Why ChatGPT content sounds like ChatGPT, and what it costs you

You paste a prompt into ChatGPT, read the output, and it looks fine. Clean grammar, sensible structure, nothing obviously wrong. So you publish it. Then someone who was about to ring you lands on that page, reads two paragraphs, and leaves. You never find out why.

The short version: they could tell. Not in a technical way, and not because they were looking for it. They just got the feeling they were reading something nobody had actually written. That costs you the enquiry, and it costs it silently, which is why almost nobody catches it.

The tells you will recognise the moment they are named

Most owners can sense AI writing without being able to point at it. Once you can name the patterns you see them everywhere, including on your own site.

The opener that clears its throat and says nothing

"In today's competitive market, finding a reliable plumber can be a daunting task. Homeowners across the region want a service they can trust." Two sentences, zero information. Nearly every AI draft starts this way, because the model produces the average of everything ever published on the topic, and the average opener is filler.

A person writing about their own trade opens with something they know. "Most of the burst pipes I get called to in winter are in places built before 1970, and the copper is usually fine - it is the joint that goes." Nobody else could have written that sentence, and it does more work than three paragraphs of preamble.

Hedging where a real tradesperson would just say it

AI is built to be agreeable and to avoid being wrong, so it softens everything. "It is generally recommended that gutters be cleaned periodically, depending on your individual circumstances." A roofer says: clean them twice a year, and if you have got a big liquidambar over the house, four.

People read your page because they want an answer. A page full of "it depends" reads like somebody who does not know, or will not commit.

Everything arrives in threes

Tidy three-item lists for things that are not lists. "Our approach is built on quality, communication and care." In reality some things have two parts, some have seven, and plenty have one. Three is a rhythm the model falls into because it reads well and costs nothing. The tell is that the list was invented to fill a shape, and the third item never carries any weight.

The same eight adjectives as every other business in the country

Seamless. Robust. Comprehensive. Tailored. Bespoke. Dedicated. Holistic. Cutting-edge. None of these are bad words. The problem is that everybody gets handed the same ones. When your clinic page promises "comprehensive, tailored care from a dedicated team", so does every other clinic page in Melbourne, so it carries no information about you.

The same cheerful register no matter the subject

AI writes at one temperature: mildly upbeat, evenly paced, faintly enthusiastic, whether the subject is a bathroom renovation, a switchboard that has been arcing, or grief counselling. Real writing changes with the subject. A psychologist writing about a first session goes quiet and careful. A sparky writing about a dangerous board gets blunt. That level register reads as somebody who does not much care.

Nothing that only you would know

This is the big one, and it contains all the others. Read a generated page and ask one question: what is on here that this business alone could have written? Usually nothing. Every sentence could be lifted onto a competitor's site with a find-and-replace on the name.

If a competitor could publish your page word for word without changing anything but the phone number, it is not your page. It is everyone's page.

- Tom McSherry

What it costs, and it is not rankings

Most people expect this to be a ranking problem. Mostly it is not. Google is not especially good at telling generated writing from written writing, and its stated position is that it cares whether a page is useful, not how it was produced. There is a real ranking cost to publishing at volume, but that is a different failure with a different cause, covered in the hidden cost of AI articles at scale.

The cost here is conversion. The page can rank beautifully and still lose you the job.

You have given the visitor no reason to choose you

Someone comparing three roofers in Christchurch, or three physios in inner-north Brisbane, opens three tabs. If all three say broadly identical things in broadly identical language, the comparison stops being about the work and becomes price, review count, or whoever answers the phone first. You had one chance to be memorable and you spent it on "quality workmanship and exceptional customer service".

It bites hardest where differentiation is already hard: multi-location groups where every branch page reads the same, clinics competing on the same handful of services, retailers up against national chains.

The writing implies nobody could be bothered

This one is worse, and owners underestimate it. If the page describing your service reads as though no effort went into it, the visitor draws the obvious conclusion about the service itself. Fair or not, people judge the parts they can see. A faded sign tells you something about a shop before you walk in.

It is sharper again for anything involving trust. Someone choosing a psychologist, or a builder for a house they will live in for twenty years, is trying to work out what kind of person they are dealing with. A page written by nobody answers that badly.

Why better prompting does not fix it

The standard response is prompt engineering. Feed it your tone of voice, give it sample paragraphs, ban the word "seamless". That helps at the margins and it is worth doing. But you can sand the surface smooth and still have a page that says nothing only you could say, because the model has nothing of yours to say.

AI is extremely good at arranging information and completely incapable of having any. It has never been on one of your jobs, never sat in one of your sessions, never had the awkward phone call about a quote that blew out. Everything it knows about your trade is the average of what has already been published about your trade, and that average is the exact thing that makes it sound generic. You cannot prompt your way out of a missing-information problem.

The four things AI cannot invent

These come from you or they do not exist. Give the model them and it will arrange them well. Withhold them and no prompt saves the page.

What you actually charge

A real number, or a real range, and an honest explanation of what moves it. Pricing is the most-wanted piece of information on most service pages and the one most businesses refuse to put there. Nobody else can write yours. Even a rough band with the variables spelled out beats "competitive pricing".

What goes wrong on jobs

The problems that keep recurring. What you find behind the wall in villas of a certain era. The reason a job runs a day over. The mistake customers make in the week before they ring you, which turns a small repair into a big one. AI guesses at these and the guesses are bland. Your version has the era of the house, the part that fails, and what it costs by then.

The thing you tell customers that they never expect

Every trade and every practice has one. The advice that visibly surprises people on the phone, usually the thing that talks you out of work: you do not need the full replacement, the cheap fix will hold another five years, come back in spring. Put that on the page and it builds more trust than the rest of the page combined.

An opinion you would defend

Something you believe about your field that a competitor might argue with. Not manufactured controversy - a real position you arrived at by doing the work. AI will not produce one, because it is trained to avoid taking sides. Ask for a strong opinion and you get a balanced overview of the opinions available. A position gives a reader something to agree with, and agreeing with someone is most of how people decide who to hire.

How to get this out of your head and onto the page

The practical version takes fifteen minutes and it happens before you open ChatGPT, not after.

• Open the voice recorder on your phone and answer five questions out loud: what did I charge on the last three of these jobs and why were they different; what went wrong on the most recent one; what do I tell customers that surprises them; what do I disagree with about how my industry works; what would I want to know if I were the one ringing.
• Get it transcribed. The transcript is your source material, not your topic.
• Give AI the transcript and ask it to organise what you said, rather than asking it to write about the subject. Only one of those produces something specific.
• Read the draft with one question in mind: could a competitor publish this sentence unchanged? Every sentence where the answer is yes is doing nothing. Cut it or replace it from the transcript.
• Keep the file and add to it after unusual jobs. Six months of that is enough raw material for every page on the site.

That reframes what AI is for. It is not a writer. It is a very fast typist and editor working from material you supply, which is roughly the whole argument in my guide to using AI for your website content without wrecking it.

Yes, this blog is written with AI

It would be a bit rich to write all of that and not say so. This post was drafted with AI doing the mechanical work: turning notes and positions into structured prose, keeping formatting consistent, catching repetition, tidying sentences that ran long.

Every judgement call is mine. What to write about, what position to take, what to cut, which client situation is a fair example and which is too identifiable to use, whether a claim is true. Every specific here came out of client work and my own head, not the model's general knowledge about SEO. The model arranged them. It did not know them.

That is the entire difference, and it is the same difference on your site. It is why I keep a human in the loop on everything, and if you want the detail of where the line sits day to day, here is how I actually use AI for SEO.

The test to run this week

Open the last page you generated and delete every sentence a competitor could publish unchanged. Whatever survives is your actual page, and even if it is three sentences it beats what you had. Build back out with things only you can say: the numbers, the failures, the surprising advice, the opinion.

Generic AI content is not usually punished by Google. It is punished by the person who read it, found no reason to pick you over the other two tabs, and rang someone else. You will never see that in your analytics, which is exactly why it is worth fixing before you publish the next one.