- A social post fails on its angle, not its grammar. A model supplies the grammar and has no access to the angle, because the angle comes from things only you saw this week.
- On 30 July 2026 LinkedIn shipped a "seems like AI slop" button and said flagged posts get less reach outside your network. Detection is now a ranking input, not just a rule.
- The three platforms police this differently. TikTok requires you to label realistic AI media, Meta applies a label from file metadata, LinkedIn has no labelling rule at all and demotes instead.
- Of the five formats a small shop actually posts, one can be drafted end to end by a model, three need a human fact at the top, and one should never be automated.
- AI detection firm Pangram put more than 40 percent of long form LinkedIn posts as fully AI generated, which is why the platform reacted rather than whether it should have.
- The constraint on posting three times a week was never writing time. It is angle supply, and a model cannot manufacture that.
Thursday evening, the shop is closed, and you are looking at an empty compose box for the third time this week because somebody told you consistency matters. You paste the product name into a chatbot, ask for a caption, and get something perfectly serviceable. It goes out. It gets four likes, three of them from people you know personally. Next Thursday the same thing happens.
This is the actual experience of using AI for social media, and the tools are not going to tell you why it happens. It is not that the writing was bad. Read it back: the grammar is fine, the length is right, there is a call to action at the end. It failed because there was no reason for anybody to stop scrolling, and no model can invent that reason from a product name.
Why does a good draft still fail?
Because a post is two things and a model only supplies one of them. The draft is the sentences. The angle is the reason those sentences exist: a thing that happened, a number that surprised you, a mistake you made, an opinion somebody might disagree with. Language models are genuinely good at the first and structurally incapable of the second, because the angle lives in information they do not have.
Think about what your best performing post ever actually was. Almost certainly it was not a description of a product. It was the supplier who let you down two days before a launch, or the customer who ordered the same thing eleven times, or the week you got the pricing wrong and said so. Every one of those is a fact from inside your business, and no amount of prompting extracts a fact that was never supplied.
Which reframes the workflow. The failure is not that you used AI. It is that you asked it for the whole post instead of the half it can do. Give it the fact and it will find you fifteen ways to say it, most of them better than your first attempt. Give it nothing and it will give you the average of everything ever written about your category, which is precisely what the platforms are now filtering out.
What did LinkedIn actually change on 30 July 2026?
It added a report button and it told everyone that flagged posts lose distribution. TechCrunch reported that the option sits in the three dot menu on every post, hides the post from the person who clicked it, and feeds a signal LinkedIn uses to tune its detection models. Chief product officer Hari Srinivasan called it a source of signal for identifying slop, and the company said new classifiers will cut how often flagged content is recommended to people outside the author's network.
Two details in that report matter more than the button. First, LinkedIn discontinued its own "enhance your post" feature, the one that rewrote your draft, and replaced it with a proofreading tool that leaves your voice alone. A platform withdrawing its own rewriting tool is a clearer statement of direction than any policy page. Second, accounts showing heavy AI usage will get a private flag in their own dashboard.
The scale it responded to came from AI detection firm Pangram, whose figures TechCrunch cites: over 40 percent of long form LinkedIn posts fully AI generated, and LinkedIn hosting roughly 62 percent of all the AI content scanned across major social networks. Whether you trust a detector's precision is a fair question. Whether the platform acted on it is not in doubt.
Reduced reach is not a ban. Nothing described here stops you posting AI assisted content, and LinkedIn's own policy page does not prohibit it. The penalty is quieter and worse: your post is shown to fewer people who do not already follow you, which is exactly the audience you were posting to reach.
What does each platform's own guidance actually say?
Three platforms, three completely different regimes, and only one of them mentions AI in its rules. The table below reads the primary sources rather than the summaries of them, because the differences decide what you have to do before you press post.
| Platform | What its own guidance requires | How it is enforced | What that means for a shop |
|---|---|---|---|
| TikTok | Label content completely generated or significantly edited by AI when it shows realistic images, audio or video | A label you apply, or a sticker or caption; the platform may apply one itself | Any AI video of a real looking person or place needs the label. Text captions do not. |
| Meta, on Facebook and Instagram | No creator rule quoted; Meta adds an AI Info label when it detects industry standard AI image indicators or when you disclose | Automatic, from file metadata, and Meta says the method may miss content edited with AI | Your generated image may get labelled without you doing anything, and may not. |
| Share only information that is real and authentic, and nothing false, misleading or intended to deceive. No AI clause. | Reader reports, classifiers, and reduced recommendation beyond your network | No box to tick. The penalty arrives as silence rather than a notice. | |
| All three | Nothing forbids using a model to draft text | Ranking, not removal | The risk is distribution, not enforcement. |
TikTok's own announcement is the clearest of the three: the requirement covers realistic images, audio or video, and the stated purpose is helping viewers contextualise what they are watching. Meta's transparency page describes labelling that happens to you rather than a rule you follow, and concedes its methodology may not catch everything edited with AI. LinkedIn's professional community policies require that you use your true identity and share only what is real and authentic, and separately prohibit repetitive unwanted promotional content, which is the clause a daily generated post is most likely to run into.
Which formats can a model carry on its own?
One of the five, and it is the one nobody enjoys writing. Here is the honest split across what a small shop actually posts, with the reason rather than a rating.
| Format | Can a model carry it alone? | What it needs from you |
|---|---|---|
| Product announcement, new item in stock | Yes, once you give it the attributes | Nothing beyond the facts and one sentence of house style |
| Behind the scenes or process post | No | The specific thing that happened, in one line. The model expands it. |
| Opinion or industry take | No | Your actual position, including the part somebody could argue with |
| Customer story or testimonial | No | The real quote and permission. Inventing either is the one that ends badly. |
| Reply to a comment or a complaint | Never | You. A generated apology reads as one within two lines. |
The pattern in that table is worth stating plainly, because it generalises past social media. A model can carry a format when the information content is already written down somewhere and the job is presentation. It cannot carry a format whose value is the information itself. Product specifications are written down. What you think about your suppliers is not.
How should a one person shop actually use it?
Reverse the order. Most people open the tool and ask for a post. Instead, keep a running note of facts from the week and hand the model one fact at a time. A note entry is a fragment: "third time this month someone asked whether the small size runs big", "returned pallet, wrong colour, second time from this supplier", "sold out of the green one in four days".
Each of those is a post. The model's job is to turn a fragment into three drafts in different registers, and yours is to pick one and change a word so it sounds like you. That takes about four minutes and it produces something nobody else could have written, which is the entire point. The same discipline applies to product copy, and we worked through it at length in the piece on what AI copywriting tools do well and where they flatten a brand.
Two further habits are worth building. Never let the model write the first line unaided, because the first line is what decides whether the rest is read, and a model defaults to a summary of what follows. And never post a draft on the same day you generated it if you can avoid it: reading it cold the next morning catches the phrases that sounded fine and are not yours.
Does anyone actually detect this, or is it vibes?
Both, and the vibes are the part that costs you. Automated detection is imperfect and every honest account of it says so, which is why LinkedIn built a human report button rather than trusting a classifier alone. But your readers are running their own detector, and it has been trained on two years of the same three sentence structures. When a post opens with a rhetorical question and closes with a lesson, people know, and they scroll.
The measurable consequence is the reach reduction TechCrunch described, which applies to recommendations beyond your own network. For a small shop that is the growth channel. Your existing followers will see you either way. The people who might become customers arrive through recommendation, and that is the traffic a slop classifier sits in front of. If you want the longer version of how detection works and what it does and does not catch, we covered it in the guide to AI content detection for shops.
What about images and video?
Here the rules get specific, and they are the only place in this article where you can be formally wrong rather than merely ignored. TikTok's requirement covers realistic images, audio or video that were generated or significantly edited by AI, and it is aimed at content that would lead a viewer to believe a person or an event is real. A stylised graphic with your prices on it is not that. A photorealistic image of a person holding your product is.
Meta's approach runs the other way, applying its AI Info label when it detects standard indicators in the file itself. This has a practical consequence people discover late: an image you generated and then edited may carry the metadata anyway, and an image you barely touched may not. You do not fully control the label, so do not build a plan that depends on it being absent.
For product photography specifically the stakes rise again, because the image sits next to a price and a buy button rather than in a feed. We set out where an illustration stops being an illustration in the article on AI product photography, and the short version is that anything a customer will use to judge what arrives in the box needs to be true.
What about automated comments and engagement?
Leave them alone entirely, and the number that explains why is in the same TechCrunch report: LinkedIn blocks hundreds of thousands of automated comment attempts every day. That is not a platform being cautious about a fringe behaviour. That is a platform that has built infrastructure against a flood, and anything you send through the same pipe is arriving in a queue designed to stop it.
The pitch for automated engagement is always the same, that commenting on fifty posts a day builds visibility. It did, briefly, in an era when nobody was doing it. Now the generic supportive comment is the single most recognisable AI artefact on any network, and the cost of being recognised is not a warning. It is that a real potential customer reads your comment, understands immediately that a machine wrote it, and forms a view of your shop that no product page will undo.
The same logic applies to bulk direct messages, which LinkedIn's own policy addresses without ever mentioning AI: repetitive unwanted promotional content is prohibited outright, whoever or whatever composed it. A model does not change what the rule covers. It just makes it much easier to break at volume, which is the pattern behind most of the accounts that quietly stop getting reach and never find out why.
How many posts a week is the right number?
As many as you have facts for, which is probably not three. The advice to post on a fixed schedule came from an era when distribution rewarded frequency, and it now collides with a ranking system that penalises the thing frequency forces you to do, which is publish when you have nothing to say.
A shop with one genuinely specific post a week and nothing else will do better than the same shop posting four times with three fillers, because the fillers train both the classifier and your followers to skip you. If the note of weekly facts is empty on Thursday, the correct action is not to generate something. It is to skip the week.
This is where the honest accounting helps. Writing time was never the bottleneck for a business this size. Four minutes a post at three posts a week is twelve minutes. Nobody abandoned social media over twelve minutes. They abandoned it because it produced nothing, and it produced nothing because the input was empty.
What this means
The platforms have stopped treating generated text as neutral and started treating it as a quality signal, and they are doing it through ranking rather than rules, which makes it invisible from the inside. You will never get a notification saying your reach was cut. You will just keep getting four likes.
The practical response is not to stop using AI, and every platform quoted here allows it. It is to move the model to the position it is actually good at: a drafting assistant that takes one true, specific fact you supply and turns it into something readable, quickly, in your voice, without flattening it into the house style of the entire internet. Keep the note of facts. Feed it one at a time.
And if the tooling around your shop is what makes collecting those facts hard, that is a solvable problem rather than a discipline problem. Knowing what sold out, what got returned twice and what customers keep asking about is ordinary shop data, which is why we build that reporting into the storefronts generated on our ecommerce website builder, and why the pricing page is built around usage rather than seats. The posts write themselves once the facts are in front of you. That was always the hard part.