How to use AI for social media marketing without it showing

Most guides to using AI for social media marketing are a list of tools with a paragraph under each. This is not that. The tool is the least important decision you will make, and we say that as people who sell one. What decides whether AI helps your social media or quietly makes it worse is the order you do things in, and what you give the model before you ask it for anything. Below is that order, the way it runs inside our own product and the way you can run it by hand with any chat model and a free scheduler.

Step one: decide what AI is for, and what it is not for

AI is good at volume, consistency and first drafts. It is bad at knowing what happened in your shop this morning, at judging whether a joke lands with your regulars, and at anything that requires having been there. So the split is not complicated. Everything that is production, the writing, the resizing, the picking of a posting time, the reminding you it is Tuesday, can go to the machine. Everything that is judgement stays with you: what you are actually offering this month, which of the drafts is worth publishing, and every reply to a customer who has written to you.

The reason to be strict about this is not principle. It is that the failures go the other way. When AI is given the judgement half, it does not refuse. It produces a confident answer, and the answer is generic because generic is the average of everything it has read. That is why AI posts sound like AI: not because the writing is bad, but because nobody told it anything specific.

Step two: feed it before you ask it for anything

This is the step nearly everybody skips, and it is the one that decides the outcome. Before the first request, the model needs to have read the business. Not a description of the business. The business.

• Your website, all of it, including the about page and the boring service pages. That is where the actual vocabulary lives.
• Ten to twenty of your real reviews. Customers describe what you do in the words other customers search for, and those are never the words on your homepage.
• Your three best posts from the last year, the ones you were pleased with, and if you can bear it, two that fell flat. The contrast teaches more than the good ones alone.
• Two things you would never say. A price you never discount, a competitor you never name, a tone you cannot stand. Negative instruction is the fastest way to move a model off the average.
• What you are actually selling this month. Not the mission. The offer.

In a chat model this is a long first message, or a saved project with the material attached. In our product it is the step where you paste a URL and the pipeline reads the site, the reviews and the pictures and infers the rest. How that inference works, and where it goes wrong, is written up separately, and if you want to do the voice part properly by hand there is a guide to writing the brand voice document the model should be given.

Step three: let it draft, in batches, against a plan

Once the model has the material, ask for a week, not a post. A week of drafts written in one pass is more consistent than seven written on seven days, because the model can see what it already said and avoid saying it again. Ask for a mix rather than a stream: the six kinds of post a small business has is a workable brief, and the model will produce a better week from "two of these, one of those" than from "seven posts".

Give it a shape for each post as well as a subject. One idea per post. A first line that names the thing a customer would type, not a slogan. Plain language. A specific detail from the material you fed it in step two, or the post does not go out. That last rule is the one that does the work. A post that could be published by any business in your category is a post that should not be published by yours.

If you are working in a chat window, the eighteen prompts that survived our own rewrites are the ones to start from, and there are versions for Claude, Gemini and Copilot written for the way each of those is actually used. Prompts are not magic and we have said so in every one of those articles. They are just the brief, written once, so you do not write it again on Sunday night.

Step four: approve like an editor, not like a reader

This is the step that separates the businesses where AI works from the ones where it fills the feed with noise, and it takes ten minutes a week if you do it in one sitting. Read every draft. Not skim, read. You are looking for four things: is it true, is it specific to us, would I say it, and does it sell what we are actually selling this month. Anything that fails one of those is skipped, not fixed. Fixing drafts is how ten minutes becomes an hour and how the habit dies.

The reason to do it weekly rather than daily is not efficiency, or not only. It is that a person reading one post a day approves on autopilot by Wednesday. A person reading seven in a row develops a standard, because the seventh is being compared to the first. We wrote up the whole routine, and the honest number for a business with a well fed model is closer to five minutes than ten. For a badly fed one, no amount of approving fixes it, and the answer is to go back to step two.

Step five: schedule it, then leave the schedule alone

Once a week is approved, it goes into a scheduler and you stop thinking about it. The tools here are cheap and interchangeable: Buffer's paid plans start at 5 dollars a channel a month and its free plan includes an AI assistant, and the free tiers across the category will carry a small business a long way. Ours does the drafting and the scheduling in one place, which is convenient and is also not the point. The point is that the week is done, on Monday, and the rest of the week is for the business.

How much to schedule is a question with a boring answer: fewer posts than the tool can produce, and more than you were managing before. Three good ones a week beats seven you did not read. The reason a scheduler matters more than the volume is that consistency is what compounds, and consistency is the thing a person with a business to run cannot supply by hand.

Step six: reply yourself, always

AI can draft a reply. It should not send one. The person who has written to you is a customer, or about to be, and they can tell the difference between an answer and a template. Comments and messages are the highest value minutes in social media marketing precisely because they are the part that does not scale, and handing them to a machine gives away the only thing you have that a chain does not. Use the model to remind you a message is waiting. Type the answer yourself.

The four ways it went wrong on our own product

These are ours, not a competitor's. Each was found in production, each was fixed, and each is a general lesson about using AI for social media rather than a bug report.

• It invented facts to fill silence. The company that did not exist, above, is the extreme case. The everyday case is a draft that mentions an opening time, an award or a years in business figure that appeared nowhere in the material, because the sentence needed one. Every draft with a number in it now gets read by a person before it goes out, and that rule applies to any tool, including ours.
• It described the pictures instead of the business. When we asked one model for a caption and a picture in the same call, the caption drifted towards describing the picture and away from the offer. Separating the two, words first and then a picture to match, fixed it. If you use a chat model for both, do them in two messages.
• It kept the logo out of the picture and did not say why. For weeks a share of generated posts came back without the customer's mark, and the interface said nothing. The reason was mundane, the crawler had not found a usable logo file, but silence made it look like a decision. Now it says what it could not find. The lesson: when a tool leaves something out, find out whether it chose to or failed to, because the fix is different.
• It wrote posts for a business rather than for the person running it. The early drafts were competent and airless, the kind of thing a franchise head office sends out, because the model had been fed the website and nothing else. Reviews and real past posts, added to what it reads, moved the drafts from correct to recognisable. Step two of this article is that fix.

Do you have to say a post was written with AI?

For ordinary text posts, no platform requires it and no reader expects it. Meta labels images, video and audio that it detects as AI generated, or that the poster discloses as such, with an "AI info" label; it does not ask you to declare a caption. Google's stated position for search is that it rewards helpful content however it was produced and penalises content made to manipulate rankings, however that was produced. The one place disclosure is not optional is the old one: if a post is an endorsement, a review or a paid relationship, the FTC's endorsement rules apply exactly as they did before, and no amount of AI changes who is responsible for the claim.

What readers do notice is not the technology. It is a post that could have been written for any business in your category. That is the failure to guard against, and it is guarded against in step two and step four, not in a disclaimer.

What this looks like in a week

• Monday, twenty minutes: tell the model what this month's offer is, ask for the week, read all of it, approve four or five, schedule them.
• Every day, five minutes: reply to anything a person has said to you. Not the model. You.
• Friday, five minutes: look at which of the week's posts got a comment or a message. Tell the model next Monday. It is the only feedback loop that matters and it is the one nobody sets up.
• Once a month, ten minutes: reread what the model has been fed. If the offer changed, the material changes. A model working from June's website writes June's posts.

That is roughly forty minutes a week, and it is forty minutes of judgement rather than forty minutes of typing. Whether the drafting is done by our product, by a chat window and a free scheduler, or by a freelancer whose rates we have written up honestly, the forty minutes do not go away. They are the job. Everything else in this article is about making sure they are the only job.