Guides

How to stop AI copy sounding generic

By Rory Wilmer, InkByte · 6 min read · 16 October 2026

You can usually spot AI-written marketing copy within a line. It's not that it's wrong. It's that it's interchangeable. Swap the brand name and the paragraph works just as well for a competitor. For a freelancer, that's the real cost: clients pay for copy that sounds like them, and generic copy has to be rewritten by hand.

This guide covers why AI copy drifts generic and the practical habits that pull it back.

Why AI copy sounds the same

A language model with no context writes what is typical. Ask for a coffee brand's homepage and you'll get the most typical coffee brand homepage: passion, craft, ritual, elevate. Those words aren't errors. They're the average of the category. The model can't be specific about your client because nothing in the prompt is specific.

So the cure is mostly about input. The more concrete, client-specific material the model has, the less it falls back on the average.

1. Ban the category's favourite words, with reasons

Every sector has its tired vocabulary. Coffee has "artisan" and "ritual". Tech has "seamless" and "leverage". Charities have "make a difference". Make a list for your client's category and put it in the instructions as a banned list.

Plain-English style guides are a good starting point for corporate filler. The GOV.UK style guide keeps a list of words to avoid with plain replacements, for example "use" instead of "utilise", "work with" instead of "collaborate", and "allow" instead of "empower". It also warns against metaphors such as "drive" and "going forward" because they slow comprehension.

Add a reason to the list. Anthropic's prompt guidance notes that explaining why a rule matters helps the model generalise beyond the exact words you listed. "Avoid 'elevate': it's vague and every competitor uses it" will also steer it away from "enhance" and "take to the next level".

2. Replace adjectives with evidence

Generic copy is heavy on adjectives and light on facts: "exceptional quality", "unrivalled service". Specific copy names the thing: the farm, the roast date, the turnaround time, the number of years trading. Give the model a short list of real facts it may use, and tell it to prefer a fact over an adjective wherever it can.

Just as important, tell it not to invent facts. Ask it to write a placeholder such as [TO CONFIRM] instead of making up a figure or a claim. Invented specifics are worse than generic ones, especially if the client is in a regulated area like health, finance or alcohol.

3. Feed it real copy, not descriptions of copy

"Witty but warm" can mean almost anything. Three paragraphs of the client's best real copy can only mean one thing. Anthropic recommends three to five examples, relevant to the real task and varied enough that the model doesn't latch onto one pattern. Mix channels so it learns the voice, not the format of a single post.

If you have it, include something the client rejected and say why. "Too salesy, too many exclamation marks" is a very clear signal.

4. Separate voice from tone

A common reason AI copy feels off is that it uses one register for everything: the same chirpy energy in an apology email as in a product launch. Mailchimp's content style guide puts the distinction neatly: you have the same voice all the time, but your tone changes with the situation and with the reader's state of mind.

Tell the model how the client's tone moves between channels. For example: "Social: playful, short. Customer service: calm, plain, no jokes. Press release: factual, third person." Then name the channel in every request.

5. Watch for the structural tells

Generic AI copy has habits beyond word choice. Watch for these and add rules against the ones your client wouldn't use:

• Everything in threes ("fast, simple and reliable") in every paragraph.
• Rhetorical questions as openers ("Ever wondered...?").
• The "it's not X, it's Y" construction, over and over.
• A summary line at the end that restates the paragraph.
• Exclamation marks and emoji the brand never uses.
• Headlines in Title Case when the client uses sentence case.

Anthropic also suggests telling the model what to do, not just what to avoid. Pair each ban with the alternative: "Open with the most useful fact, not a question."

6. Edit with a checklist, not a feeling

Even a good setup drifts. Before anything goes to the client, run a quick pass/fail check:

• Could a competitor publish this unchanged? If yes, fail.
• Any word from the banned list? Fix it.
• Any fact you can't trace to the client's material? Remove it or mark it.
• Does the tone fit this channel?
• Would the client say this sentence out loud?

When something fails, fix the instruction that should have caught it, so the next draft doesn't repeat the mistake.

7. Keep each client separate

If you use ChatGPT Custom Instructions for one client, remember OpenAI applies them to all chats. A second client's work in the same account will pick them up. Use a separate Project or custom GPT for each client, or in Claude, a separate Project for each one, so the right voice is always loaded.

The short version

AI copy goes generic when the model knows nothing specific about your client. Give it banned words with reasons, a list of real facts, three to five varied examples, and tone rules for each channel. Then check every draft against a short checklist. The writing doesn't have to be human-made to sound like the brand, but the input does have to be brand-specific.