What happens when you ask AI to write for your business?

My website was written with a lot of help from AI. It writes my first drafts, and I say so on my How I work page. I’d gone over every page more than once before launch. In July I stopped reading and counted a few patterns across the site instead, and the count found what my reading hadn’t.

What does AI-written copy get wrong?

Very little you can point to in one sentence. Every line on my pages was grammatical and clear enough. The problem was how often the same few moves came back.

I re-ran the count this month on the July version of the site. Across twelve pages and about 7,600 words, it had:

• 204 em dashes, about 2.7 for every 100 words. One sentence in four had a dash in it, and plenty of those had two.
• 22 “this, not that” constructions, like “A partnership, not a hand-off.” One of those is a good sentence. Twenty-two is a habit.
• The same phrase used over and over. “Credited toward” appeared 11 times across the site, and “page-by-page” 7.
• The insider hook, “Here’s the part most owners miss,” which promises a secret and then delivers an ordinary fact.

People use every one of these too, usually less often. One dash on its own proves nothing, and plenty of good writers use a lot of them. What matters is how often they come back.

Not everyone can see it, either. In a 2025 study of how well people spot AI-written articles, four readers who didn’t use AI to write picked out AI-written articles at about the rate of a coin toss, and were confident they’d got it right. Five readers who used it for writing a lot caught almost all of them. I think I can spot it pretty easily, and maybe that’s because I use AI myself. It’s a small study, and it tested nonfiction articles rather than business copy, so I wouldn’t lean on the numbers. It’s still a good reason not to assume nobody will notice.

Why didn’t I see it myself?

In someone else’s writing I usually can. On my own stuff I couldn’t. I could feel it wasn’t quite right, and I knew something was wrong because I kept working on it. I just couldn’t see what. Each page still read fine on its own. When the same habit runs through every page, it stops standing out — it just sounds like the way the site talks.

Reading for it hadn’t worked, so I had AI help me write a small script that counts these patterns. I run it on my posts and pages now, this one included.

My first count came out near five em dashes per 100 words. Re-running it for this post, counting only the words a visitor sees, gave 2.7. Most likely, the first count had included text each page carries in its code for search engines and link previews — that text repeats the copy, so a lot of dashes were counted twice. The 2.7 is the one I’d stand behind.

Does cutting the patterns fix it?

It helps. I took the em dashes from 204 down to 24, cut the “this, not that” constructions from 22 to 8, and thinned out the repeated phrases. The pages read better for it.

I probably overdid it. Twenty-four dashes in 7,500 words is close to none, and I think copy with every edge sanded off sounds processed in its own way. I’d now leave one of each on a page.

Counting also has a limit. A reviewer read a set of LinkedIn drafts I’d prepared, all of which passed every count, and said three of the four sounded “a little too AI like.” It was the structure, he said, a very common format on LinkedIn. All four ran the same shape: a hook, a reframe, a payoff, and a question at the end. No edit to a sentence changes the shape of the post around it.

Can AI get the facts about your business wrong?

Yes. A LinkedIn draft of mine once said “when I ran it on myself in August.” A check had run in August — an automated one — and I don’t remember doing it myself. The sentence sounded exactly like me, which is what made it easy to miss. The post was cut before it went up.

On a business website, the same thing looks like a service you don’t offer, a guarantee you don’t make, or a number nobody measured. It reads well, so nobody stops to question it.

What should you check before it goes out?

If AI drafted it, these are the checks I’d make before it goes on your site:

1. Every fact about your business. Services, prices, towns, years in business, anything that sounds like a promise. How well a sentence reads doesn’t matter if it isn’t true.
2. The last line of each section. If it sounds like it belongs on a poster, cut it and see whether the section is worse. Usually it isn’t.
3. Repeats across the whole page. Read it top to bottom in one sitting and listen for the same phrase or the same sentence shape coming back. The particular habits change as the tools change, so repetition of any kind is a better thing to listen for than a list of banned words.
4. Read it out loud. If you wouldn’t say a sentence to a customer across the table, rewrite it the way you would.
5. Leave some of it in. A dash, a short punchy line, one “this, not that” per page is fine. People have habits too.

I’m going to keep using the tool. It gets a first draft down faster than I do, and editing a draft is less work than starting from a blank page. Knowing what’s true about your business, and how you’d say it, is still your job or the job of whoever you hand it to.