Automated Content Creation: The Cost Moved From Writing to Reviewing

Automated content creation has an odd shape to its economics. One half of the job got about a thousand times cheaper in three years. The other half costs what it did before, and it is the half that decides whether any of this works.

Running our own pipeline across ten researched posts, the model usage came to $5.66 in total, or about 20 cents a post. Each one carried external research with cited sources, a full outline, and a dozen generated sections. A freelance writer quotes $100 to $400 for a 1,000-word post, and an agency $150 to $600 with retainers from $1,000 a month upward.

Read that as the writing being solved and you will publish forty mediocre posts and wonder why nothing ranks. The generation line went to near zero. The reviewing line did not move at all.

What the term covers

The phrase gets stretched across everything from a scheduling tool to a system that picks its own topics. In practice a working setup does five things, and they are worth separating because their costs are not comparable.

• Topic selection. Deciding what to write, from your own Search Console data and your competitors' rankings rather than a brainstorm.
• Research. Collecting the figures, standards and dates the piece will cite, each tied to a source that resolves.
• Drafting. Turning a brief into structured prose with headings, internal links and schema.
• Checking. Automated gates for keyword placement, reading level, working links and similarity against what you already published.
• Review and publish. A person reads it, decides, and pushes it live.

The first four now cost cents and minutes. The fifth costs a person's attention, and no amount of tooling has changed that. Most guides to this category compare tools on the first four, which is why they all end up recommending the same eight products.

The generation side, with real numbers

These are our own usage-log figures rather than a vendor's pricing page. The research call dominates, because it is the one that goes out to the web and comes back with citations.

The gap between those two totals is worth understanding, because it is the only cost lever left on this side. Every section call repeats the same style guide and business context. Sent fresh each time, that is about 44,800 input tokens per post. Held as one byte-identical prefix so the provider can cache it, the same post costs about 10,800. Anything you add to the system message that varies per section forfeits the discount.

Generated images sit outside those figures and cost more than the text does. They are also the easiest thing to switch off, and a stock photo will do for most posts.

The review side, which did not get cheaper

Survey data puts topic-to-published at 1.5 to 2.5 hours per piece once human review is counted. That number is where the honest accounting lives, and it explains the gap between what teams expect from automation and what they get.

The same research found that 41% of teams give a draft a 5 to 20 minute review, which has become the default posture. It also found what that posture costs: teams editing 20% or more of the word count report 2.7x better organic traffic than teams editing under 5%, with returns flattening past a 25% to 45% edit ratio.

Put those two findings next to each other and the picture is uncomfortable. The common practice is a skim. The practice that works is a rewrite of about a third of the piece. Most teams are doing the first and reporting the results of the second to their manager.

This is also why the tool comparison is the wrong comparison. Two products whose drafts need the same amount of editing are the same product for your purposes, whatever their feature tables say. What separates them is how much of the checking they do before a human opens the file: whether citations resolve, whether internal links work, whether the post duplicates something you published in March.

Working out your own ceiling

The arithmetic is short, and it is more useful than any tool shortlist. Take the hours you can give content review each week and divide by two, since a piece takes one to two hours of real review and you want the upper half of that range.

Four hours a week gives you two posts a week. That happens to be the cadence we recommend to clients, and it was arrived at from this direction rather than from a content-marketing benchmark. Above about forty a month the constraint stops being review time and becomes topic supply: your own search data cannot refill a queue that fast, so the briefs thin out before the writing does.

If the number you get is smaller than the number you wanted, the fix is not a faster generator. It is either more review hours or fewer, better posts. Our own publishing frequency guide covers what the cadence research says about that tradeoff.

Where Google draws the line

Google's position has been stable and it is narrower than the folklore around it. The official guidance on generative AI content says the company rewards quality regardless of production method. The policy that bites is scaled content abuse: many pages made to manipulate rankings rather than to help anyone.

Volume is the trigger in that policy, and volume is what an unreviewed pipeline produces. A team publishing forty skimmed posts a month is closer to the thing Google penalises than a team publishing eight edited ones, and the tool involved is identical in both cases.

Google also suggests telling readers how content was produced where that makes sense for your audience. A short line on your about page covers it. We go into the ranking question in more depth in does AI-generated content hurt SEO rankings.

Running it on WordPress

WordPress is the common case, and it is the easiest one, because the REST API has been in core since 2016. A pipeline can create a post with its categories, tags, featured image and SEO fields already filled in, and leave it as a draft for you. No plugin has to sit in the middle holding your content.

Two things are worth checking before you start. Your REST API needs to accept authenticated writes, which is the default unless a security plugin has closed it, and your SEO plugin needs to expose its title and description fields to the API. Yoast, RankMath and SEOPress all do.

Trying it free on your own site

BlogBERT is the engine behind the 141 guides on this site, and the WordPress version is open at no charge while early access lasts. We handle the setup and the calibration. You connect your own OpenRouter key, so the only bill is your own model usage at about the figures above.

Early access is open to businesses in the US, UK, EU, Canada, Australia and New Zealand for now. Check whether your site qualifies, or read how the managed service works if you would rather hand the whole job over.

What to take from this

Pick the tool on how much checking it does before you open the draft, not on how good its sample output looks. Set your cadence from your review hours rather than your ambitions. Then spend the money you saved on generation where it still buys something, which is the hour a person spends making the piece true.