Full Autopilot or Draft Approval: Choosing Per Site

You run three sites. One pays your rent, two are experiments. If you're hand-approving every draft on all three, you're spending review time where a bad post costs nothing. And if you've put everything on autopilot, one fabricated citation on the money site can do real damage.

This post gives you a two-axis table (site stakes versus trust in output quality) and the signals that place each of your sites on it. By the end you can set the automation level per site in one sitting. You need a ContentRails account with your sites added as projects, and an honest answer about which site can afford a bad post.

You set the automation level per site, not per account

ContentRails offers four automation levels (Off, Suggest topics, Write drafts, and Full autopilot), and the level is chosen per project, under Settings (contentrails.ai, contentrails.ai). Not per account. Not per workspace. Per project.

This is the right shape for the decision, because no single rule fits every site you run. A rule that fits your flagship blog wastes your time on a side project: you're reviewing drafts nobody would miss if they were imperfect. The reverse is worse. Full autopilot on your money site means a fabricated statistic can go live without anyone seeing it first.

The pitfall here is assuming one account-wide setting. If you've never opened Settings on each project individually, you're running someone else's default. Open Settings on every project before you decide anything.

And know what you're accepting when you turn autopilot on. ContentRails' terms make you responsible for everything published on your sites, including posts published automatically under Full autopilot. The tool checks figures against retrieved sources, but generated content can still contain errors, and review before publishing stays your job (contentrails.ai). Autopilot doesn't transfer liability. It only removes the gate.

Axis 1: Site stakes — what a bad post costs you

Multi-site operators commonly tier their sites by revenue potential, SEO risk, and brand importance, giving flagship sites the most oversight and low-stakes sites the automated pipelines (WordPress Multisite Development for Media Companies). That tiering is the first axis. Place each site on it with concrete signals, not vibes.

High-stakes signals: the site drives sales or leads. It ranks for terms your business depends on. Your name is on the masthead. A wrong figure there isn't an embarrassment, it's a cost.

Low-stakes signals: test domains, niche experiments, sites where a weak post just sits unread. If the realistic worst case is "nobody noticed," that's low stakes.

The pitfall: rating stakes by traffic instead of cost of error. A small site can still be high-stakes if it's your legal or brand face. Ten visitors on a page that misstates your pricing does more damage than ten thousand visitors on an experiment blog.

Axis 2: Trust in output quality — earned, not assumed

Trust is a property of the pipeline plus the topic, not a feeling about the model. The research on AI citations is blunt: SourceVerify reports studies consistently showing 30% to 70% of AI-generated citations are fabricated. And hallucination rates vary sharply by model — one study measured DeepSeek-V3 at 23%, GPT-5.3 at 69%, and Grok-4 at 73%. "AI wrote it" tells you nothing on its own. There is no safe default model.

What earns trust is structure, not vibes. ContentRails builds a cited fact sheet before drafting, so the draft can only cite what was retrieved. The model can't write prose first and invent evidence for it later, because there is no step where that happens. Automated checks and an editor pass run before a draft reaches you.

That structure is what makes the trust axis measurable instead of a hunch. You can open a draft's sources and check whether each figure traces back. Ten minutes across a few drafts gives you an actual answer.

The pitfall: granting trust after one good draft. Quality varies by topic. A pipeline that performed well on subjects your site covers deeply can stumble when the site moves into new subject matter. Re-check the sources whenever a site starts writing about something it hasn't covered before.

The table: four quadrants, four settings

Put the two axes together and you get four quadrants, each with a matching setting.

	Low trust	High trust
High stakes	Write drafts: every post waits for your approval, sources attached	Still Write drafts: skim faster, but approval stays on
Low stakes	Suggest topics or Write drafts with light review: read the sources, not the prose	Full autopilot

High stakes plus low trust is the default state, and Write drafts fits it: drafts are not published until you approve them unless Full autopilot is on, and every draft carries its sources (contentrails.ai). Read the sources, then approve.

High stakes plus high trust is where people get tempted. Quality is proven, so why keep approving? Because it's the money site. Approval stays on even when the pipeline has earned your confidence. You just skim faster, checking the fact sheet against the draft instead of reading every word.

Low stakes plus low trust: use Suggest topics, or Write drafts with a light review. The efficient move is to read the sources, not the prose. If the evidence holds, the post is probably fine for a site where a weak post just sits there.

Low stakes plus high trust is the only quadrant where Full autopilot is acceptable. Both conditions matter. Trust alone isn't enough — see the next section.

Why the money-site rule is not paranoia

The failures that end up in the news all share one trait: a high-stakes site, no approval gate doing real work.

CNET published dozens of AI-written financial explainers in early 2023. Fact-checkers found inaccuracies including basic math errors. CNET paused the program and issued corrections (oyova.com). In 2025, the Chicago Sun-Times and Philadelphia Inquirer ran an AI-generated summer reading list that was mostly fake books (kvia.com). Both are established publications with editors. The gate failed anyway.

And if you think a careful human read catches everything: an arXiv study found 100 fabricated citations in 53 published NeurIPS 2025 papers, each of which passed review by 3–5 expert researchers. Experts missed them. A rushed skim of your own draft will miss them too.

The lesson isn't "never use AI." It's that on a high-stakes site, the approval step has to be a real check, not a formality. And on the sites where you can't afford that check, that's exactly where autopilot belongs.

How to check it worked, and what to do next

Three checks, then one next step.

First, check the pipeline log. Each project runs the same logged sequence: plan, research, cited fact sheet, outline, draft, checks, editor. Open it on each site and confirm what actually ran. If you expected autopilot on one site and approval on another, the log shows it.

Second, spot-check one autopilot post per month. Open its sources and confirm the figures trace back to them. That's the whole audit. If a figure doesn't trace, drop that site's automation level until you understand why.

Third, confirm isolation. Each project keeps its own DNA, topics, and posts, with nothing bleeding between sites. A setting on one site can't leak to another, which is what makes per-site decisions durable.

For the next step: start on the free plan. It runs one project end to end with three posts a month (contentrails.ai). Use those three posts to calibrate the trust axis — read the sources on each draft, see whether the figures hold — before you commit any site to autopilot.

The decision rule, in one line: if a bad post on this site would cost you money or your name, it waits for your approval. If it wouldn't, let the pipeline run.

So do this now. Open Settings on your lowest-stakes site and set it to Full autopilot. Set your money site to Write drafts. Then, before you touch anything else, open the first autopilot post's sources and check that every figure traces back. That check is what earns the next site its autopilot.