An AI writes a new post every time this blog is due to publish, this one included. And over the last year we did something almost no company does with its content: instead of guessing which "hook" works better, we pulled a full year of analytics and actually measured it.
The result surprised even us: there's a 35x gap between the type of post that performs best and the one that performs worst, and it has nothing to do with posting time, length, or format. It comes down to a single decision: which idea you choose to tell.
The experiment: a full year, not two weeks
Anyone who publishes content has an opinion about what works. Almost no one checks that opinion against a full year of data, because it takes patience: a post doesn't hit its final numbers on day one, it keeps growing for weeks. One of ours went from 74,210 to 123,161 impressions (+66%) between two measurements taken weeks apart. Measuring after two days would have led to the wrong conclusion.
So we did the full exercise: 36 posts over 365 days, on a profile that reached 2,186 followers, counting only posts that had already been live for at least 5 days at the time of measurement. Each post was grouped by its angle, not its topic:
- Debunking a label: a term the market gets wrong (RAG, agent, hallucination, copilot) explained from scratch.
- Real case with a number: a before/after with a concrete figure inside (time, percentage, records).
- Our own internal engineering: how we built something under the hood, with no business figure attached.
- List of tips: best practices or warning signs, with no case and no number behind them.
The 4 angles and the 35x gap
The gap between the best and worst group isn't a stylistic nuance, it's an order of magnitude:
| Angle | Average impressions | What it looks like |
|---|---|---|
| Debunking a label | 47,101 | "RAG isn't uploading PDFs, it's this, and when it does NOT work" |
| Real case with a number | 10,927 | A verifiable before/after with a figure |
| Our own internal engineering | 1,545 | How we built something in-house, no client case |
| List of tips | 1,352 | Best practices everyone already agrees with |
A 35x gap between the top and bottom group. It's not that the worst group was badly written: one of those tips was honest, well argued, and still got exactly zero comments. The problem wasn't the quality of the sentence, it was the choice of angle before the first word was even written.
Why this happens: the decision rule
Cross-referencing the four groups against comments, saves, and shares, a clear pattern shows up. A post performs when two things are true at once:
- It leaves the reader with something concrete to take away: a number (a before/after, a percentage, a time saved) or an explicit decision rule like "if you're missing X, it's not Y." A tip everyone already agrees with doesn't give anyone a reason to react: no comments, no saves, no shares.
- It leaves room for someone to push back. The only two posts of ours that truly took off are the only two that got a third-party comment from someone adding their own experience. Closing with an open question invites exactly that; closing with a generic "has this happened to you?" doesn't.
Before we sign off on any post now, we ask two questions:
- Would an operations director forward this to their team because it gives them something to argue with?
- Does it leave room for an expert in the field to come in and add nuance?
What this means for your company, whether you post on social media or not
You don't need an AI engine writing your content for this to apply to you. It holds equally if you write it yourself, if your marketing team does, or if you use AI as support:
- Does the post name a term your customers have heard poorly explained by a vendor, not just one engineers use among themselves?
- Does it carry a real number (never made up): a before/after, a percentage, a time saved?
- If there's no number, does it at least leave a decision rule the reader can apply today?
- Does it close by inviting pushback, instead of closing flat?
- If it's a list of tips, is there a real case backing it up, or is it just generic best practices?
If most of your answers are "no," the post is probably well written and will still perform like the worst of our four groups.
How we apply this at AutoBoost
This finding directly changed the order in which our AI picks a topic every time it's due to write: first it checks whether there's a label to debunk, then a real case with a number, and only if neither is available does any other angle come into play. It's the same principle we apply on client projects, except here the "client" is our own content: measure first, decide second, never the other way around.
It's also the same approach we used on an AI-driven sales engine we built for a training company: part of that project was precisely an AI content studio that decides what to publish on social media based on what has already worked, not on what "sounds good" in a meeting. If your company publishes content and wants to stop picking the topic by gut feeling, you can see how we build this in our services and get in touch to talk about your case.
