How customers smell AI copy

Two versions of the same homepage paragraph. Both are demonstrations of style — no real shop, every bracket a blank where a real owner's facts would go. Read both at your normal speed.

Version one

We do [your trade] in [your town]. Two trucks, [number] guys on them, at it since [year]. Call before [time] and one of us can usually get out the same day. The price we give you on the phone is the price on the bill.

Somewhere around "tailored to your unique needs," you stopped listening. Not reading — listening. Every blank in version one is a fact you could fill from memory in a minute — it's a frame, steal it. Version two has no blanks, because nothing in it was ever going to be a fact.

That reaction is the whole page. Your customers have it too. Here's how to name it and keep it off your site.

The checklist we run on our own pages

Every page on this site gets audited against a list of slop tells before it ships. Here it is, translated out of web-shop language. Yours now.

The tell	What it sounds like
The same question answered all the way down	Six headings all secretly asking "why choose us?" — answered the same way six times.
A divider after every couple of sentences	Two lines, a break, two lines, a break. Real explaining runs longer than two sentences.
Everything comes in threes	"Quality, integrity, and customer satisfaction." "Fast, friendly, and affordable." The machine loves a matched set of three.
No numbers, no names, no place	Nothing anyone could check. No year, no street, no truck count, no person's name.
The logo swap	Paste a competitor's name over yours. If the page still reads fine, it was never about your business.
The confident paragraph that says nothing	It sounds sure of itself, and when it's over you can't repeat one checkable thing it said.

One tell alone proves nothing. A lone em dash or a fancy word like "delve" is no fingerprint — a 2025 review by Jisc's National Centre for AI (a UK body that tracks AI detection for schools and universities) calls most single-word "AI tells" unreliable. The list measures whether a page feels canned; it doesn't convict any one sentence. What customers smell is the pile-up.

The detector software failed. People didn't.

You'd figure software could scan for these tells. The AI industry figured that too, briefly.

OpenAI — the company behind ChatGPT — shipped its own AI-text detector in January 2023 and killed it that same summer. The shutdown notice is still sitting on the announcement page, live as of July 20, 2026, verbatim: "As of July 20, 2023, the AI classifier is no longer available due to its low rate of accuracy." While it ran, its own published numbers had it catching just 26% of AI-written text and wrongly tagging human writing as AI 9% of the time. Same page, their words: "Our classifier is not fully reliable," and "AI-written text can be edited to evade the classifier."

They haven't changed their minds. OpenAI's help center (its own timestamp: updated July 14, 2026) asks "Do AI detectors work?" and answers itself: "In short, not in our experience." Their own test detector, the same page says, "labeled human-written text like Shakespeare and the Declaration of Independence as AI-generated."

Third-party checkers exist, and schools lean on them. The same June 2025 Jisc review found AI generators "would generally outpace the development of AI detection tools," paid detectors "are relatively easy to circumvent," and some free ones post alarming false-alarm rates — one flagged human writing as AI half the time. The same market sells "humanizer" tools built to dodge the detectors and the detectors meant to catch them — and by the detectors' own published numbers, the catching half doesn't work.

Here's what does work. A 2025 peer-reviewed study published at ACL, a major language-research venue, put human readers up against the detectors. Five people who frequently use ChatGPT for writing, voting as a group, misclassified exactly 1 article out of 300 — beating most commercial and open-source detectors, even against paraphrased text and "humanizer" tools built to dodge detection. They did it by reading: for AI vocabulary, formality, originality, and clarity. People who rarely used the tools scored close to a coin flip.

Read that finding like a mechanic: the skill is mileage. People who see AI text every day smell it almost perfectly; people who never see it can't. The only question for your website is which side of that line your customers are on next year.

The fix pass

When a draft trips the tells — yours, a vendor's, or a machine's — here's the pass we run. Four moves, no software.

1. Read it out loud. Actual voice, not in your head. Slop survives on a screen and dies in the mouth.
2. Cut what you wouldn't say. If a sentence isn't something you'd say to a customer standing in front of you, it goes. Whole paragraphs will go. Let them.
3. Put a number, a name, and a place back in. Real ones, yours: the year you started, the street you're on, how many trucks, who answers the phone. If you didn't hand the machine a fact, it doesn't get to be on the page.
4. Answer each question once, well, and move on. If two sections answer the same question, one is filler. Keep the better one.

One boundary note: Google's guidance (updated December 2025) warns that using AI "to generate many pages without adding value for users" can violate its spam policy — the trigger is empty pages at scale, not AI itself. The full policy picture lives on the SEO page.

Next: the prompt method — how to load the tool with your real facts so the output isn't slop to begin with.