How to Automate Blog Posts With AI: An n8n to WordPress Pipeline That Drafts, Then Waits for You
One builder on this site runs keyword research to WordPress for under a dollar a week. Here is how to build the same pipeline, why it should stop at a draft, and what it really costs.
Two blog posts a week for less than a dollar
Most small business blogs die the same way. The owner writes three posts in January, gets busy in March, and the last post on the site is from two years ago. An agency would charge a few hundred dollars a post to keep it going, which is hard to justify when you cannot say what a post earns.
One builder documented on this site, in Full SEO Automation System: Keyword Research to Published WordPress Post for Under $1/Week, runs an n8n workflow that finds a keyword, researches it, writes the post, makes images and publishes to WordPress, twice a week, for under a dollar in total API costs. That is real and you can copy it. This guide shows you how, with one change that matters: the workflow stops at a WordPress draft, and a person who knows the business finishes the post. Without that step you are building exactly what Google's spam policies describe.
How the documented pipeline works
The listing describes four AI agents chained in n8n. A keyword step reads from a Google Sheet of target terms. A research step uses Perplexity to gather current facts and sources on the topic. A writing step uses Claude to turn that research into a structured post. A final step creates images and publishes to WordPress. Google Sheets acts as the queue and the log, so you can see what has been written and what is next.
A second case on this site shows what this kind of pipeline can do for a real business. In An AI SEO Content Pipeline Took a German Local Business From Zero to 1,400 Monthly Visits, an agency used Arvow and Claude Code to take a local client from no organic traffic to about 1,400 visits a month, which it values at roughly $1,200 a month in equivalent ad spend. Those are the agency's own figures and have not been independently verified, but they show the realistic scale: a local business does not need 100,000 visitors, it needs a few hundred people a month who are looking for exactly what it sells.
Why it should stop at a draft
Google's spam policies name scaled content abuse: generating many pages "for the primary purpose of manipulating search rankings", and the first example they give is using generative AI to produce many pages without adding value. Google's guidance on AI content says it rewards helpful content however it is produced, so AI drafting is not the problem. Publishing whatever the model writes, unread, at volume, is.
There is a plainer reason too. A drafted post about water heater replacement will say the same things as the ten posts already ranking for it, because it was built from them. What makes yours worth ranking is what only you know: what you actually charge, the mistake you see homeowners make, the permit your county requires. The machine cannot supply that. Thirty minutes of your time can.
A worked example: a plumbing company with a dead blog
Picture a four-truck plumbing company with a WordPress site and a blog last updated 18 months ago. The owner lists 24 keywords in a Google Sheet, things like tankless water heater cost in their city, signs of a slab leak, and whether a water softener is worth it in hard water areas, and marks the first eight as ready.
Every Monday and Thursday at 6am, n8n takes the next ready row. Perplexity returns a research summary with sources: current price ranges, what manufacturers say about lifespan, the questions people ask. Claude drafts a 1,200 word post using that research plus a facts document the owner wrote once: service area, licensing, real price ranges, warranty terms and three stories from jobs. n8n saves it to WordPress as a draft and marks the row drafted.
On Monday evening the owner or office manager spends 30 to 45 minutes on it. They fix the price range to match what the company quotes, add a photo from a real job, cut a paragraph that sounds like every other site, and publish. Twelve weeks later they check Search Console to see which posts get impressions. This is an illustration of the method, not a documented client, but every step uses tools and settings that exist today.
Step by step: build it this week
Step 1: Build the keyword sheet
Create a Google Sheet with columns for keyword, the question the reader is asking, status (ready, drafted, published, skip), the WordPress draft link and notes. Start with 20 to 30 keywords. Pull them from the questions customers ask on the phone, from Google's People Also Ask boxes, and from Search Console queries where you already get impressions but rank past position 10. Only mark a row ready if you could add something real to a post on it.
Step 2: Write the facts document
Write one page the model will read on every run: what you sell, where, real price ranges, what makes you different, words you never use, and three or four short stories from real jobs. Paste it into the prompt in Step 5. This single page does more for quality than any prompt trick.
Step 3: Connect WordPress to n8n
In WordPress go to Users, open your profile and create an Application Password, which WordPress has included since version 5.6 for authenticating REST API requests. Create a separate WordPress user with the Author role for this, so the workflow cannot change settings or plugins. In n8n add a WordPress credential with your site URL, that username and the application password. Your site must run on https.
Step 4: Add the trigger and the sheet lookup
In n8n create a workflow with a Schedule Trigger set to the days and time you want. Add a Google Sheets node that reads rows where status is ready and a Limit node that keeps only the first. If no row comes back, end the workflow there so it does not draft an empty post.
Step 5: Add research and drafting
Add an HTTP Request node that calls Perplexity's API with the keyword and this prompt: list the current facts, price ranges and common questions about this topic, with a source URL for each, and say where sources disagree. Perplexity moved its Sonar chat completions to its Agent API in late 2026, so follow the current API docs rather than older tutorials. Then add a Basic LLM Chain node with an Anthropic Chat Model running Claude. In the prompt include the facts document, the research and these rules: write for a homeowner, answer the question in the first two sentences, use only facts from the research or the facts document, mark anything you are unsure of with the word CHECK, and end with a short FAQ.
Step 6: Save as a draft and log it
Add a WordPress node with Resource set to Post, Operation set to Create, the title and content from Claude, and Status set to Draft, which is the node's default. The WordPress node cannot upload a featured image, so either add images by hand during editing or add an HTTP Request node that uploads to the WordPress media endpoint and sets the featured image on the post. Finish with a Google Sheets node that sets the row to drafted and stores the draft link, and an email or Slack message to whoever edits.
Step 7: Edit and publish
Search the draft for CHECK and resolve every one. Replace any number with your own. Add one photo and one paragraph from real experience. Delete anything you would not say to a customer on the phone. Then publish. If the draft is not worth fixing, set the row to skip and move on; that happens, and it is cheaper than publishing it.
Step 8: Check results after eight to twelve weeks
In Search Console, filter Pages by your blog path and look at impressions and clicks per post. Posts with impressions but few clicks need a better title. Posts with nothing after three months may be on keywords nobody searches. Adjust the sheet, not the workflow.
What this costs
As of November 2026, n8n Cloud Starter is 20 euros a month billed annually for 2,500 workflow runs, which is far more than two posts a week needs, and self-hosting n8n is free apart from your server. Claude Sonnet 5.5 through the API is $2 per million input tokens and $10 per million output tokens, so a draft with a few thousand words of input and 1,500 words of output costs a few cents. Perplexity's sonar model on its Agent API is $0.25 per million input tokens and $2.50 per million output tokens. Those are estimates from published token prices; your bill depends on prompt length.
If you would rather buy than build, Arvow's Solo plan is $69 a month, or $39 a month billed annually, for 1,000 credits and one website, with an article costing 15 credits, according to its pricing page. That is less setup and less control.
The real cost is editing time: 30 to 45 minutes a post, or roughly an hour a week at two posts. Budget it, or the pipeline becomes a spam machine with your name on it.
When not to do this
Skip it if nobody will edit the drafts. Skip it if your business gets almost all its work from referrals and repeat customers, because a blog will not move that. Skip it if your site has technical problems, such as pages that are not indexed or a broken mobile layout, because those need fixing first. And do not run it on a client's site without telling them how the posts are made and who checks them.
Frequently Asked Questions
Can I fully automate blog posts with AI and publish without reading them?▾
You can, but it is a bad idea. Google's spam policies treat mass-produced pages made mainly to rank as scaled content abuse, and unedited drafts tend to repeat what already ranks. Save as drafts and have someone who knows the business finish each one.
Does Google penalize AI-generated content?▾
Not for being AI-generated. Google says it rewards helpful content however it is made. It acts against content produced at scale mainly to manipulate rankings, whatever tool produced it.
How much does it cost to automate a blog with n8n?▾
The API costs for research and drafting are a few cents per post at November 2026 prices. n8n Cloud Starter is 20 euros a month billed annually, or you can self-host n8n for free. Editing time is the largest cost.
Can n8n post directly to WordPress?▾
Yes. n8n's WordPress node can create posts with a status of draft, pending, private, future or publish, using a WordPress application password. It cannot upload a featured image, so that needs a separate HTTP request to the WordPress media endpoint.
How many blog posts a week should a small business publish?▾
Fewer than you think. Two good posts a week on questions customers really ask is plenty for most local businesses, and one a week you have properly edited beats five you have not.
Is Arvow better than building it in n8n?▾
Arvow is faster to set up and handles publishing for you. n8n gives you control over prompts, sources and the draft-first workflow. Pick Arvow if you will not maintain a workflow, n8n if you want the editing step built in.