Using AI for your website content without wrecking it

Tom McSherry
18 July 2026 · 10 min read
The short version: when AI content damages a website, it is almost never because the writing was bad. I have read a lot of AI-generated pages that sat on sites which went backwards, and the prose was usually fine. Grammatical, clear, on topic, better than a fair chunk of what small businesses had on their sites beforehand. The damage came from everything that was missing around the writing.
This is written for you if you own the business and you have been using ChatGPT to write pages for your own website. Not for an agency running a content factory. You, on a Sunday night, pasting a prompt in because the site has been half-finished for years and this finally feels doable. That is a reasonable thing to be doing, and I want you to keep doing it without creating a problem you will pay someone to unpick later.
Four gaps cause almost all of the harm I see. No keyword map, so the buildout competes with itself. No context, so the model invents things about your business. No link plan, so internal linking is the first thing to fall over. And no real voice, so readers recognise the pattern and disengage. None of those are writing problems. All four are structure problems, and all four are fixable by you, in an afternoon, before you generate anything.
The failure is never the writing
It helps to be precise about what a large language model does well. It is a very fast, very confident writer with no memory of your business, no stake in your rankings, and no view of your website as a whole. Ask it for a page about emergency plumbing in Tauranga and you get a competent page about emergency plumbing in Tauranga. It has no way of knowing you already have three pages that half-cover the same thing, or that the phrase you actually get calls from is 'burst pipe'.
So it writes the page you asked for, in isolation, every time. Twenty pages later you do not have a content library. You have twenty islands. Each one is fine on its own and the collection is a mess, because nobody was holding the map. That is why 'the writing quality has got so much better' has not fixed anything. Better writing does not produce a better structure any more than better bricks produce a better house.
Every failure I have unpicked came from something nobody asked the AI to do, because nobody realised it was a job at all.
Gap one: no keyword map, so your AI content competes with itself
A keyword map is a boring spreadsheet that says which page on your site owns which search term. One page, one pool of demand. It is the single most valuable document a small business website can have, and almost nobody has one.
Without it, the instinct with AI is to build a page per keyword, because pages are now cheap. That instinct is exactly backwards, and the long version of why is in more pages usually means worse rankings, not better. The short version is cannibalisation (where two of your pages target the same search term and compete, so both rank worse than one strong page would have). Google picks which of your pages to show, it picks inconsistently, and your position wobbles between two mediocre results instead of settling on one good one.
AI makes this worse for a specific reason: it will happily write 'physiotherapy for back pain', 'back pain treatment' and 'lower back physio' as three separate pages, because they are three separate prompts. They are not three separate pools of demand. They are one pool with three labels, and they belong on one page. That is the cannibalisation trap in an AI buildout, and it is worth reading on its own before you generate a batch, because the checks that catch it all happen before you publish rather than after.
Here is the minimum version you can build yourself before you generate a word:
- List every page you have, with its URL and the one search term it should own.
- Write down the terms you want to rank for that no page currently owns. Those are your genuine gaps.
- Group anything that is the same question asked differently. One group equals one page, not three.
- For each new page, write the single term it owns before you write the prompt, and put that term nowhere else.
- Add a row every time you publish. That is what stops the drift six months from now.
That sheet is the thing the model cannot produce for you, because it requires knowing what you already have and what you actually want to be found for. It is thirty minutes of work and it prevents the most expensive failure on this list. If you want that step worked through properly rather than summarised, it is the whole argument in plan the site first, generate second.
Gap two: no context, so the AI invents facts about your business
Ask a model to write your about page and it will write an about page. It will not leave gaps where the facts should be. It fills them, confidently, with the most plausible-sounding version of a business like yours. That is the mechanism behind hallucination (where the AI states something untrue with complete confidence, because generating fluent text and stating facts are the same operation to it).
In practice that looks like: service areas you do not cover, an accreditation you do not hold, a founding year off by a decade, 'over 20 years of experience' when it is seven. For a clinic or anyone in a regulated field that is worse than embarrassing, because a confident claim about outcomes or qualifications is not a copy problem, it is a compliance problem.
The fix is unglamorous. Write one document about your business and paste it in at the top of every session. What belongs in it is set out properly in what AI needs to know about your business before it writes a word; mine for clients runs to a page or two:
- Exactly what you do and, just as importantly, what you do not do.
- The suburbs, towns or regions you genuinely serve, and the ones you decline.
- Real credentials, registrations, years in business, team size. Numbers you can evidence.
- Whether you publish prices, quote on site, or work to a callout fee.
- The three questions customers actually ask before they book.
- A hard line at the bottom: if a fact is not in this document, ask me rather than filling it in.
That last instruction matters more than people expect. It will not eliminate invention, so you still read every claim before it goes live, but it changes the default from 'guess plausibly' to 'flag the gap'. Read the numbers hardest. A wrong number reads perfectly. What a model invents most often about a small business, and the five minute check that catches it, are in how to stop AI inventing things about your business.
Gap three: no link plan, so internal linking is the first casualty
Internal links run from one page of your site to another, and the anchor text is the words inside the link. Together they are one of the strongest signals you fully control. They tell Google which of your pages is the important one for a given term, and they carry authority from the pages that have it to the pages that need it.
A model writing a single page cannot do this, because it does not know what else exists on your site. So you get one of two outcomes. Either it writes no internal links at all and you publish a structural orphan, or it invents links to URLs that sound like they should exist and do not. I see the second more often than you would think, because the invented URLs look so reasonable that nobody checks them. The blind spot is structural rather than something a better prompt fixes, which is why internal linking is the thing AI gets most wrong is the one gap here I would check on every page.
This is the same objection I have to tools that automate linking on your behalf, which I have set out in why giving up control of your anchor text and internal links is a red flag. Handing that decision to something with no view of your priorities smears the signal across everything instead of concentrating it where you need it.
What to do instead takes about two minutes per page. Before you generate, decide three things: which existing pages will link to this new one, what words those links will use, and which page this new one should point at. Then add the links yourself after publishing. If you want the AI to draft them, give it the actual list of your URLs and tell it to use only those.
Gap four: no real voice, so readers recognise the AI
There is a house style these models produce by default and a lot of people can now spot it at a glance. The three-item rhythm. The opening that restates the question. 'In today's fast-paced world.' Paragraphs that are perfectly balanced and say nothing you could not have guessed. It is not badly written. It is unmistakably generic, and generic has a cost that has nothing to do with Google. The tells are consistent enough to list, and I have listed them in why ChatGPT content sounds like ChatGPT. What a reader actually does in the seconds after noticing - and why that shows up in your enquiries rather than your rankings - is in readers can tell.
The cost is trust. Someone comparing three physios in Geelong or three sparkies in Christchurch is not evaluating your prose. They are looking for evidence that a real person with real judgement is behind the business. Copy that could have been written about any competitor gives them nothing to hold onto, so they keep scrolling. You have not been penalised. You have just been skipped.
'Write in a friendly, conversational tone' does almost nothing about this, because that is the tone it already defaults to. What works is giving it material only you have. The way you explain the tricky thing to customers. The objection you answer on every job. Your real position on the thing your industry disagrees about. Feed it two or three pieces you have written yourself and tell it to match them, then rewrite the worst paragraph by hand. That one rewritten paragraph often does more than the rest of the prompt combined. The fifteen minute version of gathering that material, and why voice is not a list of adjectives, is in how to actually give AI your brand voice.
The order of operations: map, context, generate, link, edit
Put it together and the sequence is: map first, context second, generate third, link fourth, edit last. The generation step is the only one the AI does, and it is the smallest part of the job. That is not a criticism of the tools, just an accurate description of where the work lives. "Edit last" is doing a lot of work in that sentence, and most people edit for style, which is the least consequential pass available - the order I actually edit in, from claims that carry liability down to taste, is in editing AI output: what actually matters.
Two things I would add for owners specifically. First, publish slowly. Volume turns a small structural error into a site-wide one, which is the whole argument in the hidden cost of AI articles at scale: the bill arrives not when you publish, but when someone has to work out what to merge and what to delete. Ten good pages beat a hundred islands, and cost far less to maintain. Rate is the earliest warning sign there is, which is why thirty posts a month is the red flag rather than the writing quality.
Second, point the effort at the right pages. For a local service business the pages that earn money are your service pages, your location pages and your Google Business Profile, not your blog. I have argued that blogs are mostly a waste of time for local businesses, and cheap generation does not change that maths. It just makes it cheaper to build the wrong thing quickly.
What I would not hand over at all
Some pages I would write myself regardless of how good the tooling gets. Anything claiming an outcome, a safety standard or a qualification. Anything a regulator could read. Your pricing page, because ambiguity there costs real money in wasted enquiries. And your about page, because being indistinguishable from your competitors is fatal to the exact thing that page exists to do. The dividing line is not how hard a page is to write, it is how much of its value comes from something only you have - the test I work through in which pages AI should never write for you.
This is also where I part company with the autonomous tools. My scepticism about set-and-forget AI SEO tools is not about writing quality. It is that they automate the judgement and leave you the labour, which is precisely the wrong way round. I would be happy to be proven wrong. I have not been yet.
How this blog is made
It would be dishonest to write all that without telling you how this post got here. This blog is produced with AI doing the mechanical work and a human making every judgement call. That is not a disclaimer bolted on the end, it is the method I have just described.
The judgement calls are mine: which posts get written and in what order, which one owns which search term, what each links to and with what words, and which paragraphs get thrown out because they say nothing. The mechanical work is drafting into a structure I have already decided on, from a context document of my own positions and phrasing. Every post gets read and edited by me before it goes live, and plenty get rewritten hard.
That is the same model I run for client work, and it is the product rather than a limitation. The reasoning is in why I keep a human in the loop, the practical detail in how I actually use AI for SEO, and the step-by-step version - what the machine does, where I make each call, and what happens to a draft that fails a step - in my actual human-in-the-loop workflow. I am comfortable telling you to use these tools because I use them, on my own site, on exactly these terms.
If you have already published a pile of it
Most people reading this are not starting clean. If you have thirty AI pages up already, do not start deleting, because that is how you lose the pages that were quietly working. Work through it in order instead - the short version is below, and the full audit, with what to merge and what to leave alone, is in auditing a site that is already full of AI content.
- Open Search Console and find any search term where more than one of your URLs shows up. Those are your cannibalisation cases, and they go first.
- For each case, pick the page that should win, merge the useful parts of the others into it, and redirect them.
- Fact-check every page for claims about credentials, coverage, guarantees and numbers.
- Check every internal link resolves to a page that exists.
- Rewrite the opening two paragraphs of your top five pages by hand.
- Then build the keyword map, so the next thirty pages do not repeat the first thirty.
Using AI for website content: it writes, you keep the structure
AI is genuinely useful for website content and I would not tell an owner to avoid it. It removes the blank page, which is what has kept half-finished websites half-finished for years. What it does not do is decide what your site should contain, how the pages relate to each other, what is true about your business, or what makes you worth choosing. Those four decisions were always the work, and they are still yours.
So use it for the writing, and keep the structure yourself. Build the map, write the context document, plan the links, put your own voice back into the first two paragraphs, and you get most of the speed with almost none of the damage. Skip all four and publish fifty pages this month, and you get a tidy-looking website that quietly performs worse than the eight-page one you started with.