Open your own blog. Pick a post you paid somebody to write. Find one sentence that could only have been written about your business: a customer’s first name, a price, a town, something that went sideways on a Tuesday in March. Can’t find it in nine hundred words? Then you already know most of what there is to know about how to tell if content was written by AI, and you got there without installing anything.
That test isn’t airtight. Plenty of humans write generic filler. But it survives contact with reality, and you can run it before your coffee gets cold.
One disclosure before this reads as finger-wagging: we use AI here every day, in production. I’ve also gone through our own blog post by post, cleaning up what happens when it runs unsupervised. Both chairs. And if this is the first you’re hearing that software writes blog posts now, that’s real, and you’re not behind for missing it. Software makes a draft. A draft and a page that earns a customer are different objects.
Is “was AI used” even the right question?
No. The question that matters is whether anyone talked to you before writing it.
A writer who spent thirty minutes on the phone with you and then used a model to shape the draft produces something with your fingerprints on it. A writer who pulled a keyword off a spreadsheet and never called produces something with the industry’s fingerprints on it. Both used AI. Only one sold you something.
So “did you use AI” is a weak opening move. A bad vendor says no with a straight face because a human retyped the output. Ask about the input instead.
How to tell if content was written by AI in fifteen minutes, with no tools
Five checks, in the order I’d run them. None is proof alone. Three at once is a pattern.
- Look at the publish dates. Read the dates in your archive, not the titles. Nine posts stamped across one Tuesday and then nothing for four months is a batch job. Innocent explanations first: a migration can restamp every date, and a scheduled queue looks the same from outside. Rule those out, then treat it as a signal. Real calendars are lumpier.
- Line up the H2s from three unrelated posts. If a post on your service area and a post on seasonal pricing both run “What Is X / Why X Matters / Benefits of X / Conclusion,” that isn’t a house style. That’s one prompt, run five times.
- Search for numbers and proper nouns. Ctrl+F for a dollar sign. Then a year. Then your town. If a full year of content holds no prices, no dates, no places and no people, nothing from an interview reached the page, and usually there wasn’t one.
- Read the connective tissue. “In the ever-evolving landscape of [your industry].” “It’s important to note that.” “In conclusion.” Any one is a tired sentence. All of them, in every post, is a tell.
- Look for the seams. Generated text nobody read leaves debris. Auditing this blog, I found live posts with an orphaned line asking whether I’d like further adjustments, case studies about businesses that don’t exist, and an unrelated industry pasted mid-article. A few stopped mid-sentence. Nobody proofread. That matters more than the drafting method did.
Are AI content detectors accurate enough to use as proof?
No. And whatever you do, keep the score out of your email.
Detector output is a probability estimate produced by another model. A guess with a percentage sign stapled to it. Clean, well-edited human writing gets flagged constantly, because “short sentences, predictable structure, low vocabulary variance” is what good editing looks like and what a language model looks like. Researchers at Stanford ran TOEFL essays by non-native English speakers through seven detectors and more than half came back labeled AI-generated. Native-speaker essays scored accurately.
Send that screenshot and the meeting becomes an argument about your evidence instead of their work. Use a detector like a smoke alarm in the next room. Reason to go look. Not the fire.
What’s the one test your agency can’t argue with?
Whether the article contains anything that could only have come from your business.
Say you run a bakery. You know the question people ask before they ask price, the delivery radius you refuse to cross, the week in the fall you stop taking custom orders. That’s the content, and none of it lands in an article written off a keyword list, because nobody called to ask.
Some of you didn’t hire anybody. You opened a chat window, wrote a few posts yourself, and they came out clean. Grammatically perfect. Better formatted than anything you’d type at eleven at night. And they read like every other page on the internet, and you couldn’t name why. I can. The model had nothing from inside your business, so it wrote the average of everyone else’s. That isn’t a prompt you got wrong. No prompt contains the week you stop taking orders, or the objection you hear on every third call. Somebody has to pull that out of you and put it in an order that sells. A job. Not a setting.
Generic writing competes with every other piece of generic writing on the topic. Losing, no matter how tidy the prose. What competitors can’t copy is what happens inside your business, and getting that out of an owner’s head and onto a page is most of the job when we write and build for a client.
What should you actually email them?
Something short, curious, and impossible to get defensive about. Ask about process, not authorship:
Hey [Name]: quick question on the blog posts from the last few months. Can you walk me through how they get written? Who does the first draft, whether AI is part of that, and what review looks like before publish. Not trying to make this a thing. I just want to understand what I’m buying.
A reasonable answer: “We record an intake call, pull the specifics out of it, draft with AI against those notes, then an editor rewrites and fact-checks before you approve.” That’s a real process. AI is in it. Fine.
A bad answer dodges the input question. Watch for “all our content is 100% human-written and passes Copyscape,” a non-answer wearing a lab coat. Copyscape checks whether text was copied off another site. It has no opinion on who typed it. If nobody can name the human who reviewed the post, you have your answer.
What does good AI-assisted content look like?
Like a transcript turned into an article. Your phrasing, your objections, your prices, the thing you refuse to do for customers and why.
Let me be precise about what the machine is good for, because the internet is not. A model is fast at structure and fast at a first pass. That’s the list. What comes out is a starting point, not a finished business asset. It doesn’t know your margins, and it can’t tell a true story from a plausible one, so it will invent a delighted customer to fill a paragraph. Everything that makes a page worth publishing happens after the draft: the interview, the cut, the fact-check, the person who signs off. AI is a tool on the bench. It isn’t the process.
Cost explains most of what’s going wrong. Our full-service content pages run $275 to $375 each, because someone has to get the specifics out of you, write it, edit it, check it and publish it properly. A vendor selling eight posts a month for less than one page costs to do right is describing the process with the price. Same arithmetic as what a website costs to build.
Three lines for your next content scope
These three are in ours. Steal them.
- “Disclose the method.” AI-assisted drafting is fine. Publishing generated text with no human review is not. Tell us which we’re buying.
- “Every article requires one input from us.” A recorded call, an interview, or written answers. Nothing ships without one.
- “Names, numbers, dates and quotes get confirmed before publish.” Anything unconfirmed comes out. That line alone kills fabricated case studies.
It’s already published and it’s already bad. What do you fix first?
Spotting it is easy. Deciding what to do with two years of it is where people stall. Don’t mass-delete. Check your traffic data, then work in this order:
- Anything factually dangerous. Outdated security advice, wrong prices, instructions that no longer work.
- Anything fabricated. Fake case studies, invented research, customers who don’t exist. The real liability, and the fastest to remove.
- Anything with traction. Posts pulling impressions get rewritten with real detail, not deleted. Killing a page that earns attention is a self-inflicted wound.
- Everything else. Consolidate, redirect, or delete, with the numbers in front of you.
Now, who does all that? A two-hundred-post archive is not a Saturday. It’s weeks of reading, rewriting and redirecting, and every hour is an hour you aren’t running the business. I’ve done it to our own site and it ran long, and this is my job. If you’re reading this at 10pm because that’s the only quiet hour you get, hand this one off. Tell us what you’re sitting on and we’ll tell you what’s worth saving. If the site under it all is slow or breaking on phones, that’s the bigger fix: small business websites.
A budget note, since content is the first line most owners cut. Cut frequency if you must. What never goes in the cut column is maintenance, backups and security updates. A blog nobody adds to costs you attention. A site nobody patches costs you the site and every post on it.
Common questions
Does Google penalize content because AI wrote it?
Google’s stated position is about quality and usefulness, not production method. What gets punished is thin, mass-produced content, and generated filler tends to be both. The fix is the old fix: make the page worth the click.
Can I get a refund if my agency used AI?
Depends what your agreement said, and I’m a web guy, not a lawyer. Practically: a refund doesn’t get you content, and a vendor who couldn’t describe their process the first time rarely runs a better one the second. What you want is pages that pull their weight. Sometimes that’s a rewrite from them with a real interview attached. Often it’s a rewrite from someone who gets on the phone with you first, which is a conversation you can start.
Is it ever fine for my agency to use AI on my content?
Yes. We do, openly. The line is whether a human who knows your business shaped the input and reviewed the output before it went live. Drafting help is a tool. Publishing unread is negligence, and your name is on the page.
Should I disclose AI use on my own website?
I’m not aware of a US rule requiring it on ordinary marketing copy, and I’m still not a lawyer. Reviews and testimonials are a separate matter with real rules attached. Otherwise what counts is that the page is accurate and a human at your company stands behind it.
The part worth remembering
It points at one outcome: content only your business could have produced, because that’s the only kind that does any work for you. The five checks tell you what you’re sitting on in fifteen minutes. They won’t fix it. Fixing means someone sitting with you, getting the real stories out, and rebuilding pages that earn their spot.
So run the checks tonight, then send me the worst thing you find. Seriously. Send it over and I’ll tell you straight whether it’s salvageable or whether you’re better off starting clean, and the calculator has the numbers if you’d rather look first. Your business has real stories in it. Somebody ought to be getting them on the page.