40 Days of Building With Claude. Then We Counted the Drafts.

Most AI case studies get written after the fact, with the losses quietly trimmed out. This one is different. We opened a session with Claude this morning, pointed it at seven client WordPress sites, and told it to pull the live numbers — published posts, draft queues, publish dates, all of it.

What follows is exactly what came back on September 9, 2026. Including the parts that make us look bad. This is original reporting on our own operation, not aggregated coverage — see our editorial standards for how we label the difference.

The scoreboard, pulled live this morning

Every number below came straight out of each site’s own WordPress install. No dashboard screenshots. No rounding up.

That is 122 drafts sitting in queues across the portfolio. Hold that number. We come back to it.

Four silent client sites, forty days, ten published pieces

Four of these sites had published nothing in months. One had been quiet since late 2024. Our own internal registry had them flagged as dormant, which is a polite word for a client paying for a website that does nothing.

Since August 1, those four sites have published ten new pieces. Here is what actually shipped.

Excel Floor Cleaning & Janitorial Service — 3 posts

A commercial cleaner in Las Vegas competing for facilities-manager contracts. The new work targets buying decisions, not vague “cleaning tips” traffic:

• Commercial Water Damage Restoration Las Vegas: The 48-Hour Rule
• How to Vet a Commercial Cleaning Company in Las Vegas
• Commercial Floor Cleaning in Las Vegas: A Facilities Manager’s Guide

Notice the vetting post. It tells a buyer exactly how to interrogate a cleaning vendor — insurance, Nevada workers’ comp, licensing jurisdiction, cancellation terms. That is a page written for someone about to sign a contract. That is the whole point.

Fathers Rights Help Line of Las Vegas — 3 posts

This site had a fully configured storefront and almost no content feeding it. The new posts are timed and procedural: school records access when school started August 10, what a first Clark County hearing actually is, and custody documentation. Every piece carries a plain disclaimer. It is not a law firm and never pretends to be.

Family Court Help Line — 3 posts

Deadline content, which is the highest-intent thing you can publish in this space. How to respond to a motion in 14 days. How Nevada actually calculates child support under NAC 425. Someone searching that phrase at 11pm has a court date.

Equal Rights for Divorced Fathers — 1 post

One piece, on Nevada SB 432 and what changed when divorce records became public by default. Statute-anchored, dated, verifiable. One good post beat five thin ones.

We thought we had a 122-post backlog. We had two broken feeds.

Ten published posts in forty days is not a miracle. A decent writer does that.

The change looked like it was upstream. When we pulled counts this morning there were 122 drafts staged across the portfolio — 88 on this site alone. The obvious read: writing stopped being the bottleneck, and reviewing became one.

So we opened the queue and classified every item in it. That read was wrong.

Fifty-two of the drafts on this site are exact duplicates — 26 articles pulled in twice because two RSS feeds overlap. Twenty-one drafts sitting on the Philippine Times of Southern Nevada are syndicated tech stories about Nvidia and Chrome patching. That is a Filipino-American community newspaper. A feed is pointed at the wrong site. Two more drafts on the advocacy site are titled, in plain text, [SUPERSEDED] and [WITHDRAWN — FACTUALLY INACCURATE].

Strip all of that out and the real queue is about 13 original pieces.

Thirteen is a Tuesday. A hundred and twenty-two is a crisis. We had been managing the wrong number, and we would have kept managing it if nobody had opened the pile and counted.

That is the actual lesson, and it is not the one we sat down to write. Cheap production does not hand you a review backlog. It hands you volume you quietly stop inspecting — and volume nobody inspects is indistinguishable from the AI slop everyone claims to be against. The dangerous number was never the size of the queue. It was that we had not looked inside it.

What actually changed in the workflow

A client registry, not a prompt

Every site has a locked profile: exact business name, exact phone number, service areas, which post types exist, what the site is legally not allowed to claim. Nothing gets written until that profile is loaded.

This kills the failure that scares agency owners most — one client’s phone number landing on another client’s page. Local rankings punish inconsistent business details hard, and an AI with no guardrails will absolutely reformat a phone number to look tidier.

Nothing goes live without a source

The family court sites cite Nevada statutes and court rules. Those get checked against the current rule, not recalled from memory. A wrong filing deadline in a custody article is not an SEO problem. It is harm to a parent.

Draft status is the default

Every write to a client site lands as a draft unless a human says publish in that same breath. Agents are eager to finish the job you gave them — that is the whole risk profile in one sentence. The default has to be safe, because the agent will take the fastest path to “done.”

That rule covers what the AI writes. It did not cover what our RSS importer does. Keep reading.

What went wrong today

Four things broke while we assembled this post, and leaving them out would make this an ad.

The rank data never arrived. Every Ahrefs call for organic keywords and traffic returned “Insufficient plan.” So this post has no ranking numbers in it. We would rather publish verified counts than dress up estimates.

The image generator was out of credits. The AI image tool on this site failed on an exhausted API balance. The graphic at the top of this article got built from the actual numbers instead. Honestly, better outcome.

Two sites are still publishing into /uncategorized/. Look at the Family Court Help Line links above. Real URLs, live right now, sitting in the default WordPress category. Fast content production exposed a site structure nobody had fixed. That is on us, and it is this week’s job.

Our own feed was publishing without us. We pulled counts twice today, about an hour apart. This site went from 1,337 published posts to 1,341. Four posts went live in that hour and no human touched any of them — the RSS importer publishes on import rather than staging a draft. We wrote an entire section above about keeping a person on the last gate, while our own aggregator quietly did not have one. Fixing that setting is the first item on the list.

Speed does not fix a broken foundation. It finds it faster.

What we would tell another agency owner

Do not start with content. Start with the registry. Write down every client’s exact business name, phone, service areas, post types, and compliance limits before an AI touches a single page. That document is the entire difference between scale and a mess.

Then go count what is actually in your draft queue. Not the number on the dashboard — open it and classify every row. We were one afternoon away from hiring against a backlog that did not exist. A dashboard count tells you how big the pile is. It does not tell you that three quarters of it is the same article twice.

Then keep a human on the last gate. The tooling market is moving fast and consolidating faster, and every platform will promise to automate that gate away. Do not sell it.

Frequently asked questions

Does Claude publish directly to client WordPress sites?

It can write to them, but everything it writes lands as a draft by default, and a person approves publication. Automated RSS syndication is a separate system with its own settings — as this post describes, ours was publishing without review until today.

How do you stop AI from inventing facts in client content?

Two rules. Business details come from a locked client registry, never from the model’s memory. Any checkable claim — a statute, a filing deadline, a percentage — gets verified against the primary source before publication, and cited in the article.

Is AI-assisted content against Google’s guidelines?

No. Google’s guidance targets content produced primarily to manipulate rankings, regardless of how it was made. Helpful, accurate, original content is fine whether a human or an AI drafted it. The standard is the content, not the tool.

How long until AI-assisted content shows results?

Longer than the volume suggests. New pages on a low-authority domain typically need several months to settle into stable positions. Forty days of publishing tells you production is working. It does not yet tell you rankings are.

The next check-in

We will run this same live pull again in 30 days and publish whatever it says. If the drafts stall or the traffic does not move, that gets published too.

Running a portfolio of sites and stuck at the production bottleneck? Tell us what your queue looks like — we will tell you honestly whether this workflow would help or just make your backlog bigger.

{“@context”:”https://schema.org”,”@type”:”FAQPage”,”mainEntity”:[
{“@type”:”Question”,”name”:”Does Claude publish directly to client WordPress sites?”,”acceptedAnswer”:{“@type”:”Answer”,”text”:”It can write to them, but everything it writes lands as a draft by default, and a person approves publication. Automated RSS syndication is a separate system with its own settings – as this post describes, ours was publishing without review until today.”}},
{“@type”:”Question”,”name”:”How do you stop AI from inventing facts in client content?”,”acceptedAnswer”:{“@type”:”Answer”,”text”:”Two rules. Business details come from a locked client registry, never from the model’s memory. Any checkable claim – a statute, a filing deadline, a percentage – gets verified against the primary source before publication, and cited in the article.”}},
{“@type”:”Question”,”name”:”Is AI-assisted content against Google’s guidelines?”,”acceptedAnswer”:{“@type”:”Answer”,”text”:”No. Google’s guidance targets content produced primarily to manipulate rankings, regardless of how it was made. Helpful, accurate, original content is fine whether a human or an AI drafted it. The standard is the content, not the tool.”}},
{“@type”:”Question”,”name”:”How long until AI-assisted content shows results?”,”acceptedAnswer”:{“@type”:”Answer”,”text”:”Longer than the volume suggests. New pages on a low-authority domain typically need several months to settle into stable positions. Forty days of publishing tells you production is working. It does not yet tell you rankings are.”}}
]}

Originally published on Digital Media Marketing Technology.