We built an AI blog generator, then turned it off

Short answer: you can wire GPT or Claude to your Shopify blog through the Admin API in about a week, and publishing the raw output is the fastest way to get the whole domain treated as low quality. Google's policy is about intent, not authorship: AI-assisted content written for people is fine, bulk-generated content written for rankings is spam. The version that works uses your data, your process knowledge and a human edit. That version does not scale to 100 posts a month, which is the part nobody selling this mentions.

We built an automatic blog generator for this site. It wrote posts on a schedule, tagged them, generated cover images and published without anyone looking.

We turned it off and deleted most of what it produced.

This is what we learned, what the mechanics actually are if you still want to build one, and the version we would build now.

Does Google penalise AI content?

Not for being AI-written. This is the most misquoted policy in SEO.

Google's position, in their own words, is that they reward high quality content "however it is produced". What they act against is scaled content abuse: generating many pages primarily to manipulate rankings rather than to help people. Automation is not the trigger. Volume without purpose is.

The March 2024 core update and the spam policy update that accompanied it hit sites publishing high volumes of thin, templated content. Many were AI-generated. The ones that survived were AI-assisted but substantive.

So the real question is not "will this get penalised". It is "would a person who knows this subject find this page worth their time". If yes, publish it. If you cannot answer confidently, you already know.

There is a second effect that matters more than the penalty risk and gets discussed less. Thin pages consume crawl budget. A site with 30 good pages and 200 generated ones spends most of its crawl allowance on the 200, and the 30 get crawled less. You can dilute yourself without ever being penalised.

How does it work mechanically?

If you are building this, the pipeline is roughly:

1. Topic source. A queue, a keyword list, a spreadsheet. The weakest link in most builds is letting the model pick topics — it produces plausible-sounding subjects that nobody searches for.

2. Research input. This is what separates a useful post from filler. Product data, real support tickets, actual configurations, your own numbers. A model given nothing but a title writes what everyone else wrote.

3. Generation. A call to the Claude or OpenAI API with a structured prompt: outline first, then sections, rather than one shot at the whole article. Multi-pass output is consistently better and costs little more.

4. Structuring. Convert to the markup Shopify expects, generate the meta title and description, pick tags, produce an excerpt.

5. Images. Generated or pulled from your product media. Alt text from the same pass.

6. Human review. The step that decides whether the whole thing is worth having.

7. Publish. Shopify Admin API articleCreate on a blog, or write to your own CMS if the store is headless.

API cost is trivial — cents per post. The cost is entirely in the research input and the review.

The three versions of this, and what each is worth

Fully automatic is what most people mean by "automate my blog", and it is the one that does not work. Not because of a penalty, but because the output says nothing that is not already on twenty other pages. Nobody links to it, nobody cites it, and it competes with your own good pages for crawl attention.

Draft and review is defensible and genuinely saves time. The model produces a structured first draft from a real outline you supplied; a person who knows the subject rewrites the claims, adds the specifics and cuts what is padding. Fifteen minutes instead of two hours is a real gain.

Research-assisted is where automation actually belongs: gathering the inputs, not writing the output. Pull the support tickets on a topic, the search queries that hit that page, the competitor coverage, your own analytics. Hand a human a research pack and let them write. This is the version we would build, and it is the one nobody sells, because it does not produce a "posts per month" number for a sales page.

What do AI assistants actually cite?

If the goal is being recommended by ChatGPT, Claude or Perplexity — and for most people asking about this, it is — generic content is exactly wrong.

Those systems retrieve and cite pages that contain something specific and answerable. A number. A procedure. A named constraint. A direct answer to the question asked.

What gets cited, in practice:

• Specific numbers with context. "Shopify retries a failed webhook over roughly 48 hours, then discards it" is citable. "Webhooks can be unreliable" is not.
• Direct answers early. A short answer in the first paragraph, before the preamble.
• Real procedures. Steps someone actually followed, including what went wrong.
• Named trade-offs. Comparisons with an actual recommendation attached.
• First-hand experience. What happened when you did it, which a model cannot generate because it does not have it.

Generated content fails all of these by construction. It has no first-hand experience to draw on, so it produces the general case — which is the one thing retrieval systems have no reason to cite, because they already have it a thousand times over.

That is the argument against fully automatic publishing, and it is stronger than the penalty argument.

If you build it anyway

Some things are worth getting right regardless of which version you choose.

Never publish unreviewed. Not "review the first ten". Every post. The moment nobody is looking is the moment a factual error goes out with your name on it.

Feed it real inputs. Product data, tickets, transcripts, your own metrics. A model with nothing to work from produces the average of the internet.

Generate the structure, write the substance. Outlines, headings, FAQ scaffolding and meta descriptions are genuine time savers and carry little risk. The claims are where a human has to be.

Cap the volume deliberately. Not because of a limit, but because your own good pages are competing with your generated ones for crawl attention.

Check facts, especially numbers. Models produce confident, wrong specifics. Version numbers, pricing, API behaviour and limits are the usual casualties.

Keep a kill switch. Ours published for weeks before anyone read the output carefully. A weekly sample review would have caught it.

Write the FAQ section properly. It is the part most likely to be extracted into a rich result or an AI answer, and the part generation handles worst, because good FAQs come from questions real customers asked.

What we do now

We write posts by hand, with AI used for research, outlining and structure. Fewer posts, and each one says something we actually know.

The blog you are reading is about 25 posts. It took longer than 250 generated ones would have. It also gets cited, which the 250 would not have.

We build content and automation tooling for Shopify stores, including the research pipelines that make writing faster without making it worthless. We will also tell you when automating your blog is the wrong project.

Thinking about automating content? Book a free 30 minute call and we will be straight about what it will and will not do. Or read where Shopify automation actually pays, or see our SEO work.