill/system prompt.
6. Editor as rewriter — Human rewrites 80% because brief was garbage. Fix: improve brief stage before blaming the draft model.

Can AI write your posts automatically? #

If you have:

• a question bank or keyword map,
• a brand voice doc,
• a human who will spend ~1 hour/post,
• and a publish path that is not tribal knowledge,

then yes. If you have none of those, AI will accelerate chaos.

---

What is the best AI tool for automated SEO content in 2026? #

There is no single best AI tool for automated SEO content in 2026—there is a best stack: Airtable (or equivalent) as source of truth, n8n as orchestrator, Cursor for long-form repo authoring, and a model roster of Claude Sonnet 5 / Opus 4.8, Gemini 3.5 Flash / 3.1 Pro, and GPT-5.4 mini / GPT-5.5. "Best tool" shopping is how teams buy another SaaS seat and still publish four posts a month.

Tool scorecard (use this, not vendor landing pages) #

Job	Winner on my builds	Runner-up	Avoid for volume SEO
Queue + editorial SoT	Airtable	Notion databases	Spreadsheets with no status API
Workflow host	n8n	Make (lighter volume)	Zapier-only for multi-step AI loops
Long-form in git/CMS repos	Cursor	Claude Code for long horizons	Chat UI copy-paste forever
Cheap/fast drafts	Claude Sonnet 5, Gemini 3.5 Flash, GPT-5.4 mini	Llama 4 (self-host / privacy)	Random "SEO writer" wrappers with stale models
Hard edit / architecture	Claude Opus 4.8, GPT-5.5, Gemini 3.1 Pro	—	Using Flash-class models for final brand voice
Research fetch	Firecrawl + APIs	Manual SERP notes	Uncited web browse hallucinations
Cover images	Image gen in pipeline	Designer batch	Stock photo graves
Rank / refresh signal	GSC + your own citation checks	Third-party rank tools	Vanity dashboard with no re-queue

Model routing I actually use #

Stage	Primary	Fallback	Why
Brief JSON	Gemini 3.5 Flash or GPT-5.4 mini	Claude Sonnet 5	Structured, cheap, fast
Draft (spokes)	Claude Sonnet 5	Gemini 3.5 Flash	Voice + instruction following
Draft (pillars)	Claude Opus 4.8 or GPT-5.5	Claude Sonnet 5 + human expand	Long coherence
Claims / skepticism pass	GPT-5.5 or Gemini 3.1 Pro	Claude Opus 4.8	Different model catches different confabulations
Meta + FAQ compression	Gemini 3.5 Flash	GPT-5.4 mini	Short-form is fine on Flash
On-prem / sensitive	Llama 4	—	When data cannot leave the VPC

Why "all-in-one SEO AI writers" usually lose #

They optimize for demo screenshots: keyword in → article out. Production SEO needs:

• cluster ownership (no cannibalization),
• parent pillar linkage,
• FAQ schema,
• internal link graphs,
• refresh loops,
• human approval audit trails,
• model upgrades without rewriting the product.

An all-in-one tool that locks you into last year's model names and a closed CMS is a ceiling, not a foundation. Build the pipeline around APIs you control.

n8n vs Make vs Zapier for this specific job #

Short version (full comparison lives in the n8n vs Make vs Zapier 2026 guide):

Need	Pick
20–40 posts/month, branching, retries, self-host option	n8n
Marketing ops owner, <10 posts/month, visual scenarios	Make
Simple "new Airtable row → Slack ping" glue	Zapier is fine as glue, not as the brain

For a 30-post cadence with AI loops, I default to n8n.

Cursor's role (people underuse this) #

Cursor is not "just a coding IDE" in this stack. For markdown-first sites (like this one), Cursor + skills/rules is the authoring surface:

• Skills encode voice, frontmatter contract, link rules, model currency.
• Agents draft into the correct path under content/blog/YYYY/MM/.
• Validation scripts (validate-blog-frontmatter, audit-blog-links) run before push.
• Airtable sync scripts close the loop (blog-sync.mjs push).

That is how you get mechanism-level consistency across hundreds of posts without a style guide PDF nobody opens.

---

Pipeline stages in depth: brief → draft → edit → schema → publish #

If you only remember one diagram from this pillar, remember the five-stage handoff with explicit artifacts at each gate. Below is the operating manual.

Stage 1 — Brief #

Artifact out: Brief JSON (or Notion/Airtable rich text) with primary query, H2s, FAQs, entities, link allowlist, claims checklist.

Brief fields I require:

{
  "slug": "example-post-slug",
  "primaryQuery": "How do I automate content creation with AI?",
  "h2Questions": ["...", "...", "..."],
  "faqQuestions": ["...", "..."],
  "entities": ["n8n", "Airtable", "Claude Sonnet 5"],
  "internalLinkAllowlist": [
    "how-to-build-an-ai-powered-newsletter-that-writes-and-sends-itself"
  ],
  "claimsToSource": [
    "Any traffic % or ranking claim"
  ],
  "serviceTrack": "ai-automation",
  "pillarPost": false,
  "targetLines": 400
}

Brief QA gate:

• Primary query is unique vs other Published posts
• Every H2 is a real question someone searches
• Internal links verified on disk before draft starts
• Claims list is non-empty if the draft will include numbers

Stage 2 — Draft #

Artifact out: Full markdown with frontmatter + body.

Draft rules that save editor hours:

1. Lead every H2 with a bold 1–2 sentence answer.
2. One structured element (table, list, mermaid, or prompt/config block) per major section.
3. No code tutorials unless they are n8n/MCP/schema/prompt configs.
4. Length targets: spokes 400–600 lines; pillars 600–1000+.
5. Model must refuse unsourced hard stats.

Draft QA gate (automated where possible):

• Frontmatter camelCase complete
• No banned AI-tell words (script)
• No stale model names (grep)
• Word count / line count within band
• Cover path matches slug

Stage 3 — Human edit #

Artifact out: Approved markdown + ClaimsValidated = true (when claims exist).

Editor checklist (print this):

1. Read the first paragraph out loud. Press-release tone? Rewrite.
2. Spot-check three factual claims.
3. Confirm every /blog/<slug> target exists.
4. Kill generic CTAs; match serviceTrack.
5. Ensure FAQ answers are 2–4 sentences with a bold lead fact.
6. Approve or send back with structured revision notes (not "make it better").

Revision notes format that models understand:

REVISE:
- H2 "Can AI write..." paragraph 2 invents a 40% figure — remove or source
- Add table comparing draft-auto vs ship-auto
- Replace closing CTA with AI automation strategy call → /contact

Stage 4 — Schema / FAQ / meta #

Artifact out: FAQ H3s present, seoTitle/seoDescription filled, entities listed, OG image path set.

Why this stage is separate: draft models under-invest in packaging. A dedicated compression pass (Flash-class model) is cheaper and cleaner.

Packaging checklist:

• seoTitle < 60 chars where possible
• seoDescription ~150–160 chars
• aioTargetQueries mirrors real H2/FAQ questions
• ≥2 FAQ H3s for FAQPage JSON-LD emission
• coverImage exists under public/images/blog/

Stage 5 — Publish #

Artifact out: Live URL + Airtable Published + Schedule date marked.

Publish sequence I use on markdown sites:

1. Validation scripts pass.
2. Cover committed.
3. blog-sync.mjs push --slug=... (or CMS publish API).
4. Spot-check live URL + FAQ schema in view-source / rich results test when stakes are high.
5. Move next Queued item to In Progress for tomorrow.

Idempotency rule: publish actions key on slug + content hash. Retries must not create duplicates.

---

Cadence design for ~30 SEO articles a month #

Thirty posts a month is ~1 post/day with weekends optional, or ~1.5 posts/weekday—either way, the constraint is editor hours and queue quality, not model tokens. Design the calendar before you buy more AI credits.

Throughput math #

Cadence	Posts/month	Editor hours @ 60 min	Notes
1/day	~30	~30h	Sustainable with one trained editor
5/week	~20–22	~20–22h	Good for smaller teams
2/day	~40–60	40–60h	Needs two editors or tighter briefs
Pillar-heavy weeks	fewer posts, more lines	same hours	Trade count for depth

Weekly operating rhythm #

Day	Machine work	Human work
Sunday	Generate next week's briefs from Queued questions	Approve / amend briefs (60–90 min)
Mon–Fri AM	Draft 1–2 posts overnight / morning batch	Edit + approve morning drafts (2–3h)
Mon–Fri PM	Package schema/meta + cover gen	Final publish click + spot checks (30–60 min)
Friday	Refresh candidates from GSC losers	Pick 2–3 URLs for refresh queue
Monthly	Cluster cannibalization report	Kill/merge overlapping PrimaryQueries

Queue depth rules #

• Keep 14–21 days of Queued briefs ready. Less = panic. More = stale SERP assumptions.
• Cap In Progress at 3–5 items so WIP does not hide blockers.
• Separate Pillar slots (1–2/month) from Spoke slots (the rest). Pillars eat editor hours differently.

Airtable fields that make cadence real #

Minimum Post fields:

• Status, Date, Slug, PrimaryQuery, Part, Category, ContentCluster
• PillarPost, ParentPillar, ServiceTrack
• TargetQuestions, FAQQuestions, ClaimsValidated
• FilePath, WordCount, CoverImage, LastModified

If Status is a free-text vibe field, your cadence will lie to you.

Burnout prevention (ops, not wellness posters) #

1. Hard WIP limits. No "just one more draft" beyond the daily cap.
2. Template the edit. Checklists beat heroic rereads.
3. Rotate claim-heavy posts. Do not schedule five data posts on the same day.
4. Refresh days count as content days. Updating old posts is part of the 30, not extracurricular. Pair with refreshing old content for the AI era.
5. Kill vanity metrics in the war room. Published count matters; "AI words generated" does not.

---

QA gates: the difference between a pipeline and a content firehose #

QA gates are binary checkpoints that can fail a ticket back to a previous stage—without them, automation only accelerates publishing mistakes. Here is the gate stack I install.

Gate 0 — Intake #

Fail if:

• PrimaryQuery duplicates an existing Published post
• Cluster/parent pillar missing on spokes
• No TargetQuestions linked

Gate 1 — Brief #

Fail if:

• Fewer than 3 H2 questions
• Internal link allowlist includes missing files
• ServiceTrack CTA mismatch vs intended offer

Gate 2 — Draft lint #

Fail if:

• Banned words present
• Stale model names present
• Frontmatter casing wrong / draft: true accidentally
• Body under minimum lines for post type

Gate 3 — Human approve #

Fail if:

• Editor selects Revise or Kill
• ClaimsValidated false while claims exist

Gate 4 — Pre-publish #

Fail if:

• audit-blog-links reports problems
• Cover image missing
• FAQ count < 2 when FAQPage expected

Gate 5 — Post-publish observe #

Not a blocker—creates tickets:

• Indexing errors
• Sharp ranking drops → refresh candidate
• Citation absence on target queries after N weeks → brief rewrite

Example n8n gate node logic (conceptual) #

{
  "name": "Draft Lint Gate",
  "type": "n8n-nodes-base.if",
  "parameters": {
    "conditions": {
      "boolean": [
        { "value1": "={{$json.bannedWordCount}}", "operation": "equal", "value2": 0 },
        { "value1": "={{$json.staleModelCount}}", "operation": "equal", "value2": 0 },
        { "value1": "={{$json.lineCount}}", "operation": "largerEqual", "value2": 250 }
      ]
    }
  }
}

Wire the false branch back to Draft with the lint report attached. Do not Slack the editor for machine-detectable failures.

---

Keyword research and briefs: automating the front of the funnel #

You can automate keyword research and brief creation by scoring questions against cluster maps and SERP features, then emitting a structured brief—humans still choose which clusters deserve budget. The automation removes spreadsheet slavery; it does not remove strategy.

Inputs that beat "dump Ahrefs into GPT" #

1. Question bank organized by Part/Category (AI Visibility / Automation / Agents).
2. Existing Post PrimaryQueries for cannibalization checks.
3. Search Console queries you already almost rank for (positions 4–20).
4. Sales/call transcripts — real buyer language beats tool keyword difficulty scores alone.
5. Competitor URL samples — for gap finding, not for cloning.

Automated brief scoring rubric #

Signal	Weight	Notes
Fits active cluster	High	Prefer depth over random topics
Cannibalization risk	High (negative)	Block near-duplicate PrimaryQuery
SERP is answerable with experience	Medium	Prefer questions you can answer with receipts
Business track alignment	High	Visibility vs Automation vs Agents CTA fit
Refresh vs new	Medium	Sometimes refresh wins over new URL

Brief generation prompt (compressed) #

Given unused questions + existing primary queries + cluster map,
propose the next spoke OR pillar brief.
Return JSON with: primaryQuery, h2Questions[3-5], faqQuestions[~8],
parentPillar, serviceTrack, whyNow (2 sentences), risks (cannibalization).
Do not invent search volumes. If volume unknown, omit.

What to keep human at the front #

• Annual/quarterly cluster bets
• Offers and CTAs tied to service tracks
• Anything that could create legal/compliance exposure
• "Should we publish this under our brand at all?"

---

Refresh loops: AI that updates old posts instead of only creating new ones #

A mature AI content pipeline spends a fixed share of capacity refreshing aging URLs—new posts alone create a graveyard. Automating refreshes means detecting decay, drafting a diff-oriented update, and re-running QA gates.

For the full framework, use refreshing old content for the AI era. The pipeline integration looks like this:

Step	Automation	Human
Detect decay (traffic, position, outdated models)	GSC export → n8n score	Confirm priority
Generate refresh brief (what changed since lastModified)	LLM	Approve scope
Draft surgical update (not full rewrite by default)	LLM	Edit
Update lastModified + claims	Script	Approve
Re-push / re-index request	Script	Spot-check

Refresh triggers worth automating #

• Model names older than current allowlist
• Stats older than 12–18 months without hedges
• Broken internal links
• Missing FAQ section on high-traffic URLs
• Primary query now dominated by AI Overview SERPs (needs answer-first rewrite)

Capacity split I recommend #

Bucket	Share of monthly posts
New spokes	60–70%
Pillars / major guides	5–10%
Refreshes	20–30%

If refreshes are "when we have time," they never happen.

---

Can AI content actually rank — and get cited? #

Yes, AI-assisted content can rank and get cited when it is answer-first, entity-clear, experience-backed, and operated under quality gates—Google and answer engines reward pages that resolve questions, not pages that brag about being written by hand or by AI. The failure mode is unedited commodity text at scale.

What still matters in mid-2026 #

Factor	Still matters?	Pipeline implication
Clear answer near the top of sections	Yes	Enforce bold lead answers in draft contract
Topical cluster depth	Yes	Airtable clusters + parent pillars
Technical crawl/index health	Yes	Separate from writing automation
Original experience / opinions	Yes	Human edit must add or preserve receipts
Thin spun duplicates	Negative	Similarity checks / editor kill power
FAQ + schema	Yes	Dedicated packaging stage
Entity consistency	Yes	entityMentions + same names sitewide
AI-generated disclaimer theater	No	Do not waste the lede on "this was written with AI"

Alignment with answer engines #

Your pipeline should produce pages that work for classic SEO and AI Overviews / answer engines. That means:

• question-shaped H2/H3s,
• concise definitions,
• tables for comparisons,
• citeable facts with sources when you state hard numbers,
• internal links to supporting spokes.

Pair this pillar with the question-first content model and the AI visibility content strategy so the automation layer is fed by a strategy layer—not the other way around.

Honest limits #

• AI will not invent domain authority.
• AI will not fix a site that cannot be crawled.
• AI will not replace subject-matter expertise in YMYL niches without heavy human control.
• AI will not make a confused offer coherent. Fix positioning first.

---

Reference stack: wiring Airtable + n8n + Cursor #

The production pattern is Airtable as editorial brain, n8n as nervous system, Cursor as hands on the repo—models are interchangeable workers behind API nodes. Here is a concrete wiring sketch.

Airtable → n8n #

Trigger options:

1. Cron every morning: fetch Status = Queued sorted by Date, limit 1–2.
2. Webhook when Status flips to In Progress.
3. Button in Airtable interface for manual runs.

n8n pulls:

• slug, title, primaryQuery, target questions, cluster, serviceTrack, pillar flag

n8n → model → artifact #

1. Build brief (or load human-approved brief).
2. Call draft model with contract prompt.
3. Write file to git branch / CMS draft.
4. Run lint commands.
5. Notify Slack with preview + Approve buttons.

Human → Cursor (when the site is markdown-first) #

For pillars and sensitive posts, I often keep draft generation inside Cursor with the authoring skill, then use n8n for queue, reminders, refresh detection, and Airtable push. Hybrid is fine. Purity is not the goal—ship rate with quality is.

n8n → Airtable push #

On publish:

• Status = Published
• FilePath, WordCount, ReadingTime, CoverImage, Excerpt, LastModified
• Flip linked Questions to Published
• Mark Schedule date complete

This site's blog-sync.mjs is that sync contract. Your CMS may use native APIs; the field names change, the loop does not.

Observability fields worth logging #

Event	Log where
Model + token usage per stage	Airtable or warehouse
Lint fail reasons	n8n execution data
Editor approve latency	Timestamp delta
Publish success/fail	Airtable + alert
Post-publish indexing	Weekly rollup

---

Team roles for a 30-post machine #

You do not need a newsroom. You need clear ownership across four seats—even if two seats are the same person on different days.

Seat	Owns	Does not own
Strategist	Clusters, PrimaryQueries, offer/CTA mapping	Line edits
Pipeline engineer	n8n, Airtable fields, lint scripts, model routing	Brand voice taste
Editor	Approve/revise/kill, claims, voice	Building workflows
SME (as needed)	Fact review on hard topics	Daily queue ops

Solo operators (me, often): wear all four hats but timebox them. Strategy on Fridays. Editing in morning blocks. Engineering when the pipeline breaks—not during every draft.

RACI for a single article #

Activity	Strategist	Engineer	Editor	SME
Pick question cluster	A	C	C	C
Brief approve	C	I	A	C
Draft generate	I	A (system)	I	I
Edit approve	I	I	A	C
Publish	I	A/C	A	I
Refresh decide	A	C	C	I

A = accountable, C = consulted, I = informed.

---

Cost shape (tokens, tools, humans) #

Token cost is usually the smallest line item once you hit 30 posts/month; editor time and tool sprawl dominate. Still, route models on purpose.

Rough unit economics (order-of-magnitude, mid-2026) #

Item	Per post (spoke)	Notes
Brief tokens	Low	Flash / mini class
Draft tokens	Medium	Sonnet / Flash class
Edit pass tokens	Low–medium	Only on revise
Image gen	Low–medium	One cover
Human edit	Highest	50–85 minutes
Tool seats	Amortized	n8n + Airtable + Cursor

Pillars cost more tokens and more editor minutes. Budget them explicitly.

Cost control tactics #

1. Do not use Opus/GPT-5.5 for every spoke draft.
2. Cache brand voice + banned lists in the system prompt; do not resend giant histories.
3. Reject at lint before human time.
4. Prefer refresh over new URL when the SERP is already yours to lose.
5. Cap parallel model calls to avoid retry storms.

---

Implementation blueprint: first 30 days #

Ship a thin pipeline in week 1–2, hit a steady 1.0 post/day by week 4, and only then add fancy agents. Most teams invert this and never publish.

Days 1–7 — Foundations #

• Airtable schema live
• Voice doc + banned list in one canonical file
• One draft prompt + one brief prompt
• Manual publish path works

Days 8–14 — First automation #

• n8n pulls Queued → drafts → Slack approve
• Lint scripts in CI or local pre-push
• Three real posts shipped through the loop

Days 15–21 — Packaging + covers #

• FAQ/meta pass automated
• Cover generation hooked
• Airtable push on publish

Days 22–30 — Cadence lock #

• Daily SLO: 1 approved publish on weekdays
• Friday refresh slot
• Retro: where did editor minutes go? Fix that stage

Exit criteria for "we have a pipeline" #

• 10+ posts shipped through the same statuses
• Mean editor time < 90 minutes/post
• Zero publishes with broken internal links in the last 10
• Model roster documented and current
• Someone besides the founder can approve a draft using the checklist

---

Anti-patterns (steal these for your kill list) #

Anti-pattern	Why it fails	Replace with
One mega-prompt "write a 2000 word SEO blog"	No stages, no gates	Five-stage pipeline
Ten SEO tools + no SoT	Status lies	Airtable (or one DB) as brain
Publishing straight from chat UI	No audit trail	Artifact in CMS/git + status
Measuring words generated	Vanity	Published + indexed + useful
No refresh budget	Decay	20–30% capacity
Identical CTA on every track	Confused conversion	serviceTrack-matched CTA
Ignoring cannibalization	Rankings fight themselves	PrimaryQuery uniqueness gate
Stale model names in prompts	Silent quality drop	Quarterly model table update

---

FAQ #

How do I build an AI content pipeline for my website? #

Start with a source-of-truth queue (Airtable), a five-stage flow (brief → draft → human edit → schema/FAQ → publish), and an orchestrator (n8n) that moves one ticket at a time through lint and approval gates. Do not begin with agent swarms. Ship three posts through a boring loop, then automate the handoffs you trust.

Can I automate keyword research and content brief creation with AI? #

Yes—score unused questions against your cluster map, cannibalization list, and GSC opportunities, then emit a structured brief JSON for human approve. Keep strategy (which clusters get budget) human. Automate the spreadsheet grind, not the bet.

How do I use AI to automatically update old blog posts for SEO? #

Detect decay with GSC/position/model-staleness signals, generate a refresh brief, draft a surgical update, re-run QA gates, and bump lastModified. Treat refreshes as first-class calendar slots. Details: refreshing old content for the AI era.

Can AI generate content that actually ranks on Google? #

AI-assisted pages rank when they are answer-first, clustered, technically healthy, and human-edited for experience and accuracy—commodity unedited AI text at scale usually does not. Automation raises throughput; it does not replace topical authority or crawl health.

How many human hours does a 30-post month really take? #

Plan ~25–40 editor hours/month once briefs and linting are mature—roughly 50–85 minutes per post—not the 120+ hours blank-page writing would consume. If you are over 2 hours/post after month two, your brief stage or voice contract is broken.

Should I use Claude, Gemini, or GPT for SEO drafts? #

Use Claude Sonnet 5 or Gemini 3.5 Flash for most drafts, escalate to Claude Opus 4.8 / GPT-5.5 / Gemini 3.1 Pro for pillars and hard edits, and keep GPT-5.4 mini for cheap structured steps. Multi-model routing beats loyalty to one logo.

Where does Cursor fit if n8n already calls the model? #

Cursor shines for markdown-first repos, skill-enforced voice/frontmatter, and multi-file fixes; n8n shines for schedule, integrations, and approvals. Many production stacks use both: Cursor for authoring quality, n8n for ops glue.

What QA checks should block publish automatically? #

Block on banned AI-tell words, stale model names, broken internal links, missing cover, failed frontmatter contract, and missing human approval token. Soft-fail (ticket only) on post-publish ranking decay and citation gaps.

How do I keep AI drafts from cannibalizing each other? #

Enforce a unique PrimaryQuery per Published post, assign ParentPillar on spokes, and reject briefs that overlap existing queries above a similarity threshold. Clusters exist to prevent your own library from competing with itself.

Do I need a separate pipeline for newsletters vs blog posts? #

Share the research → draft → approve → deliver shape, but keep separate workflows, prompts, and ESPs/CMS targets—newsletter cadence and blog SEO constraints differ. The AI newsletter pipeline guide is the email sibling of this pillar.

---

Build the factory, not another prompt #

If you want ~30 SEO articles a month without burning out, stop hunting for a magic writer tool. Install a factory: Airtable for the queue, n8n for the handoffs, Cursor for repo-quality drafts, current models for volume and edits, and humans for judgment.

I design and ship these systems as part of AI Automation + Growth work—editorial Airtable bases, n8n content workflows, lint gates, and model routing your team will actually run. If you want this mapped onto your CMS, brand voice, and monthly capacity, book an AI automation strategy call and bring: your current publishing tool, how many posts you ship today, and one URL you are proud of. We will sketch the stage map and WIP limits before anyone writes a prompt.