Still loading. On a slow connection this can take a minute. Reload

Sign in

Your whole marketing team, in one AI

Watch the film50 s

Your marketingand sales.On autopilot.

Give it a link to your business. It plans the month, makes the posts, ads and films, posts them everywhere, answers every comment and finds the buyers. You wake up to the results.Give it a link to your business. It makes the posts, ads and films, posts them everywhere and finds the buyers.

No website? Any link to your work is fine, or sign up with just your email.

From $49 a monthFree first look, no card▶ Watch the 50-second film

Listed and ranked
by MarList

The Marsell blog

How to use AI for marketing without sounding like everyone else

AI copy sounds alike because the models converge on the same answers. Nine habits that make AI-written marketing sound like your business again.

By Ilya Arbabi, 10 November 2026

To stop AI marketing sounding like everyone else's, give the model what only you have and keep the creative decisions for yourself. Paste in your own words and real examples, ban your trade's clichés, set hard constraints, choose the angle yourself (or draw it at random from lists you wrote), ask for many options, add specifics only you know, cut the tells, keep it all in a voice file and test the result on real customers. Each habit below comes with a short before-and-after for a made-up business.

Why AI copy sounds the same

The models converge on the same answers. Emily Wenger and Yoed Kenett gave three standard creativity tests to 22 language models and 102 people (Wenger and Kenett, 2025). One by one, the models scored about as well as the people, but their answers were far more alike than the people's were, even when only one model per family was counted.

Michał Brzozowski and Neo Christopher Chung asked models to invent experts (Brzozowski and Chung, 2026). Each model family reached for its own favourite names, often in fixed pairs (Elena Vasquez and Marcus Chen from Claude, Aris Thorne and Lena Petrova from Gemini), and those names now turn up across hundreds of AI-written documents, including fake testimonial sections on commercial websites.

So switching tools won't save you. Left to its defaults, any model will write something close to your competitors' copy. The full story is in Why every AI writes the same ad. The rest of this post is about what to do instead.

Nine habits that make AI copy sound like you

The examples use Halfpenny Cycles, a bike repair shop I made up for this post.

1. Give it your words and real examples

A model that only knows your business name writes about an average business. Paste in three to five things you wrote that customers responded to: a reply to an enquiry, your About page, an email that got answers, a transcript of you describing the shop out loud. Anthropic's prompting guide for Claude calls examples one of the most reliable ways to steer tone and structure, and suggests three to five (Anthropic). Ask for your rhythm and vocabulary, not your sentences copied back.

2. List your trade's clichés and ban them

Every trade has phrases everyone uses. Read ten competitors' websites and write down what repeats. For a bike shop that might be "passionate about cycling", "keep you on the road", "all your cycling needs", "expert team" and "ride with confidence". Put the list in your prompt as phrases to avoid, then search the draft for them anyway, because models don't always follow a ban. A list from your own market beats a generic one: it catches what your customers are tired of reading.

3. Set hard constraints

Open prompts get average copy. Constraints push the model off its defaults: a word limit, a number that must appear, a named problem, a banned kind of adjective, one specific reader.

4. Make the choices yourself, or draw them at random

Sameness comes from the decisions you leave to the model: the angle, the format, the setting, the character, the joke. So don't leave them open. Pick the angle yourself, or write short lists of your own (ten angles, ten formats, ten settings), choose one from each with a dice roll or a spreadsheet's RANDBETWEEN, and give the model the result.

A draw of "the cost of putting it off", "a one-line confession" and "the school run in the rain" gives Halfpenny Cycles a post other shops are unlikely to have: "Confession: our mechanic did Tuesday's school run in the rain on brake pads he'd been meaning to change since June. Pads are £18 fitted. Don't be him."

Marsell works on the same idea. Before any model writes, it draws the choices a model would otherwise make by habit (the angle, a constraint, an era, a genre, the lead character, the place) in code, from large pools, with a seeded random generator. The draws steer away from the brand's recent ones and stay within the brand's own tone. The chaos engine explains how it works.

5. Ask for many options, then choose

Ask for 30 headlines, not three, and assume the obvious ones are the ones your competitors got too. Cross out anything another business in your trade could put its name to. In a 2024 Wharton working paper, GPT-4's product ideas were less varied than a pool of MBA students' ideas, but the gap nearly closed when the model was asked to list 100 short ideas, then make them bold and different, then describe them (Meincke, Mollick and Terwiesch). The choosing stays with you. A model can rank its own options, but its taste is the average you're trying to get away from.

6. Feed it specifics only you know

A model can't know your customers, your numbers, your streets or your worst week, and those are what make copy yours. Keep a running list: questions customers ask, the words they use (ask before you name or quote anyone), numbers you can prove, local places, stories from the job.

7. Cut the tells

Some words have become giveaways. A study of more than 15 million biomedical abstracts found "delves" used 28 times as often in 2024 as the trend before ChatGPT predicted (Kobak and colleagues). Cut:

The last one matters most. The invented names from the Ghost Couple study turn up in fake testimonials, and the US Federal Trade Commission's 2024 rule bans reviews and testimonials from people who don't exist, "such as AI-generated fake reviews" (FTC). The plain guide to AI marketing covers the other rules.

8. Keep a voice file

Put all of this on one page that you paste in at the start of every session, or save as the instructions in your tool's project settings:

Each time you make the same edit twice, add it to the file.

9. Check with real customers

The final test is whether it sounds like you to people who know you. Show two versions to five regular customers and ask which one sounds like the shop. Send two subject lines to two halves of your list and count replies, not opens. If a regular says "that doesn't sound like you", believe them over any tool.

FAQ

Why does AI marketing copy all sound the same?

Because the models behind the tools converge on the same answers. In a 2025 study, answers from 22 language models to creativity tests were far more alike than answers from 102 people, even though each model scored about as well as the people did on its own.

Will a different AI model make my copy more original?

Not by much. The similarity Wenger and Kenett measured held across model families, and Brzozowski and Chung found that each family has its own stock of invented names. What you give the model, and what you decide yourself, matters more than which model you use.

How do I get AI to write in my brand's voice?

Give it a voice file: who you're for, words you use and avoid, facts it must get right and three to five real examples of your writing. Then edit every draft, and add each fix you repeat to the file.

Can I let AI write customer testimonials?

No. A testimonial has to come from a real customer describing a real experience. In the US, the FTC's rule bans fake reviews and testimonials, AI-generated ones included, and the UK has banned posting or commissioning fake reviews since April 2025. Use AI to draft the message asking real customers for a review instead.

Sources