Why Your AI Social Media Posts Sound Like AI (And What to Do About It)
Here's a post you'll recognise. Business name in capitals, flanked by sparkle emojis. A short intro about being reliable and friendly. Then nine services listed one by one, each with its own cute emoji: a broom for cleaning, a house for property handovers, a key for key handovers. It closes with "I'm happy to help. Send me a message for more information or availability."
You've seen this post before. You've probably seen it dozens of times from dozens of different businesses. That's the problem.
Obviously you can't miss the emojis, but one of the other biggest giveaways? It's the em-dash (—). Right there in the second line, something like "Based near Poitiers, covering up to 1 hour surrounding areas." Except it reads "Based near Poitiers — covering up to 1 hour surrounding areas." That long dash between two clauses is something almost no one writes naturally in a social media post. AI uses it constantly because it's common in the polished written English the model was trained on. It's practically a signature at this point. If you spot an em-dash in a Facebook post from a local business, there's a very good chance a human didn't write it.
The other tells are less precise but they stack up. The systematic one-emoji-per-service formatting reads like a template because it is one. The AI assigned a relevant icon to each line because that's what posts like this tend to do. "Help take the stress away" sounds warm but could apply to literally any service in any industry. "I'm happy to help" is how AI closes things when it has nothing specific to say. "Whether you need help getting your property guest-ready, preparing to move, or simply keeping on top of things" is a sentence construction AI reaches for constantly, three parallel clauses held together with "whether."
None of this is a problem with AI, it's 2026 and I'm not here to judge. But sounding like everyone else? I doubt anyone wants that. It comes down to a problem with the prompt.
When you open ChatGPT or Claude and type "write a social media post for my cleaning business," the tool has nothing to go on. It doesn't know how you speak, what you find funny, what makes your service different from every other cleaner in the area, or what you'd never say in a million years. So it produces the average, the statistical centre of every cleaning business post it has ever seen. Generic phrasing, template structure, emojis as bullet points, a closing line that could belong to anyone.
The fix is not to stop using AI for content. The fix is to give it enough to work from before you ask it to write anything. That means your own voice written down in plain language. Not your job title, your actual voice. How you normally talk to customers. The things you'd never say. Examples of messages you've written that felt right. Specific details about your situation, your location, your clients, the kind of work you actually do. When the AI has that context it stops averaging and starts writing toward something specific. A great way to provide more context is to pass the agent copies of your own previously written emails.
It also means being deliberate in the prompt itself. "Write a social media post for my cleaning business" is close to the worst prompt you can give. Something like "write a post for a one-person cleaning service based near Poitiers, mainly doing gite handovers and end-of-stay cleans, who wants to come across as practical and straightforward rather than corporate and cheerful" gives the model enough to deviate from the template. It still won't sound exactly like you. But it will sound less like everyone else.
And tell it what not to do. If you wouldn't use nine different emojis in a note to a neighbour, say so. If you'd never use "I'm happy to help" as a sign-off, say so. If you hate em-dashes, say so. The model defaults to these things because they're common in the genre. You have to specifically ask it not to.
There's a broader point here about how most people use these tools. They treat them as text generators: describe a topic, get a post. The results tend to be fine in a forgettable sort of way. The people getting better output are treating the AI more like a collaborator with a short memory that needs briefing every single time. Give it your voice, give it the context, tell it what you don't want, then edit what comes out toward something that sounds like you actually wrote it.
The all-caps headline, the emoji list, the generic CTA, the em-dash. None of these things on their own are a disaster. But together they signal to anyone paying attention that no actual person sat down and thought about this post. In a local market where most businesses are generating the same AI-average content, the bar for standing out is actually quite low. You don't need to be a great writer. You just need to sound like a person.
Andrew Watts, web developer based in Nouvelle-Aquitaine, France.