Open the websites of five companies in your sector. Read the homepage headline and the first paragraph of each. There is a fair chance you could shuffle them and nobody would notice, including the companies.
This was a problem before AI. Marketing copy has always drifted towards whatever the rest of the industry was saying. What has changed is the speed. Everyone now has access to the same few writing tools, and those tools pull in one direction.
Why models regress to the average
A language model writes by predicting which words are likely to come next, based on patterns in an enormous amount of existing text. Likely is the important word. The most likely way to describe an accountancy firm is the way thousands of accountancy firms have already been described.
So when you ask for “a homepage intro for a B2B software company”, you get the centre of the distribution: competent, familiar and belonging to nobody. The model is doing exactly what it was built to do.
Three things make it worse in practice:
- Thin prompts. The less you tell the model about you, the more it fills in with the generic.
- Shared prompts. Your competitors are typing much the same request into much the same tool.
- Politeness by default. These tools are tuned to be agreeable and balanced, which strips out the edge that makes a voice recognisable.
What that does to your brand
For a company with a turnover of a few million, differentiation is rarely about being the cheapest or the biggest. It is usually something softer: you are easier to deal with, you know one niche better than anyone, you say what the larger firms will not. Those things are carried almost entirely by how you sound.
When your copy converges with everyone else’s, a buyer comparing shortlists has nothing to go on except price. That is a poor place to compete from.
There is a search cost too. A page that says what every other page says gives a search engine, or an AI answer tool, no particular reason to choose yours.
A method for keeping your voice
None of this means you should stop using AI for writing. It means the model needs material it cannot get anywhere else, and a person needs to finish the job. Five things, in order of effort.
1. Write voice notes
One page, not a brand book. Describe how the company talks in terms someone could act on. “Friendly and professional” is useless because every company claims it. Try pairs that force a choice:
- Plain before clever
- Direct, even when the news is unwelcome
- Dry humour allowed. Exclamation marks are not
- We explain the jargon or we do not use it
Then add three short samples of real writing you consider right: an email, a paragraph from a proposal, a social post. Models imitate examples far better than they follow adjectives. So do freelancers.
2. Keep a banned-words list
Every sector has its clichés, and AI tools have their own favourites on top. Start a list and add to it whenever a word makes you wince. Include:
- Your industry’s stock phrases (“solutions”, “end-to-end”, “best in class”)
- The words AI drafts overuse
- Anything your customers would never say aloud
Paste the list into every brief. Then search for the words anyway before publishing, because the model will ignore the instruction more often than you would expect.
3. Use real customer language
Your customers describe their problem in their own words, and those words are almost never the ones on your website. Collect them from:
- Sales call notes and recordings (with permission)
- Support tickets and enquiry form messages
- Reviews, yours and competitors’
- The exact phrases people type into your site search
Keep a running document of verbatim phrases. When a customer says “we kept getting passed around”, that beats “fragmented service experience” every time. Feed these into your briefs, with any personal details taken out first.
4. Have opinions
The quickest way to sound unlike a machine is to take a position. A model asked to write about your industry will give a balanced overview. You are allowed to say what you think is overrated, which common advice is wrong for companies like your customers, and what you would do instead.
Write down five things your company believes that a competitor might disagree with. If you cannot find five, that is a positioning conversation worth having with your CEO, and it is more valuable than any tool.
A test: if no reasonable person could disagree with the sentence, it is not an opinion, and it is not doing any work.
5. Add the specifics only you know
This is the part no model can supply. The detail that comes from doing the work:
- What goes wrong on the third week of a typical project
- The question every new customer asks, and the honest answer
- Numbers from your own operation (real ones, with permission)
- Named places, products, processes and trade-offs
One paragraph of this is worth more than a page of well-formed generalities.
Before and after
Here is a generic line of the kind any tool will produce for a commercial cleaning firm:
“We provide tailored cleaning solutions that help businesses maintain a safe, healthy and productive working environment.”
And a version built from voice notes, customer language and a specific:
“Most offices change cleaner because nobody turned up twice in one month. Ours clock in on site, and you get a text if someone is running more than fifteen minutes late.”
The second one is a hypothetical, but notice what it contains: a customer’s actual complaint, an opinion about why contracts fail, and an operational detail a competitor would have to match or avoid mentioning.
Putting it into a workflow
A voice document only helps if it is used. Build it into how work gets made:
- Keep the voice notes, banned list and customer phrases in one file.
- Paste that file into the brief every time, for a model or a person. There is a template in how to brief AI and humans.
- Let the tool draft.
- Edit by hand for opinion and specifics. This is the step that cannot be skipped.
- Run the swap test before publishing: put a competitor’s name on it and see whether anything has to change. More checks are in how to spot AI slop.
Review the file every quarter. Voices shift, and the banned list always grows.
The part that takes a person
You can give a model your voice notes and get something closer to right. What it will not do is decide what your company thinks, notice that a sentence is true but dull, or know that your managing director would never say “passionate”. Someone who knows the business has to read it and make those calls.
That editing step is where we spend most of our time at DGTL. The drafts are AI-assisted and a named person rewrites and signs off each one. If you would like help building the voice file in the first place, get in touch.